From 8d6f9eaeb0ae277b30f976f61a33894b4c7cd8cb Mon Sep 17 00:00:00 2001
From: dylanyunlon <dylanyunlon@users.noreply.github.com>
Date: Thu, 13 Aug 2026 09:24:42 +0000
Subject: [PATCH] =?UTF-8?q?feat:=20CCCL=20CachingDeviceAllocator=20preload?=
 =?UTF-8?q?=20=E2=80=94=20=E5=AE=8C=E6=95=B4=E4=BE=9D=E8=B5=96=E9=93=BE=20?=
 =?UTF-8?q?288=20files?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

从 cccl_upstream 递归追踪 cub/util_allocator.cuh 的全部 include 依赖:
  cub/         9 files (config, util_*, version, detect_cuda_runtime)
  cuda/        libcudacxx type_traits, concepts, algorithm, iterator...
  nv/          target macros, preprocessor

总计 288 个头文件 (1.4MB)，打包到 include/ 目录，编译时 -I include
即可完全脱离 CCCL 原始目录结构。

.cu 文件直接 #include <cub/util_allocator.cuh>，
走原版 CUB CachingDeviceAllocator，零 mock。

BI-V100 参数: growth=2 bins=[8..32] max_cached=8GB/device
---
 .../cccl_preload/build_cccl_preload.sh        |  116 +-
 .../cccl_preload/cccl_allocator_preload.cu    |  465 +--
 .../cccl_preload/include/cub/config.cuh       |   29 +
 .../cub/detail/detect_cuda_runtime.cuh        |   45 +
 .../include/cub/util_allocator.cuh            |  901 +++++
 .../cccl_preload/include/cub/util_arch.cuh    |  219 ++
 .../include/cub/util_cpp_dialect.cuh          |   71 +
 .../cccl_preload/include/cub/util_debug.cuh   |  187 ++
 .../cccl_preload/include/cub/util_macro.cuh   |   67 +
 .../include/cub/util_namespace.cuh            |  172 +
 .../cccl_preload/include/cub/version.cuh      |   65 +
 .../cccl_preload/include/cuda/__cccl_config   |   38 +
 .../include/cuda/__cmath/ceil_div.h           |  123 +
 .../include/cuda/__cmath/round_up.h           |  104 +
 .../cuda/__device/compute_capability.h        |  272 ++
 .../cccl_preload/include/cuda/__fwd/complex.h |   48 +
 .../cccl_preload/include/cuda/__fwd/devices.h |   47 +
 .../include/cuda/__memory/address_space.h     |  259 ++
 .../include/cuda/__memory/check_address.h     |  111 +
 .../cccl_preload/include/cuda/__nvtx/nvtx.h   |  150 +
 .../cccl_preload/include/cuda/__nvtx/nvtx3.h  | 2977 +++++++++++++++++
 .../__ptx/instructions/generated/get_sreg.h   |  949 ++++++
 .../cuda/__ptx/instructions/get_sreg.h        |   43 +
 .../include/cuda/__ptx/ptx_dot_variants.h     |  230 ++
 .../include/cuda/__ptx/ptx_helper_functions.h |  178 +
 .../__type_traits/is_trivially_copyable.h     |  115 +
 .../include/cuda/std/__algorithm/clamp.h      |   53 +
 .../include/cuda/std/__algorithm/comp.h       |   58 +
 .../cuda/std/__algorithm/comp_ref_type.h      |   85 +
 .../include/cuda/std/__algorithm/equal.h      |  132 +
 .../include/cuda/std/__algorithm/fill_n.h     |   51 +
 .../include/cuda/std/__algorithm/iter_swap.h  |   85 +
 .../std/__algorithm/iterator_operations.h     |  179 +
 .../std/__algorithm/lexicographical_compare.h |   70 +
 .../include/cuda/std/__algorithm/max.h        |   63 +
 .../cuda/std/__algorithm/max_element.h        |   67 +
 .../include/cuda/std/__algorithm/min.h        |   63 +
 .../cuda/std/__algorithm/min_element.h        |   87 +
 .../std/__algorithm/ranges_iterator_concept.h |   65 +
 .../cuda/std/__algorithm/swap_ranges.h        |   78 +
 .../cuda/std/__algorithm/unwrap_iter.h        |   95 +
 .../include/cuda/std/__bit/bit_cast.h         |   86 +
 .../include/cuda/std/__cccl/architecture.h    |  128 +
 .../include/cuda/std/__cccl/assert.h          |  169 +
 .../include/cuda/std/__cccl/attributes.h      |  221 ++
 .../include/cuda/std/__cccl/builtin.h         |  474 +++
 .../include/cuda/std/__cccl/compiler.h        |  238 ++
 .../cuda/std/__cccl/cuda_capabilities.h       |  118 +
 .../include/cuda/std/__cccl/cuda_toolkit.h    |   56 +
 .../include/cuda/std/__cccl/deprecated.h      |   87 +
 .../include/cuda/std/__cccl/diagnostic.h      |  145 +
 .../include/cuda/std/__cccl/dialect.h         |  230 ++
 .../include/cuda/std/__cccl/epilogue.h        |  390 +++
 .../include/cuda/std/__cccl/exceptions.h      |   42 +
 .../include/cuda/std/__cccl/execution_space.h |   87 +
 .../cuda/std/__cccl/extended_data_types.h     |  148 +
 .../include/cuda/std/__cccl/host_std_lib.h    |   61 +
 .../std/__cccl/is_non_narrowing_convertible.h |   71 +
 .../cccl_preload/include/cuda/std/__cccl/os.h |  120 +
 .../include/cuda/std/__cccl/preprocessor.h    | 1366 ++++++++
 .../include/cuda/std/__cccl/prologue.h        |  348 ++
 .../include/cuda/std/__cccl/ptx_isa.h         |  369 ++
 .../include/cuda/std/__cccl/rtti.h            |   72 +
 .../include/cuda/std/__cccl/sequence_access.h |   83 +
 .../include/cuda/std/__cccl/system_header.h   |   38 +
 .../include/cuda/std/__cccl/unreachable.h     |   31 +
 .../include/cuda/std/__cccl/version.h         |   26 +
 .../include/cuda/std/__cccl/visibility.h      |  198 ++
 .../include/cuda/std/__concepts/arithmetic.h  |   56 +
 .../include/cuda/std/__concepts/assignable.h  |   64 +
 .../cuda/std/__concepts/boolean_testable.h    |   63 +
 .../cuda/std/__concepts/class_or_enum.h       |   45 +
 .../std/__concepts/common_reference_with.h    |   69 +
 .../cuda/std/__concepts/concept_macros.h      |  389 +++
 .../cuda/std/__concepts/constructible.h       |  174 +
 .../cuda/std/__concepts/convertible_to.h      |   70 +
 .../include/cuda/std/__concepts/copyable.h    |   60 +
 .../cuda/std/__concepts/derived_from.h        |   56 +
 .../cuda/std/__concepts/destructible.h        |   76 +
 .../cuda/std/__concepts/equality_comparable.h |   98 +
 .../include/cuda/std/__concepts/invocable.h   |   80 +
 .../include/cuda/std/__concepts/movable.h     |   58 +
 .../include/cuda/std/__concepts/predicate.h   |   54 +
 .../include/cuda/std/__concepts/regular.h     |   54 +
 .../include/cuda/std/__concepts/relation.h    |   77 +
 .../include/cuda/std/__concepts/same_as.h     |   39 +
 .../include/cuda/std/__concepts/semiregular.h |   54 +
 .../include/cuda/std/__concepts/swappable.h   |  209 ++
 .../cuda/std/__concepts/totally_ordered.h     |  101 +
 .../include/cuda/std/__cstddef/byte.h         |  113 +
 .../include/cuda/std/__cstddef/types.h        |   52 +
 .../include/cuda/std/__cstring/memcpy.h       |   61 +
 .../cuda/std/__exception/exception_macros.h   |  126 +
 .../include/cuda/std/__exception/terminate.h  |   82 +
 .../cuda/std/__floating_point/format.h        |  157 +
 .../cuda/std/__floating_point/properties.h    |  229 ++
 .../cuda/std/__floating_point/storage.h       |  260 ++
 .../cuda/std/__floating_point/traits.h        |  171 +
 .../cuda/std/__functional/binary_function.h   |   64 +
 .../include/cuda/std/__functional/identity.h  |   57 +
 .../include/cuda/std/__functional/invoke.h    |  298 ++
 .../cuda/std/__functional/reference_wrapper.h |  116 +
 .../cuda/std/__functional/unary_function.h    |   63 +
 .../cuda/std/__functional/unwrap_ref.h        |   56 +
 .../cuda/std/__functional/weak_result_type.h  |  262 ++
 .../include/cuda/std/__fwd/array.h            |   68 +
 .../include/cuda/std/__fwd/complex.h          |   75 +
 .../include/cuda/std/__fwd/format.h           |  120 +
 .../cccl_preload/include/cuda/std/__fwd/fp.h  |   37 +
 .../cccl_preload/include/cuda/std/__fwd/get.h |  132 +
 .../include/cuda/std/__fwd/iterator.h         |   45 +
 .../include/cuda/std/__fwd/pair.h             |   53 +
 .../cuda/std/__fwd/reference_wrapper.h        |   52 +
 .../include/cuda/std/__fwd/span.h             |   45 +
 .../include/cuda/std/__fwd/subrange.h         |   65 +
 .../include/cuda/std/__fwd/tuple.h            |   52 +
 .../include/cuda/std/__host_stdlib/cstdio     |   28 +
 .../include/cuda/std/__host_stdlib/math.h     |   49 +
 .../include/cuda/std/__host_stdlib/memory     |   36 +
 .../include/cuda/std/__host_stdlib/new        |   29 +
 .../include/cuda/std/__host_stdlib/stdexcept  |   28 +
 .../include/cuda/std/__internal/atomic.h      |   55 +
 .../include/cuda/std/__internal/cpp_dialect.h |   44 +
 .../include/cuda/std/__internal/features.h    |  127 +
 .../include/cuda/std/__internal/namespaces.h  |  188 ++
 .../include/cuda/std/__internal/pstl_config.h |   34 +
 .../include/cuda/std/__internal/thread_api.h  |   58 +
 .../include/cuda/std/__internal/version.h     |   52 +
 .../include/cuda/std/__iterator/access.h      |  140 +
 .../include/cuda/std/__iterator/advance.h     |  230 ++
 .../include/cuda/std/__iterator/concepts.h    |  718 ++++
 .../include/cuda/std/__iterator/data.h        |   61 +
 .../include/cuda/std/__iterator/distance.h    |  130 +
 .../include/cuda/std/__iterator/empty.h       |   53 +
 .../std/__iterator/incrementable_traits.h     |  143 +
 .../include/cuda/std/__iterator/iter_move.h   |  165 +
 .../include/cuda/std/__iterator/iter_swap.h   |  185 +
 .../include/cuda/std/__iterator/iterator.h    |   44 +
 .../cuda/std/__iterator/iterator_traits.h     |  618 ++++
 .../include/cuda/std/__iterator/next.h        |  104 +
 .../include/cuda/std/__iterator/prev.h        |   93 +
 .../cuda/std/__iterator/readable_traits.h     |  156 +
 .../cuda/std/__iterator/reverse_access.h      |  154 +
 .../cuda/std/__iterator/reverse_iterator.h    |  375 +++
 .../include/cuda/std/__iterator/size.h        |   69 +
 .../cuda/std/__limits/numeric_limits.h        |  612 ++++
 .../cuda/std/__limits/numeric_limits_ext.h    |  753 +++++
 .../include/cuda/std/__memory/addressof.h     |   96 +
 .../cuda/std/__memory/allocator_arg_t.h       |   85 +
 .../cuda/std/__memory/pointer_traits.h        |  247 ++
 .../cuda/std/__memory/uses_allocator.h        |   59 +
 .../include/cuda/std/__new/device_new.h       |   30 +
 .../include/cuda/std/__ranges/access.h        |  318 ++
 .../include/cuda/std/__ranges/concepts.h      |  303 ++
 .../include/cuda/std/__ranges/data.h          |  143 +
 .../cuda/std/__ranges/enable_borrowed_range.h |   41 +
 .../include/cuda/std/__ranges/enable_view.h   |   78 +
 .../include/cuda/std/__ranges/size.h          |  212 ++
 .../cuda/std/__tuple_dir/make_tuple_types.h   |  143 +
 .../cuda/std/__tuple_dir/sfinae_helpers.h     |   94 +
 .../cuda/std/__tuple_dir/tuple_constraints.h  |  643 ++++
 .../cuda/std/__tuple_dir/tuple_element.h      |   88 +
 .../cuda/std/__tuple_dir/tuple_indices.h      |   44 +
 .../include/cuda/std/__tuple_dir/tuple_like.h |  103 +
 .../include/cuda/std/__tuple_dir/tuple_size.h |  101 +
 .../cuda/std/__tuple_dir/tuple_types.h        |   53 +
 .../cuda/std/__type_traits/add_const.h        |   40 +
 .../std/__type_traits/add_lvalue_reference.h  |   62 +
 .../cuda/std/__type_traits/add_pointer.h      |   65 +
 .../std/__type_traits/add_rvalue_reference.h  |   62 +
 .../__type_traits/aggregate_members_all_of.h  |  135 +
 .../cuda/std/__type_traits/always_false.h     |   35 +
 .../cuda/std/__type_traits/common_reference.h |  262 ++
 .../cuda/std/__type_traits/common_type.h      |  173 +
 .../cuda/std/__type_traits/conditional.h      |   65 +
 .../cuda/std/__type_traits/conjunction.h      |   67 +
 .../include/cuda/std/__type_traits/copy_cv.h  |   50 +
 .../cuda/std/__type_traits/copy_cvref.h       |  148 +
 .../include/cuda/std/__type_traits/decay.h    |   83 +
 .../cuda/std/__type_traits/disjunction.h      |   77 +
 .../cuda/std/__type_traits/enable_if.h        |   43 +
 .../include/cuda/std/__type_traits/extent.h   |   67 +
 .../include/cuda/std/__type_traits/fold.h     |   47 +
 .../std/__type_traits/integral_constant.h     |   53 +
 .../cuda/std/__type_traits/is_aggregate.h     |   42 +
 .../cuda/std/__type_traits/is_arithmetic.h    |   42 +
 .../include/cuda/std/__type_traits/is_array.h |   62 +
 .../cuda/std/__type_traits/is_assignable.h    |   78 +
 .../cuda/std/__type_traits/is_base_of.h       |   42 +
 .../cuda/std/__type_traits/is_callable.h      |   61 +
 .../include/cuda/std/__type_traits/is_class.h |   42 +
 .../cuda/std/__type_traits/is_comparable.h    |   79 +
 .../cuda/std/__type_traits/is_complete.h      |   53 +
 .../include/cuda/std/__type_traits/is_const.h |   60 +
 .../cuda/std/__type_traits/is_constructible.h |  174 +
 .../cuda/std/__type_traits/is_convertible.h   |  211 ++
 .../std/__type_traits/is_copy_assignable.h    |   60 +
 .../std/__type_traits/is_copy_constructible.h |   58 +
 .../std/__type_traits/is_core_convertible.h   |   47 +
 .../__type_traits/is_default_constructible.h  |   40 +
 .../cuda/std/__type_traits/is_destructible.h  |  115 +
 .../include/cuda/std/__type_traits/is_empty.h |   42 +
 .../include/cuda/std/__type_traits/is_enum.h  |   42 +
 .../is_extended_floating_point.h              |   84 +
 .../std/__type_traits/is_floating_point.h     |   53 +
 .../cuda/std/__type_traits/is_function.h      |   64 +
 .../is_implicitly_default_constructible.h     |   61 +
 .../cuda/std/__type_traits/is_integral.h      |  127 +
 .../is_member_function_pointer.h              |   79 +
 .../__type_traits/is_member_object_pointer.h  |   57 +
 .../std/__type_traits/is_member_pointer.h     |   57 +
 .../std/__type_traits/is_move_assignable.h    |   59 +
 .../std/__type_traits/is_move_constructible.h |   56 +
 .../std/__type_traits/is_nothrow_assignable.h |   70 +
 .../__type_traits/is_nothrow_constructible.h  |   72 +
 .../__type_traits/is_nothrow_convertible.h    |   59 +
 .../is_nothrow_copy_assignable.h              |   60 +
 .../is_nothrow_copy_constructible.h           |   58 +
 .../is_nothrow_default_constructible.h        |   54 +
 .../__type_traits/is_nothrow_destructible.h   |   82 +
 .../is_nothrow_move_assignable.h              |   60 +
 .../is_nothrow_move_constructible.h           |   57 +
 .../cuda/std/__type_traits/is_null_pointer.h  |   43 +
 .../cuda/std/__type_traits/is_object.h        |   57 +
 .../cuda/std/__type_traits/is_pointer.h       |   60 +
 .../std/__type_traits/is_primary_template.h   |  121 +
 .../cuda/std/__type_traits/is_reference.h     |   89 +
 .../cuda/std/__type_traits/is_referenceable.h |   55 +
 .../include/cuda/std/__type_traits/is_same.h  |   88 +
 .../cuda/std/__type_traits/is_scalar.h        |   60 +
 .../cuda/std/__type_traits/is_signed.h        |   76 +
 .../std/__type_traits/is_signed_integer.h     |   59 +
 .../cuda/std/__type_traits/is_swappable.h     |  211 ++
 .../std/__type_traits/is_trivially_copyable.h |   42 +
 .../std/__type_traits/is_unbounded_array.h    |   60 +
 .../include/cuda/std/__type_traits/is_union.h |   42 +
 .../cuda/std/__type_traits/is_unsigned.h      |   76 +
 .../std/__type_traits/is_unsigned_integer.h   |   59 +
 .../std/__type_traits/is_valid_expansion.h    |   41 +
 .../include/cuda/std/__type_traits/is_void.h  |   70 +
 .../cuda/std/__type_traits/is_volatile.h      |   60 +
 .../include/cuda/std/__type_traits/lazy.h     |   35 +
 .../std/__type_traits/make_const_lvalue_ref.h |   36 +
 .../cuda/std/__type_traits/make_signed.h      |  155 +
 .../cuda/std/__type_traits/make_unsigned.h    |  166 +
 .../include/cuda/std/__type_traits/nat.h      |   39 +
 .../include/cuda/std/__type_traits/negation.h |   43 +
 .../reference_constructs_from_temporary.h     |   57 +
 .../std/__type_traits/remove_all_extents.h    |   71 +
 .../cuda/std/__type_traits/remove_const.h     |   59 +
 .../cuda/std/__type_traits/remove_cv.h        |   57 +
 .../cuda/std/__type_traits/remove_cvref.h     |   57 +
 .../cuda/std/__type_traits/remove_extent.h    |   65 +
 .../cuda/std/__type_traits/remove_pointer.h   |   73 +
 .../cuda/std/__type_traits/remove_reference.h |   72 +
 .../cuda/std/__type_traits/remove_volatile.h  |   58 +
 .../cuda/std/__type_traits/sfinae_traits.h    |   63 +
 .../cuda/std/__type_traits/type_identity.h    |   40 +
 .../cuda/std/__type_traits/type_list.h        | 1098 ++++++
 .../include/cuda/std/__type_traits/type_set.h |  131 +
 .../cuda/std/__type_traits/underlying_type.h  |   52 +
 .../include/cuda/std/__type_traits/void_t.h   |   34 +
 .../include/cuda/std/__utility/auto_cast.h    |   34 +
 .../cuda/std/__utility/convert_to_integral.h  |  101 +
 .../include/cuda/std/__utility/declval.h      |   77 +
 .../include/cuda/std/__utility/exchange.h     |   46 +
 .../include/cuda/std/__utility/forward.h      |   90 +
 .../cuda/std/__utility/integer_sequence.h     |  308 ++
 .../include/cuda/std/__utility/move.h         |  140 +
 .../include/cuda/std/__utility/pair.h         |  943 ++++++
 .../cuda/std/__utility/piecewise_construct.h  |   37 +
 .../include/cuda/std/__utility/priority_tag.h |   40 +
 .../include/cuda/std/__utility/swap.h         |   64 +
 .../cuda/std/__utility/to_underlying.h        |   40 +
 .../include/cuda/std/__utility/undefined.h    |   34 +
 .../include/cuda/std/__utility/unreachable.h  |   37 +
 .../cccl_preload/include/cuda/std/array       |  553 +++
 .../cccl_preload/include/cuda/std/cassert     |   27 +
 .../cccl_preload/include/cuda/std/cfloat      |   59 +
 .../cccl_preload/include/cuda/std/climits     |   61 +
 .../cccl_preload/include/cuda/std/cstddef     |   28 +
 .../cccl_preload/include/cuda/std/cstdint     |  204 ++
 .../include/cuda/std/detail/__config          |   22 +
 .../include/cuda/std/initializer_list         |   48 +
 .../cccl_preload/include/cuda/std/limits      |   28 +
 .../cccl_preload/include/cuda/std/version     |  243 ++
 .../cccl_preload/include/cuda/version         |   16 +
 .../include/nv/detail/__preprocessor          |  169 +
 .../include/nv/detail/__target_macros         |  794 +++++
 .../cccl_preload/include/nv/target            |  250 ++
 290 files changed, 39904 insertions(+), 455 deletions(-)
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/config.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/detail/detect_cuda_runtime.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_allocator.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_arch.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_cpp_dialect.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_debug.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_macro.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/util_namespace.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cub/version.cuh
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__cccl_config
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__cmath/ceil_div.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__cmath/round_up.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__device/compute_capability.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__fwd/complex.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__fwd/devices.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__memory/address_space.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__memory/check_address.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx3.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/generated/get_sreg.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/get_sreg.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_dot_variants.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_helper_functions.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/__type_traits/is_trivially_copyable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/clamp.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp_ref_type.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/equal.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/fill_n.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iter_swap.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iterator_operations.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/lexicographical_compare.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max_element.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min_element.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/ranges_iterator_concept.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/swap_ranges.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/unwrap_iter.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__bit/bit_cast.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/architecture.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/assert.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/attributes.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/builtin.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/compiler.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_capabilities.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_toolkit.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/deprecated.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/diagnostic.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/dialect.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/epilogue.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/exceptions.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/execution_space.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/extended_data_types.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/host_std_lib.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/is_non_narrowing_convertible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/os.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/preprocessor.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/prologue.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/ptx_isa.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/rtti.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/sequence_access.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/system_header.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/unreachable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/version.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/visibility.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/arithmetic.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/boolean_testable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/class_or_enum.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/common_reference_with.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/concept_macros.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/convertible_to.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/copyable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/derived_from.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/destructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/equality_comparable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/invocable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/movable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/predicate.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/regular.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/relation.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/same_as.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/semiregular.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/swappable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/totally_ordered.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/byte.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/types.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__cstring/memcpy.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/exception_macros.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/terminate.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/format.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/properties.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/storage.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/binary_function.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/identity.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/invoke.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/reference_wrapper.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unary_function.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unwrap_ref.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/weak_result_type.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/array.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/complex.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/format.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/fp.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/get.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/iterator.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/pair.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/reference_wrapper.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/span.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/subrange.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/tuple.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/cstdio
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/math.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/memory
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/new
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/stdexcept
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/atomic.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/cpp_dialect.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/features.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/namespaces.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/pstl_config.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/thread_api.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/version.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/access.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/advance.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/concepts.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/data.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/distance.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/empty.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/incrementable_traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_move.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_swap.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator_traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/next.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/prev.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/readable_traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_access.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_iterator.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/size.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits_ext.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/addressof.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/allocator_arg_t.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/pointer_traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/uses_allocator.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__new/device_new.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/access.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/concepts.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/data.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_borrowed_range.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_view.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/size.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/make_tuple_types.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/sfinae_helpers.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_constraints.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_element.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_indices.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_like.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_size.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_types.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_const.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_lvalue_reference.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_rvalue_reference.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/aggregate_members_all_of.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/always_false.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_reference.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_type.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conditional.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conjunction.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cv.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cvref.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/decay.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/disjunction.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/enable_if.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/extent.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/fold.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/integral_constant.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_aggregate.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_arithmetic.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_array.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_base_of.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_callable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_class.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_comparable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_complete.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_const.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_convertible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_core_convertible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_default_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_destructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_empty.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_enum.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_extended_floating_point.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_floating_point.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_function.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_implicitly_default_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_integral.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_function_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_object_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_convertible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_default_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_destructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_assignable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_constructible.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_null_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_object.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_primary_template.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_reference.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_referenceable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_same.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_scalar.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed_integer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_swappable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_trivially_copyable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unbounded_array.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_union.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned_integer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_valid_expansion.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_void.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_volatile.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/lazy.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_const_lvalue_ref.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_signed.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_unsigned.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/nat.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/negation.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/reference_constructs_from_temporary.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_all_extents.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_const.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cv.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cvref.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_extent.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_pointer.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_reference.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_volatile.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/sfinae_traits.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_identity.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_list.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_set.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/underlying_type.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/void_t.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/auto_cast.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/convert_to_integral.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/declval.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/exchange.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/forward.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/integer_sequence.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/move.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/pair.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/piecewise_construct.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/priority_tag.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/swap.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/to_underlying.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/undefined.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/unreachable.h
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/array
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/cassert
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/cfloat
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/climits
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/cstddef
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/cstdint
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/detail/__config
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/initializer_list
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/limits
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/std/version
 create mode 100644 qwen3_6_scripts/cccl_preload/include/cuda/version
 create mode 100644 qwen3_6_scripts/cccl_preload/include/nv/detail/__preprocessor
 create mode 100644 qwen3_6_scripts/cccl_preload/include/nv/detail/__target_macros
 create mode 100644 qwen3_6_scripts/cccl_preload/include/nv/target

diff --git a/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh b/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
index 0d7c53c7..b97fde7c 100755
--- a/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
+++ b/qwen3_6_scripts/cccl_preload/build_cccl_preload.sh
@@ -1,41 +1,43 @@
 #!/usr/bin/env bash
-# Build libcccl_allocator.so — LD_PRELOAD .so for CUB CachingDeviceAllocator
+# Build libcccl_allocator.so
+#
+# Full CCCL dependency chain (288 headers) in ./include/
+# Source: cccl_upstream/cub/cub/util_allocator.cuh + transitive deps
 #
 # Usage:
 #   bash build_cccl_preload.sh [output_dir]
-#
-# On BI-V100 with CoreX SDK:
-#   bash build_cccl_preload.sh /workspace/qwen3_6_scripts/cccl_preload
-#
-# The .so intercepts cudaMalloc/cudaFree and routes through CUB's
-# caching allocator, bypassing CoreX's "expandable segment not supported"
-# ASSERT in CUDACachingAllocator.cpp:545.
 
 set -euo pipefail
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 OUTPUT_DIR="${1:-${SCRIPT_DIR}}"
 SRC="${SCRIPT_DIR}/cccl_allocator_preload.cu"
+INC="${SCRIPT_DIR}/include"
 OUT="${OUTPUT_DIR}/libcccl_allocator.so"
 
-# Find CoreX clang++ (preferred) or system g++
-if [[ -x /usr/local/corex-3.2.3/bin/clang++ ]]; then
-    CXX=/usr/local/corex-3.2.3/bin/clang++
-    echo "[build] Using CoreX clang++: ${CXX}"
-elif [[ -x /usr/local/corex/bin/clang++ ]]; then
-    CXX=/usr/local/corex/bin/clang++
-    echo "[build] Using CoreX clang++ (alt): ${CXX}"
-else
-    CXX=g++
-    echo "[build] CoreX clang++ not found, falling back to g++"
-fi
+[[ -d "${INC}/cub" ]] || { echo "CCCL include tree missing: ${INC}/cub"; exit 2; }
+[[ -d "${INC}/cuda" ]] || { echo "CCCL include tree missing: ${INC}/cuda"; exit 2; }
 
-# Find CUDA include path
+# Find compiler
+CXX=""
+for candidate in \
+    /usr/local/corex-3.2.3/bin/clang++ \
+    /usr/local/corex/bin/clang++ \
+    /usr/local/corex/lib64/clang/16/bin/clang++ \
+    ; do
+    if [[ -x "${candidate}" ]]; then
+        CXX="${candidate}"
+        break
+    fi
+done
+[[ -n "${CXX}" ]] || { CXX=g++; echo "[build] no CoreX clang++, falling back to g++"; }
+echo "[build] CXX=${CXX}"
+
+# Find CUDA headers (for cuda_runtime_api.h)
 CUDA_INC=""
 for candidate in \
     /usr/local/corex/include \
     /usr/local/cuda/include \
-    /usr/local/corex/lib64/clang/16/include \
     ; do
     if [[ -f "${candidate}/cuda_runtime_api.h" ]]; then
         CUDA_INC="${candidate}"
@@ -43,7 +45,7 @@ for candidate in \
     fi
 done
 
-# Find CUDA lib path for linking
+# Find CUDA libs
 CUDA_LIB=""
 for candidate in \
     /usr/local/corex/lib64 \
@@ -55,57 +57,43 @@ for candidate in \
     fi
 done
 
-if [[ -z "${CUDA_INC}" ]]; then
-    echo "[WARN] cuda_runtime_api.h not found — trying compile anyway"
-fi
-
 echo "[build] CUDA include: ${CUDA_INC:-system}"
-echo "[build] CUDA lib: ${CUDA_LIB:-system}"
-echo "[build] Source: ${SRC}"
-echo "[build] Output: ${OUT}"
+echo "[build] CUDA lib:     ${CUDA_LIB:-system}"
+echo "[build] CCCL include: ${INC} ($(find "${INC}" -type f | wc -l) files)"
+echo "[build] Source:        ${SRC}"
+echo "[build] Output:        ${OUT}"
+
+COMMON_FLAGS=(
+    -shared -fPIC -O2 -std=c++17
+    -I"${INC}"
+    ${CUDA_INC:+-I"${CUDA_INC}"}
+    ${CUDA_LIB:+-L"${CUDA_LIB}"}
+    -lcudart -ldl
+    # Suppress CCCL warnings that don't affect correctness
+    -Wno-unused-function
+    -Wno-unknown-pragmas
+    # CUB needs these for non-NVCC compilers
+    -D_CCCL_COMPILER_GCC=1
+    -D__CUDA_ARCH_LIST__=700
+    -DCUB_DISABLE_NAMESPACE_MAGIC
+    -DCUB_WRAPPED_NAMESPACE=cccl_preload
+)
 
-# Build as shared library
-# -x cuda or -x c++ depending on compiler
 if [[ "${CXX}" == *clang++* ]]; then
-    # CoreX clang++ can compile .cu natively
-    ${CXX} \
-        -shared -fPIC \
-        -O2 \
-        ${CUDA_INC:+-I"${CUDA_INC}"} \
-        ${CUDA_LIB:+-L"${CUDA_LIB}"} \
-        -lcudart \
-        -ldl \
-        -std=c++17 \
-        -o "${OUT}" \
-        "${SRC}"
+    "${CXX}" "${COMMON_FLAGS[@]}" -x c++ -o "${OUT}" "${SRC}" 2>&1
 else
-    # g++ needs .cu renamed or treated as C++
-    # cuda_runtime_api.h should still work with host compiler
-    ${CXX} \
-        -shared -fPIC \
-        -O2 \
-        ${CUDA_INC:+-I"${CUDA_INC}"} \
-        ${CUDA_LIB:+-L"${CUDA_LIB}"} \
-        -lcudart \
-        -ldl \
-        -std=c++17 \
-        -x c++ \
-        -o "${OUT}" \
-        "${SRC}"
+    "${CXX}" "${COMMON_FLAGS[@]}" -x c++ -o "${OUT}" "${SRC}" 2>&1
 fi
 
 if [[ -f "${OUT}" ]]; then
-    SIZE=$(stat -c%s "${OUT}" 2>/dev/null || stat -f%z "${OUT}" 2>/dev/null || echo "?")
+    SIZE=$(stat -c%s "${OUT}" 2>/dev/null || echo "?")
+    echo ""
     echo "[build] SUCCESS: ${OUT} (${SIZE} bytes)"
     echo ""
-    echo "Usage:"
-    echo "  LD_PRELOAD=${OUT} CCCL_ALLOC_DEBUG=1 python3 -c 'import torch; t=torch.zeros(1024, device=\"cuda\")'"
-    echo ""
-    echo "In computility-run.yaml, add to env:"
-    echo "  - name: LD_PRELOAD"
-    echo "    value: /workspace/qwen3_6_scripts/cccl_preload/libcccl_allocator.so"
-    echo "  - name: PYTORCH_CUDA_ALLOC_CONF"
-    echo "    value: expandable_segments:True"
+    echo "Test:"
+    echo "  LD_PRELOAD=${OUT} CCCL_ALLOC_DEBUG=1 \\"
+    echo "  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \\"
+    echo "  python3 verify_preload.py"
 else
     echo "[build] FAILED"
     exit 1
diff --git a/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu b/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
index 2812d2ae..50061632 100644
--- a/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
+++ b/qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu
@@ -1,409 +1,90 @@
 /*
  * cccl_allocator_preload.cu
  *
- * LD_PRELOAD .so that replaces PyTorch's CUDA memory allocator with
- * CUB's CachingDeviceAllocator (extracted from CCCL upstream).
+ * LD_PRELOAD .so — CUB CachingDeviceAllocator from CCCL upstream.
+ * Full dependency chain (288 files) extracted into include/.
  *
- * Purpose: CoreX's CUDACachingAllocator.cpp:545 asserts
- * "expandable segment not supported". Instead of patching libtorch,
- * we intercept cudaMalloc/cudaFree at the dynamic linker level and
- * route them through CUB's battle-tested caching allocator.
+ * Intercepts cudaMalloc/cudaFree, routes through CUB's geometric-bin
+ * caching allocator. Strips expandable_segments from
+ * PYTORCH_CUDA_ALLOC_CONF before libtorch reads it.
  *
- * Source: cccl_upstream/cub/cub/util_allocator.cuh
- * License: BSD-3 (NVIDIA/CUB)
- *
- * Build (on BI-V100 with CoreX clang++):
- *   bash build_cccl_preload.sh
- *
- * Usage:
- *   LD_PRELOAD=/workspace/qwen3_6_scripts/cccl_preload/libcccl_allocator.so \
- *   CCCL_ALLOC_DEBUG=0 \
- *   PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512 \
- *   python3 -m vllm.entrypoints.openai.api_server ...
+ * Source: CCCL cub/cub/util_allocator.cuh (BSD-3, NVIDIA)
+ * Build:  bash build_cccl_preload.sh
  */
 
-#include <cuda_runtime_api.h>
+/* ---- CCCL include chain (288 files from cccl_upstream) ---- */
+#include <cub/util_allocator.cuh>
+
+/* ---- System ---- */
 #include <dlfcn.h>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
-#include <map>
-#include <mutex>
-#include <set>
+#include <string>
 
 /* ========================================================================
- * CUB CachingDeviceAllocator — extracted from CCCL
- * cccl_upstream/cub/cub/util_allocator.cuh
+ * Configuration for BI-V100 (32GB × 4 cards)
  *
- * All CUB/CCCL macro dependencies replaced with plain C++.
+ * CUB CachingDeviceAllocator parameters:
+ *   bin_growth = 2   (power-of-2 bins: 256B, 512B, 1KB, ... 4GB)
+ *   min_bin    = 8   (2^8 = 256B minimum allocation)
+ *   max_bin    = 32  (2^32 = 4GB maximum cached bin)
+ *   max_cached = 8GB per device
+ *
+ * More granular bins (growth=2) than CUB default (growth=8) because
+ * PyTorch tensor sizes vary widely in inference.
  * ======================================================================== */
 
-static bool g_cccl_debug = false;
+static constexpr unsigned int ALLOC_BIN_GROWTH   = 2;
+static constexpr unsigned int ALLOC_MIN_BIN      = 8;   /* 256 bytes */
+static constexpr unsigned int ALLOC_MAX_BIN      = 32;  /* 4 GB */
+static constexpr size_t       ALLOC_MAX_CACHED   = (size_t)8 * 1024 * 1024 * 1024; /* 8GB */
 
-#define CcclDebug(e) (e)
-#define CcclLog(...)               \
-    do {                           \
-        if (g_cccl_debug) {        \
-            fprintf(stderr, "[cccl_alloc] "); \
-            fprintf(stderr, __VA_ARGS__); \
-        }                          \
-    } while (0)
-
-struct CachingDeviceAllocator
-{
-    static constexpr unsigned int INVALID_BIN = (unsigned int) -1;
-    static constexpr size_t INVALID_SIZE = (size_t) -1;
-    static constexpr int INVALID_DEVICE_ORDINAL = -1;
-
-    struct BlockDescriptor
-    {
-        void* d_ptr;
-        size_t bytes;
-        unsigned int bin;
-        int device;
-        cudaStream_t associated_stream;
-        cudaEvent_t ready_event;
-
-        BlockDescriptor(void* d_ptr_, int device_)
-            : d_ptr(d_ptr_), bytes(0), bin(INVALID_BIN), device(device_),
-              associated_stream(nullptr), ready_event(nullptr) {}
-
-        BlockDescriptor(int device_)
-            : d_ptr(nullptr), bytes(0), bin(INVALID_BIN), device(device_),
-              associated_stream(nullptr), ready_event(nullptr) {}
-
-        static bool PtrCompare(const BlockDescriptor& a, const BlockDescriptor& b) {
-            return (a.device == b.device) ? (a.d_ptr < b.d_ptr) : (a.device < b.device);
-        }
-        static bool SizeCompare(const BlockDescriptor& a, const BlockDescriptor& b) {
-            return (a.device == b.device) ? (a.bytes < b.bytes) : (a.device < b.device);
-        }
-    };
-
-    using Compare = bool (*)(const BlockDescriptor&, const BlockDescriptor&);
-
-    struct TotalBytes { size_t free; size_t live; TotalBytes() : free(0), live(0) {} };
-
-    using CachedBlocks = std::multiset<BlockDescriptor, Compare>;
-    using BusyBlocks = std::multiset<BlockDescriptor, Compare>;
-    using GpuCachedBytes = std::map<int, TotalBytes>;
-
-    static unsigned int IntPow(unsigned int base, unsigned int exp) {
-        unsigned int retval = 1;
-        while (exp > 0) {
-            if (exp & 1) retval *= base;
-            base *= base;
-            exp >>= 1;
-        }
-        return retval;
-    }
-
-    void NearestPowerOf(unsigned int& power, size_t& rounded_bytes,
-                        unsigned int base, size_t value) {
-        power = 0;
-        rounded_bytes = 1;
-        if (value * base < value) {
-            power = sizeof(size_t) * 8;
-            rounded_bytes = size_t(0) - 1;
-            return;
-        }
-        while (rounded_bytes < value) {
-            rounded_bytes *= base;
-            power++;
-        }
-    }
-
-    std::mutex mutex;
-    unsigned int bin_growth;
-    unsigned int min_bin;
-    unsigned int max_bin;
-    size_t min_bin_bytes;
-    size_t max_bin_bytes;
-    size_t max_cached_bytes;
-    bool skip_cleanup;
-    GpuCachedBytes cached_bytes;
-    CachedBlocks cached_blocks;
-    BusyBlocks live_blocks;
-
-    /*
-     * Constructor tuned for BI-V100 (32GB per card, 4 cards):
-     *   bin_growth=8, min_bin=3 (512B), max_bin=13 (~550MB)
-     *   max_cached_bytes = 4GB per device (reasonable for 32GB card)
-     *
-     * This replaces PyTorch's expandable_segments with a proven
-     * geometric-bin caching strategy from CUB/CCCL.
-     */
-    CachingDeviceAllocator()
-        : bin_growth(8)
-        , min_bin(3)           /* 8^3 = 512B minimum allocation */
-        , max_bin(13)          /* 8^13 = ~550MB maximum cached bin */
-        , min_bin_bytes(IntPow(8, 3))
-        , max_bin_bytes(IntPow(8, 13))
-        , max_cached_bytes((size_t)4 * 1024 * 1024 * 1024)  /* 4GB per device */
-        , skip_cleanup(true)   /* CoreX may tear down CUDA before our dtor */
-        , cached_blocks(BlockDescriptor::SizeCompare)
-        , live_blocks(BlockDescriptor::PtrCompare)
-    {
-        CcclLog("CachingDeviceAllocator init: bin_growth=%u min_bin=%u "
-                "max_bin=%u max_cached=%.1fGB\n",
-                bin_growth, min_bin, max_bin,
-                (double)max_cached_bytes / (1024.0*1024.0*1024.0));
-    }
-
-    /* ---- Real cudaMalloc/cudaFree via dlsym(RTLD_NEXT) ---- */
-    using RealMalloc_t = cudaError_t (*)(void**, size_t);
-    using RealFree_t = cudaError_t (*)(void*);
-
-    static RealMalloc_t get_real_malloc() {
-        static RealMalloc_t fn = (RealMalloc_t)dlsym(RTLD_NEXT, "cudaMalloc");
-        return fn;
-    }
-    static RealFree_t get_real_free() {
-        static RealFree_t fn = (RealFree_t)dlsym(RTLD_NEXT, "cudaFree");
-        return fn;
-    }
-
-    cudaError_t DeviceAllocate(int device, void** d_ptr, size_t bytes,
-                               cudaStream_t active_stream = nullptr)
-    {
-        *d_ptr = nullptr;
-        int entrypoint_device = INVALID_DEVICE_ORDINAL;
-        cudaError_t error = cudaSuccess;
-
-        if (device == INVALID_DEVICE_ORDINAL) {
-            error = cudaGetDevice(&entrypoint_device);
-            if (error != cudaSuccess) return error;
-            device = entrypoint_device;
-        }
-
-        bool found = false;
-        BlockDescriptor search_key(device);
-        search_key.associated_stream = active_stream;
-        NearestPowerOf(search_key.bin, search_key.bytes, bin_growth, bytes);
-
-        if (search_key.bin > max_bin) {
-            search_key.bin = INVALID_BIN;
-            search_key.bytes = bytes;
-        } else {
-            mutex.lock();
-            if (search_key.bin < min_bin) {
-                search_key.bin = min_bin;
-                search_key.bytes = min_bin_bytes;
-            }
-
-            CachedBlocks::iterator block_itr = cached_blocks.lower_bound(search_key);
-            while ((block_itr != cached_blocks.end()) &&
-                   (block_itr->device == device) &&
-                   (block_itr->bin == search_key.bin))
-            {
-                bool is_reusable = false;
-                if (active_stream == block_itr->associated_stream) {
-                    is_reusable = true;
-                } else {
-                    cudaError_t event_status = cudaEventQuery(block_itr->ready_event);
-                    if (event_status != cudaErrorNotReady) {
-                        is_reusable = true;
-                    }
-                }
-
-                if (is_reusable) {
-                    found = true;
-                    search_key = *block_itr;
-                    search_key.associated_stream = active_stream;
-                    live_blocks.insert(search_key);
-                    cached_bytes[device].free -= search_key.bytes;
-                    cached_bytes[device].live += search_key.bytes;
-
-                    CcclLog("reuse %p (%zu bytes) dev=%d\n",
-                            search_key.d_ptr, search_key.bytes, device);
-                    cached_blocks.erase(block_itr);
-                    break;
-                }
-                block_itr++;
-            }
-            mutex.unlock();
-        }
-
-        if (!found) {
-            if (device != entrypoint_device) {
-                if (entrypoint_device == INVALID_DEVICE_ORDINAL)
-                    cudaGetDevice(&entrypoint_device);
-                cudaSetDevice(device);
-            }
-
-            /* Use real cudaMalloc, not ourselves */
-            error = get_real_malloc()(&search_key.d_ptr, search_key.bytes);
-
-            if (error == cudaErrorMemoryAllocation) {
-                CcclLog("OOM for %zu bytes on dev=%d, freeing cache...\n",
-                        search_key.bytes, device);
-                cudaGetLastError();  /* reset */
-
-                mutex.lock();
-                BlockDescriptor free_key(device);
-                CachedBlocks::iterator block_itr = cached_blocks.lower_bound(free_key);
-                while ((block_itr != cached_blocks.end()) &&
-                       (block_itr->device == device))
-                {
-                    error = get_real_free()(block_itr->d_ptr);
-                    if (error != cudaSuccess) break;
-                    cudaEventDestroy(block_itr->ready_event);
-                    cached_bytes[device].free -= block_itr->bytes;
-                    block_itr = cached_blocks.erase(block_itr);
-                }
-                mutex.unlock();
-
-                if (error != cudaSuccess) return error;
-                error = get_real_malloc()(&search_key.d_ptr, search_key.bytes);
-                if (error != cudaSuccess) return error;
-            } else if (error != cudaSuccess) {
-                return error;
-            }
-
-            cudaEventCreateWithFlags(&search_key.ready_event, cudaEventDisableTiming);
-
-            mutex.lock();
-            live_blocks.insert(search_key);
-            cached_bytes[device].live += search_key.bytes;
-            mutex.unlock();
-
-            CcclLog("alloc %p (%zu bytes, bin=%u) dev=%d\n",
-                    search_key.d_ptr, search_key.bytes, search_key.bin, device);
-
-            if ((entrypoint_device != INVALID_DEVICE_ORDINAL) &&
-                (entrypoint_device != device))
-                cudaSetDevice(entrypoint_device);
-        }
-
-        *d_ptr = search_key.d_ptr;
-        return cudaSuccess;
-    }
-
-    cudaError_t DeviceAllocate(void** d_ptr, size_t bytes,
-                               cudaStream_t active_stream = nullptr) {
-        return DeviceAllocate(INVALID_DEVICE_ORDINAL, d_ptr, bytes, active_stream);
-    }
-
-    cudaError_t DeviceFree(int device, void* d_ptr)
-    {
-        int entrypoint_device = INVALID_DEVICE_ORDINAL;
-        cudaError_t error = cudaSuccess;
-
-        if (d_ptr == nullptr) return cudaSuccess;
-
-        if (device == INVALID_DEVICE_ORDINAL) {
-            error = cudaGetDevice(&entrypoint_device);
-            if (error != cudaSuccess) return error;
-            device = entrypoint_device;
-        }
-
-        mutex.lock();
-        bool recached = false;
-        BlockDescriptor search_key(d_ptr, device);
-        BusyBlocks::iterator block_itr = live_blocks.find(search_key);
-
-        if (block_itr != live_blocks.end()) {
-            search_key = *block_itr;
-            live_blocks.erase(block_itr);
-            cached_bytes[device].live -= search_key.bytes;
-
-            if ((search_key.bin != INVALID_BIN) &&
-                (cached_bytes[device].free + search_key.bytes <= max_cached_bytes))
-            {
-                recached = true;
-                cached_blocks.insert(search_key);
-                cached_bytes[device].free += search_key.bytes;
-                CcclLog("cache %p (%zu bytes) dev=%d\n",
-                        d_ptr, search_key.bytes, device);
-            }
-        }
-        mutex.unlock();
-
-        if (device != entrypoint_device) {
-            if (entrypoint_device == INVALID_DEVICE_ORDINAL)
-                cudaGetDevice(&entrypoint_device);
-            cudaSetDevice(device);
-        }
-
-        if (recached) {
-            cudaEventRecord(search_key.ready_event, search_key.associated_stream);
-        } else {
-            /* Not tracked or cache full — real free */
-            CcclLog("free %p dev=%d (not cached)\n", d_ptr, device);
-            error = get_real_free()(d_ptr);
-            if (block_itr != live_blocks.end())
-                cudaEventDestroy(search_key.ready_event);
-        }
-
-        if ((entrypoint_device != INVALID_DEVICE_ORDINAL) &&
-            (entrypoint_device != device))
-            cudaSetDevice(entrypoint_device);
-
-        return error;
-    }
-
-    cudaError_t DeviceFree(void* d_ptr) {
-        return DeviceFree(INVALID_DEVICE_ORDINAL, d_ptr);
-    }
-
-    cudaError_t FreeAllCached()
-    {
-        cudaError_t error = cudaSuccess;
-        int entrypoint_device = INVALID_DEVICE_ORDINAL;
-        int current_device = INVALID_DEVICE_ORDINAL;
-
-        mutex.lock();
-        while (!cached_blocks.empty()) {
-            CachedBlocks::iterator begin = cached_blocks.begin();
-            if (entrypoint_device == INVALID_DEVICE_ORDINAL)
-                cudaGetDevice(&entrypoint_device);
-            if (begin->device != current_device) {
-                cudaSetDevice(begin->device);
-                current_device = begin->device;
-            }
-            get_real_free()(begin->d_ptr);
-            cudaEventDestroy(begin->ready_event);
-            cached_bytes[current_device].free -= begin->bytes;
-            cached_blocks.erase(begin);
-        }
-        mutex.unlock();
-
-        if (entrypoint_device != INVALID_DEVICE_ORDINAL)
-            cudaSetDevice(entrypoint_device);
-        return error;
-    }
-};
-
-/* ========================================================================
- * Global singleton + LD_PRELOAD intercepts
- * ======================================================================== */
-
-static CachingDeviceAllocator& get_allocator() {
-    static CachingDeviceAllocator instance;
+/* ---- Global allocator singleton ---- */
+static cub::CachingDeviceAllocator& get_allocator() {
+    static cub::CachingDeviceAllocator instance(
+        ALLOC_BIN_GROWTH,
+        ALLOC_MIN_BIN,
+        ALLOC_MAX_BIN,
+        ALLOC_MAX_CACHED,
+        true  /* skip_cleanup: CoreX may tear down CUDA before our dtor */
+    );
     return instance;
 }
 
 static bool g_preload_active = false;
+static bool g_debug = false;
 
-/* Called once at .so load time */
+/* ---- Real cudaMalloc/cudaFree via dlsym(RTLD_NEXT) ---- */
+using RealMalloc_t = cudaError_t (*)(void**, size_t);
+using RealFree_t   = cudaError_t (*)(void*);
+
+static RealMalloc_t get_real_malloc() {
+    static RealMalloc_t fn = (RealMalloc_t)dlsym(RTLD_NEXT, "cudaMalloc");
+    return fn;
+}
+static RealFree_t get_real_free() {
+    static RealFree_t fn = (RealFree_t)dlsym(RTLD_NEXT, "cudaFree");
+    return fn;
+}
+
+/* ========================================================================
+ * Constructor: runs at LD_PRELOAD load time
+ * ======================================================================== */
 __attribute__((constructor))
 static void cccl_preload_init() {
     const char* debug_env = getenv("CCCL_ALLOC_DEBUG");
-    g_cccl_debug = (debug_env && atoi(debug_env) > 0);
+    g_debug = (debug_env && atoi(debug_env) > 0);
 
     const char* disable_env = getenv("CCCL_ALLOC_DISABLE");
     if (disable_env && atoi(disable_env) > 0) {
         fprintf(stderr, "[cccl_alloc] DISABLED by CCCL_ALLOC_DISABLE=1\n");
-        g_preload_active = false;
         return;
     }
 
-    /* Strip expandable_segments from PYTORCH_CUDA_ALLOC_CONF
-     * so CoreX's allocator doesn't hit the assert.
-     * We handle the caching ourselves. */
+    /* Strip expandable_segments from PYTORCH_CUDA_ALLOC_CONF */
     const char* alloc_conf = getenv("PYTORCH_CUDA_ALLOC_CONF");
     if (alloc_conf) {
-        /* Build a new conf string without expandable_segments */
         std::string conf(alloc_conf);
         std::string clean;
         size_t pos = 0;
@@ -411,49 +92,51 @@ static void cccl_preload_init() {
             size_t comma = conf.find(',', pos);
             if (comma == std::string::npos) comma = conf.size();
             std::string token = conf.substr(pos, comma - pos);
-            /* Skip expandable_segments:* */
             if (token.find("expandable_segments") == std::string::npos) {
                 if (!clean.empty()) clean += ",";
                 clean += token;
             }
             pos = comma + 1;
         }
-        if (clean.empty()) {
+        if (clean.empty())
             unsetenv("PYTORCH_CUDA_ALLOC_CONF");
-        } else {
+        else
             setenv("PYTORCH_CUDA_ALLOC_CONF", clean.c_str(), 1);
-        }
-        fprintf(stderr, "[cccl_alloc] stripped expandable_segments from "
-                "PYTORCH_CUDA_ALLOC_CONF: \"%s\" -> \"%s\"\n",
+
+        fprintf(stderr, "[cccl_alloc] PYTORCH_CUDA_ALLOC_CONF: \"%s\" -> \"%s\"\n",
                 alloc_conf, clean.empty() ? "(unset)" : clean.c_str());
     }
 
-    /* Force-initialize the allocator singleton */
-    (void)get_allocator();
+    /* Initialize allocator */
+    auto& alloc = get_allocator();
+    if (g_debug) {
+        alloc.debug = true;
+    }
+
     g_preload_active = true;
-    fprintf(stderr, "[cccl_alloc] LD_PRELOAD active — CUB CachingDeviceAllocator "
-            "replacing cudaMalloc/cudaFree\n");
+    fprintf(stderr,
+        "[cccl_alloc] LD_PRELOAD active — CUB CachingDeviceAllocator "
+        "(growth=%u, bins=[%u..%u], max_cached=%.1fGB)\n",
+        ALLOC_BIN_GROWTH, ALLOC_MIN_BIN, ALLOC_MAX_BIN,
+        (double)ALLOC_MAX_CACHED / (1024.0*1024.0*1024.0));
 }
 
-/* ---- cudaMalloc intercept ---- */
+/* ========================================================================
+ * cudaMalloc / cudaFree intercepts
+ * ======================================================================== */
+
 extern "C" cudaError_t cudaMalloc(void** devPtr, size_t size)
 {
     if (!g_preload_active) {
-        /* Fallback to real cudaMalloc during init or if disabled */
-        static auto real_fn = (CachingDeviceAllocator::RealMalloc_t)
-            dlsym(RTLD_NEXT, "cudaMalloc");
-        return real_fn(devPtr, size);
+        return get_real_malloc()(devPtr, size);
     }
     return get_allocator().DeviceAllocate(devPtr, size);
 }
 
-/* ---- cudaFree intercept ---- */
 extern "C" cudaError_t cudaFree(void* devPtr)
 {
     if (!g_preload_active || devPtr == nullptr) {
-        static auto real_fn = (CachingDeviceAllocator::RealFree_t)
-            dlsym(RTLD_NEXT, "cudaFree");
-        return real_fn(devPtr);
+        return get_real_free()(devPtr);
     }
     return get_allocator().DeviceFree(devPtr);
 }
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/config.cuh b/qwen3_6_scripts/cccl_preload/include/cub/config.cuh
new file mode 100644
index 00000000..a3bc8443
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/config.cuh
@@ -0,0 +1,29 @@
+// SPDX-FileCopyrightText: Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/**
+ * \file
+ * Static configuration header for the CUB project.
+ */
+
+#pragma once
+
+// For _CCCL_IMPLICIT_SYSTEM_HEADER
+#include <cuda/__cccl_config> // IWYU pragma: export
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/util_arch.cuh> // IWYU pragma: export
+#include <cub/util_cpp_dialect.cuh> // IWYU pragma: export
+#include <cub/util_macro.cuh> // IWYU pragma: export
+#include <cub/util_namespace.cuh> // IWYU pragma: export
+
+#if !_CCCL_COMPILER(NVRTC)
+#  include <cuda/__nvtx/nvtx.h>
+#endif // !_CCCL_COMPILER(NVRTC)
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/detail/detect_cuda_runtime.cuh b/qwen3_6_scripts/cccl_preload/include/cub/detail/detect_cuda_runtime.cuh
new file mode 100644
index 00000000..ef53fe85
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/detail/detect_cuda_runtime.cuh
@@ -0,0 +1,45 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2021, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/**
+ * @file
+ * Utilities for CUDA dynamic parallelism.
+ */
+
+#pragma once
+
+// We cannot use `cub/config.cuh` here due to circular dependencies
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#ifdef _CCCL_DOXYGEN_INVOKED // Only parse this during doxygen passes:
+//! Defined if RDC is enabled and CUB_DISABLE_CDP is not defined.
+//! Deprecated [Since 3.2]
+#  define CUB_RDC_ENABLED
+
+//! If defined, support for device-side usage of CUB is disabled.
+//! Deprecated [Since 3.2]. Use CCCL_DISABLE_CDP instead.
+#  define CUB_DISABLE_CDP
+
+//! Execution space for functions that use the CUDA runtime API, e.g. to launch kernels. Such functions are `__host__
+//! __device__` when compiling with RDC, otherwise only `__host__`.
+//! Deprecated [Since 3.2]
+#  define CUB_RUNTIME_FUNCTION
+#else // Non-doxygen pass:
+
+#  if _CCCL_HAS_CDP()
+#    define CUB_RDC_ENABLED
+#  endif // _CCCL_HAS_CDP()
+
+#  ifndef CUB_RUNTIME_FUNCTION
+#    define CUB_RUNTIME_FUNCTION _CCCL_CDP_API
+#  endif // CUB_RUNTIME_FUNCTION predefined
+#endif // Do not document
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_allocator.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_allocator.cuh
new file mode 100644
index 00000000..6b2d26d4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_allocator.cuh
@@ -0,0 +1,901 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2018, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/******************************************************************************
+ * Simple caching allocator for device memory allocations. The allocator is
+ * thread-safe and capable of managing device allocations on multiple devices.
+ ******************************************************************************/
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#ifndef CCCL_DISABLE_NVRTC_COMPATIBILITY_CHECK
+#  if _CCCL_COMPILER(NVRTC)
+#    error \
+      "Including <cub/util_allocator.cuh> is not supported when compiling with NVRTC, which supports device code only. You can define CCCL_DISABLE_NVRTC_COMPATIBILITY_CHECK to disable this warning."
+#  endif // _CCCL_COMPILER(NVRTC)
+#endif // CCCL_DISABLE_NVRTC_COMPATIBILITY_CHECK
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/util_debug.cuh>
+#include <cub/util_namespace.cuh>
+
+#include <cuda/std/__host_stdlib/math.h>
+
+#include <map>
+#include <mutex>
+#include <set>
+
+CUB_NAMESPACE_BEGIN
+
+/******************************************************************************
+ * CachingDeviceAllocator (host use)
+ ******************************************************************************/
+
+/**
+ * @brief A simple caching allocator for device memory allocations.
+ *
+ * @par Overview
+ * The allocator is thread-safe and stream-safe and is capable of managing cached
+ * device allocations on multiple devices.  It behaves as follows:
+ *
+ * @par
+ * - Allocations from the allocator are associated with an @p active_stream. Once freed,
+ *   the allocation becomes available immediately for reuse within the @p active_stream
+ *   with which it was associated with during allocation, and it becomes available for
+ *   reuse within other streams when all prior work submitted to @p active_stream has completed.
+ * - Allocations are categorized and cached by bin size. A new allocation request of
+ *   a given size will only consider cached allocations within the corresponding bin.
+ * - Bin limits progress geometrically in accordance with the growth factor
+ *   @p bin_growth provided during construction. Unused device allocations within
+ *   a larger bin cache are not reused for allocation requests that categorize to
+ *   smaller bin sizes.
+ * - Allocation requests below ( @p bin_growth ^ @p min_bin ) are rounded up to
+ *   ( @p bin_growth ^ @p min_bin ).
+ * - Allocations above ( @p bin_growth ^ @p max_bin ) are not rounded up to the nearest
+ *   bin and are simply freed when they are deallocated instead of being returned
+ *   to a bin-cache.
+ * - If the total storage of cached allocations on a given device will exceed
+ *   @p max_cached_bytes, allocations for that device are simply freed when they are
+ *   deallocated instead of being returned to their bin-cache.
+ *
+ * @par
+ * For example, the default-constructed CachingDeviceAllocator is configured with:
+ * - @p bin_growth          = 8
+ * - @p min_bin             = 3
+ * - @p max_bin             = 7
+ * - @p max_cached_bytes    = 6MB - 1B
+ *
+ * @par
+ * which delineates five bin-sizes: 512B, 4KB, 32KB, 256KB, and 2MB
+ * and sets a maximum of 6,291,455 cached bytes per device
+ *
+ */
+struct CachingDeviceAllocator
+{
+  //---------------------------------------------------------------------
+  // Constants
+  //---------------------------------------------------------------------
+
+  /// Out-of-bounds bin
+  static constexpr unsigned int INVALID_BIN = (unsigned int) -1;
+
+  /// Invalid size
+  static constexpr size_t INVALID_SIZE = (size_t) -1;
+
+#ifndef _CCCL_DOXYGEN_INVOKED // Do not document
+
+  /// Invalid device ordinal
+  static constexpr int INVALID_DEVICE_ORDINAL = -1;
+
+  //---------------------------------------------------------------------
+  // Type definitions and helper types
+  //---------------------------------------------------------------------
+
+  /**
+   * Descriptor for device memory allocations
+   */
+  struct BlockDescriptor
+  {
+    // Device pointer
+    void* d_ptr;
+
+    // Size of allocation in bytes
+    size_t bytes;
+
+    // Bin enumeration
+    unsigned int bin;
+
+    // device ordinal
+    int device;
+
+    // Associated associated_stream
+    cudaStream_t associated_stream;
+
+    // Signal when associated stream has run to the point at which this block was freed
+    cudaEvent_t ready_event;
+
+    // Constructor (suitable for searching maps for a specific block, given its pointer and
+    // device)
+    BlockDescriptor(void* d_ptr, int device)
+        : d_ptr(d_ptr)
+        , bytes(0)
+        , bin(INVALID_BIN)
+        , device(device)
+        , associated_stream(nullptr)
+        , ready_event(nullptr)
+    {}
+
+    // Constructor (suitable for searching maps for a range of suitable blocks, given a device)
+    BlockDescriptor(int device)
+        : d_ptr(nullptr)
+        , bytes(0)
+        , bin(INVALID_BIN)
+        , device(device)
+        , associated_stream(nullptr)
+        , ready_event(nullptr)
+    {}
+
+    // Comparison functor for comparing device pointers
+    static bool PtrCompare(const BlockDescriptor& a, const BlockDescriptor& b)
+    {
+      if (a.device == b.device)
+      {
+        return (a.d_ptr < b.d_ptr);
+      }
+      else
+      {
+        return (a.device < b.device);
+      }
+    }
+
+    // Comparison functor for comparing allocation sizes
+    static bool SizeCompare(const BlockDescriptor& a, const BlockDescriptor& b)
+    {
+      if (a.device == b.device)
+      {
+        return (a.bytes < b.bytes);
+      }
+      else
+      {
+        return (a.device < b.device);
+      }
+    }
+  };
+
+  /// BlockDescriptor comparator function interface
+  using Compare = bool (*)(const BlockDescriptor&, const BlockDescriptor&);
+
+  class TotalBytes
+  {
+  public:
+    size_t free;
+    size_t live;
+    TotalBytes()
+    {
+      free = live = 0;
+    }
+  };
+
+  /// Set type for cached blocks (ordered by size)
+  using CachedBlocks = std::multiset<BlockDescriptor, Compare>;
+
+  /// Set type for live blocks (ordered by ptr)
+  using BusyBlocks = std::multiset<BlockDescriptor, Compare>;
+
+  /// Map type of device ordinals to the number of cached bytes cached by each device
+  using GpuCachedBytes = std::map<int, TotalBytes>;
+
+  //---------------------------------------------------------------------
+  // Utility functions
+  //---------------------------------------------------------------------
+
+  /**
+   * Integer pow function for unsigned base and exponent
+   */
+  static unsigned int IntPow(unsigned int base, unsigned int exp)
+  {
+    unsigned int retval = 1;
+    while (exp > 0)
+    {
+      if (exp & 1)
+      {
+        retval = retval * base; // multiply the result by the current base
+      }
+      base = base * base; // square the base
+      exp  = exp >> 1; // divide the exponent in half
+    }
+    return retval;
+  }
+
+  /**
+   * Round up to the nearest power-of
+   */
+  void NearestPowerOf(unsigned int& power, size_t& rounded_bytes, unsigned int base, size_t value)
+  {
+    power         = 0;
+    rounded_bytes = 1;
+
+    if (value * base < value)
+    {
+      // Overflow
+      power         = sizeof(size_t) * 8;
+      rounded_bytes = size_t(0) - 1;
+      return;
+    }
+
+    while (rounded_bytes < value)
+    {
+      rounded_bytes *= base;
+      power++;
+    }
+  }
+
+  //---------------------------------------------------------------------
+  // Fields
+  //---------------------------------------------------------------------
+
+  /// Mutex for thread-safety
+  std::mutex mutex;
+
+  /// Geometric growth factor for bin-sizes
+  unsigned int bin_growth;
+
+  /// Minimum bin enumeration
+  unsigned int min_bin;
+
+  /// Maximum bin enumeration
+  unsigned int max_bin;
+
+  /// Minimum bin size
+  size_t min_bin_bytes;
+
+  /// Maximum bin size
+  size_t max_bin_bytes;
+
+  /// Maximum aggregate cached bytes per device
+  size_t max_cached_bytes;
+
+  /// Whether or not to skip a call to FreeAllCached() when destructor is called.
+  /// (The CUDA runtime may have already shut down for statically declared allocators)
+  const bool skip_cleanup;
+
+  /// Whether or not to print (de)allocation events to stdout
+  bool debug;
+
+  /// Map of device ordinal to aggregate cached bytes on that device
+  GpuCachedBytes cached_bytes;
+
+  /// Set of cached device allocations available for reuse
+  CachedBlocks cached_blocks;
+
+  /// Set of live device allocations currently in use
+  BusyBlocks live_blocks;
+
+#endif // _CCCL_DOXYGEN_INVOKED
+
+  //---------------------------------------------------------------------
+  // Methods
+  //---------------------------------------------------------------------
+
+  /**
+   * @brief Constructor.
+   *
+   * @param bin_growth
+   *   Geometric growth factor for bin-sizes
+   *
+   * @param min_bin
+   *   Minimum bin (default is bin_growth ^ 1)
+   *
+   * @param max_bin
+   *   Maximum bin (default is no max bin)
+   *
+   * @param max_cached_bytes
+   *   Maximum aggregate cached bytes per device (default is no limit)
+   *
+   * @param skip_cleanup
+   *   Whether or not to skip a call to @p FreeAllCached() when the destructor is called (default
+   *   is to deallocate)
+   */
+  CachingDeviceAllocator(
+    unsigned int bin_growth,
+    unsigned int min_bin    = 1,
+    unsigned int max_bin    = INVALID_BIN,
+    size_t max_cached_bytes = INVALID_SIZE,
+    bool skip_cleanup       = false)
+      : bin_growth(bin_growth)
+      , min_bin(min_bin)
+      , max_bin(max_bin)
+      , min_bin_bytes(IntPow(bin_growth, min_bin))
+      , max_bin_bytes(IntPow(bin_growth, max_bin))
+      , max_cached_bytes(max_cached_bytes)
+      , skip_cleanup(skip_cleanup)
+      , debug(false)
+      , cached_blocks(BlockDescriptor::SizeCompare)
+      , live_blocks(BlockDescriptor::PtrCompare)
+  {}
+
+  /**
+   * @brief Default constructor.
+   *
+   * Configured with:
+   * @par
+   * - @p bin_growth          = 8
+   * - @p min_bin             = 3
+   * - @p max_bin             = 7
+   * - @p max_cached_bytes    = ( @p bin_growth ^ @p max_bin) * 3 ) - 1 = 6,291,455 bytes
+   *
+   * which delineates five bin-sizes: 512B, 4KB, 32KB, 256KB, and 2MB and
+   * sets a maximum of 6,291,455 cached bytes per device
+   */
+  CachingDeviceAllocator(bool skip_cleanup = false, bool debug = false)
+      : bin_growth(8)
+      , min_bin(3)
+      , max_bin(7)
+      , min_bin_bytes(IntPow(bin_growth, min_bin))
+      , max_bin_bytes(IntPow(bin_growth, max_bin))
+      , max_cached_bytes((max_bin_bytes * 3) - 1)
+      , skip_cleanup(skip_cleanup)
+      , debug(debug)
+      , cached_blocks(BlockDescriptor::SizeCompare)
+      , live_blocks(BlockDescriptor::PtrCompare)
+  {}
+
+  /**
+   * @brief Sets the limit on the number bytes this allocator is allowed to cache per device.
+   *
+   * Changing the ceiling of cached bytes does not cause any allocations (in-use or
+   * cached-in-reserve) to be freed.  See \p FreeAllCached().
+   */
+  cudaError_t SetMaxCachedBytes(size_t max_cached_bytes_)
+  {
+    // Lock
+    mutex.lock();
+
+#ifdef CUB_DEBUG_LOG
+    _CubLog(
+      "Changing max_cached_bytes (%lld -> %lld)\n", (long long) this->max_cached_bytes, (long long) max_cached_bytes_);
+#endif
+
+    this->max_cached_bytes = max_cached_bytes_;
+
+    // Unlock
+    mutex.unlock();
+
+    return cudaSuccess;
+  }
+
+  /**
+   * @brief Provides a suitable allocation of device memory for the given size on the specified
+   *        device.
+   *
+   * Once freed, the allocation becomes available immediately for reuse within the @p
+   * active_stream with which it was associated with during allocation, and it becomes available
+   * for reuse within other streams when all prior work submitted to @p active_stream has
+   * completed.
+   *
+   * @param[in] device
+   *   Device on which to place the allocation
+   *
+   * @param[out] d_ptr
+   *   Reference to pointer to the allocation
+   *
+   * @param[in] bytes
+   *   Minimum number of bytes for the allocation
+   *
+   * @param[in] active_stream
+   *   The stream to be associated with this allocation
+   */
+  cudaError_t DeviceAllocate(int device, void** d_ptr, size_t bytes, cudaStream_t active_stream = nullptr)
+  {
+    *d_ptr                = nullptr;
+    int entrypoint_device = INVALID_DEVICE_ORDINAL;
+    cudaError_t error     = cudaSuccess;
+
+    if (device == INVALID_DEVICE_ORDINAL)
+    {
+      error = CubDebug(cudaGetDevice(&entrypoint_device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+
+      device = entrypoint_device;
+    }
+
+    // Create a block descriptor for the requested allocation
+    bool found = false;
+    BlockDescriptor search_key(device);
+    search_key.associated_stream = active_stream;
+    NearestPowerOf(search_key.bin, search_key.bytes, bin_growth, bytes);
+
+    if (search_key.bin > max_bin)
+    {
+      // Bin is greater than our maximum bin: allocate the request
+      // exactly and give out-of-bounds bin.  It will not be cached
+      // for reuse when returned.
+      search_key.bin   = INVALID_BIN;
+      search_key.bytes = bytes;
+    }
+    else
+    {
+      // Search for a suitable cached allocation: lock
+      mutex.lock();
+
+      if (search_key.bin < min_bin)
+      {
+        // Bin is less than minimum bin: round up
+        search_key.bin   = min_bin;
+        search_key.bytes = min_bin_bytes;
+      }
+
+      // Iterate through the range of cached blocks on the same device in the same bin
+      CachedBlocks::iterator block_itr = cached_blocks.lower_bound(search_key);
+      while ((block_itr != cached_blocks.end()) && (block_itr->device == device) && (block_itr->bin == search_key.bin))
+      {
+        // To prevent races with reusing blocks returned by the host but still
+        // in use by the device, only consider cached blocks that are
+        // either (from the active stream) or (from an idle stream)
+        bool is_reusable = false;
+        if (active_stream == block_itr->associated_stream)
+        {
+          is_reusable = true;
+        }
+        else
+        {
+          const cudaError_t event_status = cudaEventQuery(block_itr->ready_event);
+          if (event_status != cudaErrorNotReady)
+          {
+            CubDebug(event_status);
+            is_reusable = true;
+          }
+        }
+
+        if (is_reusable)
+        {
+          // Reuse existing cache block.  Insert into live blocks.
+          found                        = true;
+          search_key                   = *block_itr;
+          search_key.associated_stream = active_stream;
+          live_blocks.insert(search_key);
+
+          // Remove from free blocks
+          cached_bytes[device].free -= search_key.bytes;
+          cached_bytes[device].live += search_key.bytes;
+
+#ifdef CUB_DEBUG_LOG
+          _CubLog("\tDevice %d reused cached block at %p (%lld bytes) for stream %lld (previously associated with "
+                  "stream %lld).\n",
+                  device,
+                  search_key.d_ptr,
+                  (long long) search_key.bytes,
+                  (long long) search_key.associated_stream,
+                  (long long) block_itr->associated_stream);
+#endif
+
+          cached_blocks.erase(block_itr);
+
+          break;
+        }
+        block_itr++;
+      }
+
+      // Done searching: unlock
+      mutex.unlock();
+    }
+
+    // Allocate the block if necessary
+    if (!found)
+    {
+      // Set runtime's current device to specified device (entrypoint may not be set)
+      if (device != entrypoint_device)
+      {
+        error = CubDebug(cudaGetDevice(&entrypoint_device));
+        if (cudaSuccess != error)
+        {
+          return error;
+        }
+
+        error = CubDebug(cudaSetDevice(device));
+        if (cudaSuccess != error)
+        {
+          return error;
+        }
+      }
+
+      // Attempt to allocate
+      error = CubDebug(cudaMalloc(&search_key.d_ptr, search_key.bytes));
+      if (error == cudaErrorMemoryAllocation)
+      {
+        // The allocation attempt failed: free all cached blocks on device and retry
+#ifdef CUB_DEBUG_LOG
+        _CubLog("\tDevice %d failed to allocate %lld bytes for stream %lld, retrying after freeing cached allocations",
+                device,
+                (long long) search_key.bytes,
+                (long long) search_key.associated_stream);
+#endif
+
+        error = cudaSuccess; // Reset the error we will return
+        cudaGetLastError(); // Reset CUDART's error
+
+        // Lock
+        mutex.lock();
+
+        // Iterate the range of free blocks on the same device
+        BlockDescriptor free_key(device);
+        CachedBlocks::iterator block_itr = cached_blocks.lower_bound(free_key);
+
+        while ((block_itr != cached_blocks.end()) && (block_itr->device == device))
+        {
+          // No need to worry about synchronization with the device: cudaFree is
+          // blocking and will synchronize across all kernels executing
+          // on the current device
+
+          // Free device memory and destroy stream event.
+          error = CubDebug(cudaFree(block_itr->d_ptr));
+          if (cudaSuccess != error)
+          {
+            break;
+          }
+
+          error = CubDebug(cudaEventDestroy(block_itr->ready_event));
+          if (cudaSuccess != error)
+          {
+            break;
+          }
+
+          // Reduce balance and erase entry
+          cached_bytes[device].free -= block_itr->bytes;
+
+#ifdef CUB_DEBUG_LOG
+          _CubLog("\tDevice %d freed %lld bytes.\n\t\t  %lld available blocks cached (%lld bytes), %lld live blocks "
+                  "(%lld bytes) outstanding.\n",
+                  device,
+                  (long long) block_itr->bytes,
+                  (long long) cached_blocks.size(),
+                  (long long) cached_bytes[device].free,
+                  (long long) live_blocks.size(),
+                  (long long) cached_bytes[device].live);
+#endif
+
+          block_itr = cached_blocks.erase(block_itr);
+        }
+
+        // Unlock
+        mutex.unlock();
+
+        // Return under error
+        if (error)
+        {
+          return error;
+        }
+
+        // Try to allocate again
+        error = CubDebug(cudaMalloc(&search_key.d_ptr, search_key.bytes));
+        if (cudaSuccess != error)
+        {
+          return error;
+        }
+      }
+
+      // Create ready event
+      error = CubDebug(cudaEventCreateWithFlags(&search_key.ready_event, cudaEventDisableTiming));
+
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+
+      // Insert into live blocks
+      mutex.lock();
+      live_blocks.insert(search_key);
+      cached_bytes[device].live += search_key.bytes;
+      mutex.unlock();
+
+#ifdef CUB_DEBUG_LOG
+      _CubLog("\tDevice %d allocated new device block at %p (%lld bytes associated with stream %lld).\n",
+              device,
+              search_key.d_ptr,
+              (long long) search_key.bytes,
+              (long long) search_key.associated_stream);
+#endif
+
+      // Attempt to revert back to previous device if necessary
+      if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device))
+      {
+        error = CubDebug(cudaSetDevice(entrypoint_device));
+        if (cudaSuccess != error)
+        {
+          return error;
+        }
+      }
+    }
+
+    // Copy device pointer to output parameter
+    *d_ptr = search_key.d_ptr;
+
+#ifdef CUB_DEBUG_LOG
+    if (debug)
+    {
+      _CubLog("\t\t%lld available blocks cached (%lld bytes), %lld live blocks outstanding(%lld bytes).\n",
+              (long long) cached_blocks.size(),
+              (long long) cached_bytes[device].free,
+              (long long) live_blocks.size(),
+              (long long) cached_bytes[device].live);
+    }
+#endif
+
+    return error;
+  }
+
+  /**
+   * @brief Provides a suitable allocation of device memory for the given size on the current
+   *        device.
+   *
+   * Once freed, the allocation becomes available immediately for reuse within the @p
+   * active_stream with which it was associated with during allocation, and it becomes available
+   * for reuse within other streams when all prior work submitted to @p active_stream has
+   * completed.
+   *
+   * @param[out] d_ptr
+   *   Reference to pointer to the allocation
+   *
+   * @param[in] bytes
+   *   Minimum number of bytes for the allocation
+   *
+   * @param[in] active_stream
+   *   The stream to be associated with this allocation
+   */
+  cudaError_t DeviceAllocate(void** d_ptr, size_t bytes, cudaStream_t active_stream = nullptr)
+  {
+    return DeviceAllocate(INVALID_DEVICE_ORDINAL, d_ptr, bytes, active_stream);
+  }
+
+  /**
+   * @brief Frees a live allocation of device memory on the specified device, returning it to the
+   *        allocator.
+   *
+   * Once freed, the allocation becomes available immediately for reuse within the
+   * @p active_stream with which it was associated with during allocation, and it becomes
+   * available for reuse within other streams when all prior work submitted to @p active_stream
+   * has completed.
+   */
+  cudaError_t DeviceFree(int device, void* d_ptr)
+  {
+    int entrypoint_device = INVALID_DEVICE_ORDINAL;
+    cudaError_t error     = cudaSuccess;
+
+    if (device == INVALID_DEVICE_ORDINAL)
+    {
+      error = CubDebug(cudaGetDevice(&entrypoint_device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+      device = entrypoint_device;
+    }
+
+    // Lock
+    mutex.lock();
+
+    // Find corresponding block descriptor
+    bool recached = false;
+    BlockDescriptor search_key(d_ptr, device);
+    BusyBlocks::iterator block_itr = live_blocks.find(search_key);
+    if (block_itr != live_blocks.end())
+    {
+      // Remove from live blocks
+      search_key = *block_itr;
+      live_blocks.erase(block_itr);
+      cached_bytes[device].live -= search_key.bytes;
+
+      // Keep the returned allocation if bin is valid and we won't exceed the max cached threshold
+      if ((search_key.bin != INVALID_BIN) && (cached_bytes[device].free + search_key.bytes <= max_cached_bytes))
+      {
+        // Insert returned allocation into free blocks
+        recached = true;
+        cached_blocks.insert(search_key);
+        cached_bytes[device].free += search_key.bytes;
+
+#ifdef CUB_DEBUG_LOG
+        _CubLog("\tDevice %d returned %lld bytes from associated stream %lld.\n\t\t %lld available blocks cached (%lld "
+                "bytes), %lld live blocks outstanding. (%lld bytes)\n",
+                device,
+                (long long) search_key.bytes,
+                (long long) search_key.associated_stream,
+                (long long) cached_blocks.size(),
+                (long long) cached_bytes[device].free,
+                (long long) live_blocks.size(),
+                (long long) cached_bytes[device].live);
+#endif
+      }
+    }
+
+    // Unlock
+    mutex.unlock();
+
+    // First set to specified device (entrypoint may not be set)
+    if (device != entrypoint_device)
+    {
+      error = CubDebug(cudaGetDevice(&entrypoint_device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+
+      error = CubDebug(cudaSetDevice(device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+    }
+
+    if (recached)
+    {
+      // Insert the ready event in the associated stream (must have current device set properly)
+      error = CubDebug(cudaEventRecord(search_key.ready_event, search_key.associated_stream));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+    }
+
+    if (!recached)
+    {
+      // Free the allocation from the runtime and cleanup the event.
+      error = CubDebug(cudaFree(d_ptr));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+
+      error = CubDebug(cudaEventDestroy(search_key.ready_event));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+
+#ifdef CUB_DEBUG_LOG
+      _CubLog("\tDevice %d freed %lld bytes from associated stream %lld.\n\t\t  %lld available blocks cached (%lld "
+              "bytes), %lld live blocks (%lld bytes) outstanding.\n",
+              device,
+              (long long) search_key.bytes,
+              (long long) search_key.associated_stream,
+              (long long) cached_blocks.size(),
+              (long long) cached_bytes[device].free,
+              (long long) live_blocks.size(),
+              (long long) cached_bytes[device].live);
+#endif
+    }
+
+    // Reset device
+    if ((entrypoint_device != INVALID_DEVICE_ORDINAL) && (entrypoint_device != device))
+    {
+      error = CubDebug(cudaSetDevice(entrypoint_device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+    }
+
+    return error;
+  }
+
+  /**
+   * @brief Frees a live allocation of device memory on the current device, returning it to the
+   *        allocator.
+   *
+   * Once freed, the allocation becomes available immediately for reuse within the @p
+   * active_stream with which it was associated with during allocation, and it becomes available
+   * for reuse within other streams when all prior work submitted to @p active_stream has
+   * completed.
+   */
+  cudaError_t DeviceFree(void* d_ptr)
+  {
+    return DeviceFree(INVALID_DEVICE_ORDINAL, d_ptr);
+  }
+
+  /**
+   * @brief Frees all cached device allocations on all devices
+   */
+  cudaError_t FreeAllCached()
+  {
+    cudaError_t error     = cudaSuccess;
+    int entrypoint_device = INVALID_DEVICE_ORDINAL;
+    int current_device    = INVALID_DEVICE_ORDINAL;
+
+    mutex.lock();
+
+    while (!cached_blocks.empty())
+    {
+      // Get first block
+      CachedBlocks::iterator begin = cached_blocks.begin();
+
+      // Get entry-point device ordinal if necessary
+      if (entrypoint_device == INVALID_DEVICE_ORDINAL)
+      {
+        error = CubDebug(cudaGetDevice(&entrypoint_device));
+        if (cudaSuccess != error)
+        {
+          break;
+        }
+      }
+
+      // Set current device ordinal if necessary
+      if (begin->device != current_device)
+      {
+        error = CubDebug(cudaSetDevice(begin->device));
+        if (cudaSuccess != error)
+        {
+          break;
+        }
+        current_device = begin->device;
+      }
+
+      // Free device memory
+      error = CubDebug(cudaFree(begin->d_ptr));
+      if (cudaSuccess != error)
+      {
+        break;
+      }
+
+      error = CubDebug(cudaEventDestroy(begin->ready_event));
+      if (cudaSuccess != error)
+      {
+        break;
+      }
+
+      // Reduce balance and erase entry
+      const size_t block_bytes = begin->bytes;
+      cached_bytes[current_device].free -= block_bytes;
+      cached_blocks.erase(begin);
+
+#ifdef CUB_DEBUG_LOG
+      _CubLog("\tDevice %d freed %lld bytes.\n\t\t  %lld available blocks cached (%lld bytes), %lld live blocks (%lld "
+              "bytes) outstanding.\n",
+              current_device,
+              (long long) block_bytes,
+              (long long) cached_blocks.size(),
+              (long long) cached_bytes[current_device].free,
+              (long long) live_blocks.size(),
+              (long long) cached_bytes[current_device].live);
+#endif
+    }
+
+    mutex.unlock();
+
+    // Attempt to revert back to entry-point device if necessary
+    if (entrypoint_device != INVALID_DEVICE_ORDINAL)
+    {
+      error = CubDebug(cudaSetDevice(entrypoint_device));
+      if (cudaSuccess != error)
+      {
+        return error;
+      }
+    }
+
+    return error;
+  }
+
+  /**
+   * @brief Destructor
+   */
+  virtual ~CachingDeviceAllocator()
+  {
+    if (!skip_cleanup)
+    {
+      FreeAllCached();
+    }
+  }
+};
+
+CUB_NAMESPACE_END
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_arch.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_arch.cuh
new file mode 100644
index 00000000..c5b38bb3
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_arch.cuh
@@ -0,0 +1,219 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2025, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/**
+ * \file
+ * Static architectural properties by SM version.
+ */
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/util_cpp_dialect.cuh> // IWYU pragma: export
+#include <cub/util_macro.cuh>
+#include <cub/util_namespace.cuh>
+
+#include <cuda/__cmath/ceil_div.h>
+#include <cuda/__cmath/round_up.h>
+#include <cuda/__device/compute_capability.h>
+#include <cuda/std/__algorithm/clamp.h>
+#include <cuda/std/__algorithm/max.h>
+#include <cuda/std/__algorithm/min.h>
+
+// Legacy include; this functionality used to be defined in here.
+#include <cub/detail/detect_cuda_runtime.cuh>
+
+CUB_NAMESPACE_BEGIN
+
+#ifndef _CCCL_DOXYGEN_INVOKED // Do not document
+
+/// In device code, CUB_PTX_ARCH expands to the PTX version for which we are
+/// compiling. In host code, CUB_PTX_ARCH's value is implementation defined.
+#  ifndef CUB_PTX_ARCH
+// deprecated in 3.1
+#    if _CCCL_CUDA_COMPILER(NVHPC)
+// NV_TARGET_MINIMUM_SM_INTEGER is the oldest target PTX version, and is defined when compiling both host code and
+// device code.
+#      define CUB_PTX_ARCH (NV_TARGET_MINIMUM_SM_INTEGER * 10)
+#    else // ^^^ _CCCL_CUDA_COMPILER(NVHPC) ^^^ / vvv !_CCCL_CUDA_COMPILER(NVHPC) vvv
+#      define CUB_PTX_ARCH _CCCL_PTX_ARCH()
+#    endif // ^^^ !_CCCL_CUDA_COMPILER(NVHPC) ^^^
+#  endif
+
+/// Maximum number of devices supported.
+#  ifndef CUB_MAX_DEVICES
+//! Deprecated [Since 3.0]
+#    define CUB_MAX_DEVICES (128)
+#  endif
+static_assert(CUB_MAX_DEVICES > 0, "CUB_MAX_DEVICES must be greater than 0.");
+
+/// Number of threads per warp
+#  ifndef CUB_LOG_WARP_THREADS
+//! Deprecated [Since 3.0]
+#    define CUB_LOG_WARP_THREADS(unused) (5)
+//! Deprecated [Since 3.0]
+#    define CUB_WARP_THREADS(unused) (1 << CUB_LOG_WARP_THREADS(0))
+
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_WARP_THREADS CUB_WARP_THREADS(0)
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_LOG_WARP_THREADS CUB_LOG_WARP_THREADS(0)
+#  endif
+
+/// Number of smem banks
+#  ifndef CUB_LOG_SMEM_BANKS
+//! Deprecated [Since 3.0]
+#    define CUB_LOG_SMEM_BANKS(unused) (5)
+//! Deprecated [Since 3.0]
+#    define CUB_SMEM_BANKS(unused) (1 << CUB_LOG_SMEM_BANKS(0))
+
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_LOG_SMEM_BANKS CUB_LOG_SMEM_BANKS(0)
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_SMEM_BANKS CUB_SMEM_BANKS
+#  endif
+
+/// Oversubscription factor
+#  ifndef CUB_SUBSCRIPTION_FACTOR
+//! Deprecated [Since 3.0]
+#    define CUB_SUBSCRIPTION_FACTOR(unused) (5)
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_SUBSCRIPTION_FACTOR CUB_SUBSCRIPTION_FACTOR(0)
+#  endif
+
+/// Prefer padding overhead vs X-way conflicts greater than this threshold
+#  ifndef CUB_PREFER_CONFLICT_OVER_PADDING
+//! Deprecated [Since 3.0]
+#    define CUB_PREFER_CONFLICT_OVER_PADDING(unused) (1)
+//! Deprecated [Since 3.0]
+#    define CUB_PTX_PREFER_CONFLICT_OVER_PADDING CUB_PREFER_CONFLICT_OVER_PADDING(0)
+#  endif
+
+namespace detail
+{
+inline constexpr int max_devices       = CUB_MAX_DEVICES;
+inline constexpr int warp_threads      = CUB_PTX_WARP_THREADS;
+inline constexpr int log2_warp_threads = CUB_PTX_LOG_WARP_THREADS;
+inline constexpr int smem_banks        = CUB_SMEM_BANKS(0);
+inline constexpr int log2_smem_banks   = CUB_PTX_LOG_SMEM_BANKS;
+
+inline constexpr int subscription_factor           = CUB_PTX_SUBSCRIPTION_FACTOR;
+inline constexpr bool prefer_conflict_over_padding = CUB_PTX_PREFER_CONFLICT_OVER_PADDING;
+
+// The maximum amount of shared memory available per thread block for eternity. Every current and future CUDA
+// architecture has and will have at least this amount of shared memory. This is also the maximum size of total static
+// shared memory in a kernel. Note that dynamic shared memory may be larger than this amount.
+static constexpr ::cuda::std::size_t max_smem_per_block = 48 * 1024;
+
+// The size in bytes of the largest machine word that can be atomically read/written in a single instruction, so we can
+// use it to pass messages from one thread to another using strong loads (acquire) and stores (release).
+inline constexpr int largest_atomic_message_size = 16;
+
+struct scaling_result
+{
+  int items_per_thread;
+  int threads_per_block;
+};
+
+[[nodiscard]] _CCCL_HOST_DEVICE_API inline constexpr auto
+scale_reg_bound(int nominal_4B_threads_per_block, int nominal_4B_items_per_thread, int target_type_size)
+  -> scaling_result
+{
+  const int items_per_thread =
+    (::cuda::std::max) (1, nominal_4B_items_per_thread * 4 / (::cuda::std::max) (4, target_type_size));
+  const int threads_per_block =
+    (::cuda::std::min) (nominal_4B_threads_per_block,
+                        ::cuda::ceil_div(int{max_smem_per_block} / (target_type_size * items_per_thread), 32) * 32);
+  return {items_per_thread, threads_per_block};
+}
+
+template <int Nominal4ByteThreadsPerBlock, int Nominal4ByteItemsPerThread, typename T>
+struct RegBoundScaling
+{
+private:
+  static constexpr auto result =
+    scale_reg_bound(Nominal4ByteThreadsPerBlock, Nominal4ByteItemsPerThread, int{sizeof(T)});
+
+public:
+  static constexpr int ITEMS_PER_THREAD = result.items_per_thread;
+  static constexpr int BLOCK_THREADS    = result.threads_per_block;
+};
+
+[[nodiscard]] _CCCL_HOST_DEVICE_API inline constexpr auto
+scale_mem_bound(int nominal_4B_threads_per_block, int nominal_4B_items_per_thread, int target_type_size)
+  -> scaling_result
+{
+  const int items_per_thread =
+    ::cuda::std::clamp(nominal_4B_items_per_thread * 4 / target_type_size, 1, nominal_4B_items_per_thread * 2);
+  const int threads_per_block =
+    (::cuda::std::min) (nominal_4B_threads_per_block,
+                        ::cuda::round_up(int{max_smem_per_block} / (target_type_size * items_per_thread), 32));
+  return {items_per_thread, threads_per_block};
+}
+
+template <int Nominal4ByteThreadsPerBlock, int Nominal4ByteItemsPerThread, typename T>
+struct MemBoundScaling
+{
+private:
+  static constexpr auto result =
+    scale_mem_bound(Nominal4ByteThreadsPerBlock, Nominal4ByteItemsPerThread, int{sizeof(T)});
+
+public:
+  static constexpr int ITEMS_PER_THREAD = result.items_per_thread;
+  static constexpr int BLOCK_THREADS    = result.threads_per_block;
+};
+
+template <int Nominal4ByteThreadsPerBlock, int Nominal4ByteItemsPerThread, typename = void>
+struct NoScaling
+{
+  static constexpr int ITEMS_PER_THREAD = Nominal4ByteItemsPerThread;
+  static constexpr int BLOCK_THREADS    = Nominal4ByteThreadsPerBlock;
+};
+
+[[nodiscard]] _CCCL_HOST_DEVICE_API constexpr ::cuda::compute_capability current_tuning_cc() noexcept
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC)
+  return ::cuda::compute_capability(NV_TARGET_MINIMUM_SM_INTEGER);
+#  elif _CCCL_DEVICE_COMPILATION()
+  return ::cuda::device::current_compute_capability();
+#  else
+  // clang 22+ supports __CUDA_ARCH_LIST__ and also instantiates tuning policies inside kernels during the **host**
+  // pass (e.g. to compute the value for __launch_bounds__), where we rely on current_tuning_cc(), which is then passed
+  // to the policy selector. In the rare case that the policy selector is an adapter over a policy hub and invokes
+  // ChainedPolicy (e.g. test cub.test.device.histogram_custom_policy_hub.lid_0), it will fail to compile during
+  // constant evaluation, since it cannot find a policy for a PTX version of zero. As a workaround, we return the oldest
+  // CC we are compiling for during the host pass. And for consistency, we do the same for all compilers.
+#    if _CCCL_CUDA_COMPILER(CLANG)
+  return ::cuda::__target_compute_capabilities().front();
+#    else // ^^^ _CCCL_CUDA_COMPILER(CLANG) ^^^ / vvv !_CCCL_CUDA_COMPILER(CLANG) vvv
+  return {};
+#    endif // ^^^ !_CCCL_CUDA_COMPILER(CLANG) ^^^
+#  endif
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class PolicySelector>
+[[nodiscard]] _CCCL_HOST_DEVICE_API constexpr auto select_policy(::cuda::compute_capability cc)
+{
+  return PolicySelector{}(cc);
+}
+
+template <class PolicySelector>
+[[nodiscard]] _CCCL_DEVICE_API constexpr auto current_policy()
+{
+  return select_policy<PolicySelector>(current_tuning_cc());
+}
+} // namespace detail
+#endif // Do not document
+
+CUB_NAMESPACE_END
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_cpp_dialect.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_cpp_dialect.cuh
new file mode 100644
index 00000000..34fbcc46
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_cpp_dialect.cuh
@@ -0,0 +1,71 @@
+// SPDX-FileCopyrightText: Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+//! @file
+//! Detect the version of the C++ standard used by the compiler.
+
+#pragma once
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#ifndef _CCCL_DOXYGEN_INVOKED // Do not document
+
+// Deprecation warnings may be silenced by defining the following macros. These
+// may be combined.
+// - CCCL_IGNORE_DEPRECATED_COMPILER
+//   Ignore deprecation warnings when using deprecated compilers. Compiling
+//   with deprecated C++ dialects will still issue warnings.
+
+//! Deprecated [Since 3.0]
+#  define CUB_CPP_DIALECT _CCCL_STD_VER
+
+// Define CUB_COMPILER_DEPRECATION macro:
+#  if _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#    define CUB_COMP_DEPR_IMPL(msg) _CCCL_PRAGMA(message(__FILE__ ":" _CCCL_TO_STRING(__LINE__) ": warning: " #msg))
+#  else // clang / gcc:
+#    define CUB_COMP_DEPR_IMPL(msg) _CCCL_PRAGMA(GCC warning #msg)
+#  endif
+
+// Compiler checks:
+// clang-format off
+#  define CUB_COMPILER_DEPRECATION(REQ) \
+    CUB_COMP_DEPR_IMPL(CUB requires at least REQ. Define CCCL_IGNORE_DEPRECATED_COMPILER to suppress this message.)
+
+#  define CUB_COMPILER_DEPRECATION_SOFT(REQ, CUR)                                                        \
+    CUB_COMP_DEPR_IMPL(                                                                                  \
+      CUB requires at least REQ. CUR is deprecated but still supported. CUR support will be removed in a \
+        future release. Define CCCL_IGNORE_DEPRECATED_COMPILER to suppress this message.)
+// clang-format on
+
+#  ifndef CCCL_IGNORE_DEPRECATED_COMPILER
+#    if _CCCL_COMPILER(GCC, <, 7)
+CUB_COMPILER_DEPRECATION(GCC 7.0);
+#    elif _CCCL_COMPILER(CLANG, <, 7)
+CUB_COMPILER_DEPRECATION(Clang 7.0);
+#    elif _CCCL_COMPILER(MSVC, <, 19, 10)
+// <2017. Hard upgrade message:
+CUB_COMPILER_DEPRECATION(MSVC 2019(19.20 / 16.0 / 14.20));
+#    endif
+#  endif // CCCL_IGNORE_DEPRECATED_COMPILER
+
+#  undef CUB_COMPILER_DEPRECATION_SOFT
+#  undef CUB_COMPILER_DEPRECATION
+
+// C++17 dialect check:
+#  ifndef CCCL_IGNORE_DEPRECATED_CPP_DIALECT
+#    if _CCCL_STD_VER < 2017
+#      error CUB requires at least C++17. Define CCCL_IGNORE_DEPRECATED_CPP_DIALECT to suppress this message.
+#    endif // _CCCL_STD_VER < 2017
+#  endif
+
+#  undef CUB_COMP_DEPR_IMPL
+
+#endif // !_CCCL_DOXYGEN_INVOKED
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_debug.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_debug.cuh
new file mode 100644
index 00000000..bb58c46a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_debug.cuh
@@ -0,0 +1,187 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2022, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/**
+ * \file
+ * Error and event logging routines.
+ *
+ * The following macros definitions are supported:
+ * - \p CUB_LOG.  Simple event messages are printed to \p stdout.
+ */
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <nv/target>
+
+#ifdef _CCCL_DOXYGEN_INVOKED // Only parse this during doxygen passes:
+
+/**
+ * @def CUB_DEBUG_LOG
+ *
+ * Causes kernel launch configurations to be printed to the console
+ */
+#  define CUB_DEBUG_LOG
+
+/**
+ * @def CUB_DEBUG_SYNC
+ *
+ * Causes synchronization of the stream after every kernel launch to check
+ * for errors. Also causes kernel launch configurations to be printed to the
+ * console.
+ */
+#  define CUB_DEBUG_SYNC
+
+/**
+ * @def CUB_DEBUG_ALL
+ *
+ * Causes host and device-side precondition assertions to be checked. Apart
+ * from that, causes synchronization of the stream after every kernel launch to
+ * check for errors. Also causes kernel launch configurations to be printed to
+ * the console.
+ */
+#  define CUB_DEBUG_ALL
+
+#endif // _CCCL_DOXYGEN_INVOKED
+
+// CUB_DEBUG_SYNC also enables CUB_DEBUG_LOG
+#ifdef CUB_DEBUG_SYNC
+#  ifndef CUB_DEBUG_LOG
+#    define CUB_DEBUG_LOG
+#  endif
+#endif
+
+// CUB_DEBUG_ALL = CUB_DEBUG_LOG + CUB_DEBUG_SYNC
+#ifdef CUB_DEBUG_ALL
+#  ifndef CUB_DEBUG_LOG
+#    define CUB_DEBUG_LOG
+#  endif // CUB_DEBUG_LOG
+#  ifndef CUB_DEBUG_SYNC
+#    define CUB_DEBUG_SYNC
+#  endif // CUB_DEBUG_SYNC
+#endif // CUB_DEBUG_ALL
+
+/// CUB error reporting macro (prints error messages to stderr)
+#if (defined(DEBUG) || defined(_DEBUG)) && !defined(CUB_STDERR)
+#  define CUB_STDERR
+#endif
+
+#if defined(CUB_STDERR) || defined(CUB_DEBUG_LOG)
+#  include <cuda/std/__host_stdlib/cstdio>
+#endif
+
+CUB_NAMESPACE_BEGIN
+
+/**
+ * \brief %If \p CUB_STDERR is defined and \p error is not \p cudaSuccess, the
+ * corresponding error message is printed to \p stderr (or \p stdout in device
+ * code) along with the supplied source context.
+ *
+ * \return The CUDA error.
+ */
+_CCCL_HOST_DEVICE _CCCL_FORCEINLINE cudaError_t
+Debug(cudaError_t error, [[maybe_unused]] const char* filename, [[maybe_unused]] int line)
+{
+  // Clear the global CUDA error state which may have been set by the last
+  // call. Otherwise, errors may "leak" to unrelated kernel launches.
+
+  // clang-format off
+  #ifndef CUB_RDC_ENABLED
+  #define CUB_TEMP_DEVICE_CODE
+  #else
+  #define CUB_TEMP_DEVICE_CODE last_error = cudaGetLastError()
+  #endif
+
+  cudaError_t last_error = cudaSuccess;
+
+  NV_IF_ELSE_TARGET(
+    NV_IS_HOST,
+    (last_error = cudaGetLastError();),
+    (CUB_TEMP_DEVICE_CODE;)
+  );
+
+  #undef CUB_TEMP_DEVICE_CODE
+  // clang-format on
+
+  if (error == cudaSuccess && last_error != cudaSuccess)
+  {
+    error = last_error;
+  }
+
+#ifdef CUB_STDERR
+  if (error)
+  {
+    NV_IF_ELSE_TARGET(
+      NV_IS_HOST,
+      (fprintf(stderr, "CUDA error %d [%s, %d]: %s\n", error, filename, line, cudaGetErrorString(error));
+       fflush(stderr);),
+      (printf("CUDA error %d [block (%d,%d,%d) thread (%d,%d,%d), %s, %d]\n",
+              error,
+              blockIdx.z,
+              blockIdx.y,
+              blockIdx.x,
+              threadIdx.z,
+              threadIdx.y,
+              threadIdx.x,
+              filename,
+              line);));
+  }
+#endif
+
+  return error;
+}
+
+/**
+ * \brief Debug macro
+ */
+#ifndef CubDebug
+#  define CubDebug(e) CUB_NS_QUALIFIER::Debug((cudaError_t) (e), __FILE__, __LINE__)
+#endif
+
+/**
+ * \brief Debug macro with exit
+ */
+#ifndef CubDebugExit
+#  define CubDebugExit(e)                                               \
+    if (CUB_NS_QUALIFIER::Debug((cudaError_t) (e), __FILE__, __LINE__)) \
+    {                                                                   \
+      exit(1);                                                          \
+    }
+#endif
+
+/**
+ * \brief Log macro for printf statements.
+ */
+#if !defined(_CubLog)
+#  if _CCCL_HOSTJIT()
+#    define _CubLog(format, ...) (void(0))
+#  else // ^^^ _CCCL_HOSTJIT() ^^^ / vvv !_CCCL_HOSTJIT() vvv
+#    define _CubLog(format, ...)                                    \
+      do                                                            \
+      {                                                             \
+        NV_IF_ELSE_TARGET(                                          \
+          NV_IS_HOST,                                               \
+          (printf(format, __VA_ARGS__);),                           \
+          (printf("[block (%d,%d,%d), thread (%d,%d,%d)]: " format, \
+                  blockIdx.z,                                       \
+                  blockIdx.y,                                       \
+                  blockIdx.x,                                       \
+                  threadIdx.z,                                      \
+                  threadIdx.y,                                      \
+                  threadIdx.x,                                      \
+                  __VA_ARGS__);));                                  \
+      } while (false)
+#  endif // !_CCCL_HOSTJIT()
+#endif // !defined(_CubLog)
+
+CUB_NAMESPACE_END
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_macro.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_macro.cuh
new file mode 100644
index 00000000..8b9c43bd
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_macro.cuh
@@ -0,0 +1,67 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2024, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/******************************************************************************
+ * Common C/C++ macro utilities
+ ******************************************************************************/
+
+#pragma once
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/detail/detect_cuda_runtime.cuh> // IWYU pragma: export
+#include <cub/util_namespace.cuh> // IWYU pragma: export
+
+CUB_NAMESPACE_BEGIN
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CUB_DISABLE_KERNEL_VISIBILITY_WARNING_SUPPRESSION
+#endif
+
+/**
+ * @def CUB_DISABLE_KERNEL_VISIBILITY_WARNING_SUPPRESSION
+ * If defined, the default suppression of kernel visibility attribute warning is disabled.
+ */
+#if !defined(CUB_DISABLE_KERNEL_VISIBILITY_WARNING_SUPPRESSION)
+_CCCL_DIAG_SUPPRESS_GCC("-Wattributes")
+_CCCL_DIAG_SUPPRESS_CLANG("-Wattributes")
+#  if !_CCCL_CUDA_COMPILER(NVHPC)
+_CCCL_DIAG_SUPPRESS_NVHPC(attribute_requires_external_linkage)
+#  endif // !_CCCL_CUDA_COMPILER(NVHPC)
+#endif // !CUB_DISABLE_KERNEL_VISIBILITY_WARNING_SUPPRESSION
+
+#ifndef CUB_DEFINE_KERNEL_GETTER
+#  define CUB_DEFINE_KERNEL_GETTER(name, ...)                                               \
+    _CCCL_HIDE_FROM_ABI CUB_RUNTIME_FUNCTION static constexpr decltype(&__VA_ARGS__) name() \
+    {                                                                                       \
+      return &__VA_ARGS__;                                                                  \
+    }
+#endif
+
+// TODO(bgruber): drop in CCCL 4.0 when we drop the public dispatchers
+#ifndef CUB_DEFINE_SUB_POLICY_GETTER
+#  define CUB_DEFINE_SUB_POLICY_GETTER(name)                            \
+    _CCCL_HOST_DEVICE static constexpr auto name()                      \
+    {                                                                   \
+      return MakePolicyWrapper(typename StaticPolicyT::name##Policy()); \
+    }
+#endif
+
+#if defined(CUB_DEFINE_RUNTIME_POLICIES)
+#  define CUB_DETAIL_STATIC_ISH_ASSERT(expr, msg) _CCCL_ASSERT(expr, msg)
+#  define CUB_DETAIL_CONSTEXPR_ISH
+#else // ^^^ CUB_DEFINE_RUNTIME_POLICIES ^^^ / vvv !CUB_DEFINE_RUNTIME_POLICIES vvv
+#  define CUB_DETAIL_STATIC_ISH_ASSERT(expr, msg) static_assert(expr, msg);
+#  define CUB_DETAIL_CONSTEXPR_ISH                constexpr
+#endif // !(CUB_DEFINE_RUNTIME_POLICIES)
+
+CUB_NAMESPACE_END
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/util_namespace.cuh b/qwen3_6_scripts/cccl_preload/include/cub/util_namespace.cuh
new file mode 100644
index 00000000..60ddf454
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/util_namespace.cuh
@@ -0,0 +1,172 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011, Duane Merrill. All rights reserved.
+// SPDX-FileCopyrightText: Copyright (c) 2011-2021, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/**
+ * \file util_namespace.cuh
+ * \brief Utilities that allow `cub::` to be placed inside an
+ * application-specific namespace.
+ */
+
+#pragma once
+
+// This is not used by this file; this is a hack so that we can detect the
+// CUB version from Thrust on older versions of CUB that did not have
+// version.cuh.
+#include <cub/version.cuh>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cub/detail/detect_cuda_runtime.cuh>
+
+// Prior to 1.13.1, only the PREFIX/POSTFIX macros were used. Notify users
+// that they must now define the qualifier macro, too.
+#if (defined(CUB_NS_PREFIX) || defined(CUB_NS_POSTFIX)) && !defined(CUB_NS_QUALIFIER)
+#  error CUB requires a definition of CUB_NS_QUALIFIER when CUB_NS_PREFIX/POSTFIX are defined.
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define THRUST_CUB_WRAPPED_NAMESPACE
+#endif
+
+/**
+ * \def THRUST_CUB_WRAPPED_NAMESPACE
+ * If defined, this value will be used as the name of a namespace that wraps the
+ * `thrust::` and `cub::` namespaces.
+ * This macro should not be used with any other CUB namespace macros.
+ */
+#ifdef THRUST_CUB_WRAPPED_NAMESPACE
+#  define CUB_WRAPPED_NAMESPACE THRUST_CUB_WRAPPED_NAMESPACE
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CUB_WRAPPED_NAMESPACE
+#endif
+
+/**
+ * \def CUB_WRAPPED_NAMESPACE
+ * If defined, this value will be used as the name of a namespace that wraps the
+ * `cub::` namespace.
+ * If THRUST_CUB_WRAPPED_NAMESPACE is set, this will inherit that macro's value.
+ * This macro should not be used with any other CUB namespace macros.
+ */
+#ifdef CUB_WRAPPED_NAMESPACE
+#  define CUB_NS_PREFIX             \
+    namespace CUB_WRAPPED_NAMESPACE \
+    {
+#  define CUB_NS_POSTFIX }
+
+#  define CUB_NS_QUALIFIER ::CUB_WRAPPED_NAMESPACE::cub
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CUB_NS_PREFIX
+#endif
+
+/**
+ * \def CUB_NS_PREFIX
+ * This macro is inserted prior to all `namespace cub { ... }` blocks. It is
+ * derived from CUB_WRAPPED_NAMESPACE, if set, and will be empty otherwise.
+ * It may be defined by users, in which case CUB_NS_PREFIX,
+ * CUB_NS_POSTFIX, and CUB_NS_QUALIFIER must all be set consistently.
+ */
+#ifndef CUB_NS_PREFIX
+#  define CUB_NS_PREFIX
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CUB_NS_POSTFIX
+#endif
+
+/**
+ * \def CUB_NS_POSTFIX
+ * This macro is inserted following the closing braces of all
+ * `namespace cub { ... }` block. It is defined appropriately when
+ * CUB_WRAPPED_NAMESPACE is set, and will be empty otherwise. It may be
+ * defined by users, in which case CUB_NS_PREFIX, CUB_NS_POSTFIX, and
+ * CUB_NS_QUALIFIER must all be set consistently.
+ */
+#ifndef CUB_NS_POSTFIX
+#  define CUB_NS_POSTFIX
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CUB_NS_QUALIFIER
+#endif
+
+/**
+ * \def CUB_NS_QUALIFIER
+ * This macro is used to qualify members of cub:: when accessing them from
+ * outside of their namespace. By default, this is just `::cub`, and will be
+ * set appropriately when CUB_WRAPPED_NAMESPACE is defined. This macro may be
+ * defined by users, in which case CUB_NS_PREFIX, CUB_NS_POSTFIX, and
+ * CUB_NS_QUALIFIER must all be set consistently.
+ */
+#ifndef CUB_NS_QUALIFIER
+#  define CUB_NS_QUALIFIER ::cub
+#endif
+
+#if defined(CUB_DISABLE_NAMESPACE_MAGIC) || defined(CUB_WRAPPED_NAMESPACE)
+#  if !defined(CUB_WRAPPED_NAMESPACE)
+#    if !defined(CUB_IGNORE_NAMESPACE_MAGIC_ERROR)
+#      error "Disabling namespace magic is unsafe without wrapping namespace"
+#    endif // !defined(CUB_IGNORE_NAMESPACE_MAGIC_ERROR)
+#  endif // !defined(CUB_WRAPPED_NAMESPACE)
+#  define CUB_DETAIL_MAGIC_NS_BEGIN
+#  define CUB_DETAIL_MAGIC_NS_END
+#else // not defined(CUB_DISABLE_NAMESPACE_MAGIC)
+#  if defined(_NVHPC_CUDA)
+#    define CUB_DETAIL_MAGIC_NS_BEGIN                                                                                  \
+      inline namespace _CCCL_PP_CAT(                                                                                   \
+        _CCCL_PP_CAT(_CCCL_PP_CAT(_V_, CUB_VERSION), _CCCL_PP_SPLICE_WITH(_, _SM, NV_TARGET_SM_INTEGER_LIST)), _NVHPC) \
+      {
+#    define CUB_DETAIL_MAGIC_NS_END }
+#  else // not defined(_NVHPC_CUDA)
+#    define CUB_DETAIL_MAGIC_NS_BEGIN                                                                                 \
+      inline namespace _CCCL_PP_CAT(_CCCL_PP_CAT(_V_, CUB_VERSION), _CCCL_PP_SPLICE_WITH(_, _SM, __CUDA_ARCH_LIST__)) \
+      {
+#    define CUB_DETAIL_MAGIC_NS_END }
+#  endif // not defined(_NVHPC_CUDA)
+#endif // not defined(CUB_DISABLE_NAMESPACE_MAGIC)
+
+/**
+ * \def CUB_NAMESPACE_BEGIN
+ * This macro is used to open a `cub::` namespace block, along with any
+ * enclosing namespaces requested by CUB_WRAPPED_NAMESPACE, etc.
+ * This macro is defined by CUB and may not be overridden.
+ */
+#define CUB_NAMESPACE_BEGIN \
+  CUB_NS_PREFIX             \
+  namespace cub             \
+  {                         \
+  CUB_DETAIL_MAGIC_NS_BEGIN
+
+/**
+ * \def CUB_NAMESPACE_END
+ * This macro is used to close a `cub::` namespace block, along with any
+ * enclosing namespaces requested by CUB_WRAPPED_NAMESPACE, etc.
+ * This macro is defined by CUB and may not be overridden.
+ */
+#define CUB_NAMESPACE_END   \
+  CUB_DETAIL_MAGIC_NS_END   \
+  } /* end namespace cub */ \
+  CUB_NS_POSTFIX
+
+// Declare these namespaces here for the purpose of Doxygenating them
+CUB_NS_PREFIX
+
+/*! \namespace cub
+ *  \brief \p cub is the top-level namespace which contains all CUB
+ *         functions and types.
+ */
+namespace cub
+{
+}
+
+CUB_NS_POSTFIX
diff --git a/qwen3_6_scripts/cccl_preload/include/cub/version.cuh b/qwen3_6_scripts/cccl_preload/include/cub/version.cuh
new file mode 100644
index 00000000..986c2575
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cub/version.cuh
@@ -0,0 +1,65 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2022, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/*! \file version.cuh
+ *  \brief Compile-time macros encoding CUB release version
+ *
+ *         <cub/version.h> is the only CUB header that is guaranteed to
+ *         change with every CUB release.
+ *
+ */
+
+#pragma once
+
+// For _CCCL_IMPLICIT_SYSTEM_HEADER
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/version>
+
+/*! \def CUB_VERSION
+ *  \brief The preprocessor macro \p CUB_VERSION encodes the version
+ *         number of the CUB library as MMMmmmpp.
+ *
+ *  \note CUB_VERSION is formatted as `MMMmmmpp`, which differs from `CCCL_VERSION` that uses `MMMmmmppp`.
+ *
+ *         <tt>CUB_VERSION % 100</tt> is the sub-minor version.
+ *         <tt>CUB_VERSION / 100 % 1000</tt> is the minor version.
+ *         <tt>CUB_VERSION / 100000</tt> is the major version.
+ */
+#define CUB_VERSION 300500 // macro expansion with ## requires this to be a single value
+
+/*! \def CUB_MAJOR_VERSION
+ *  \brief The preprocessor macro \p CUB_MAJOR_VERSION encodes the
+ *         major version number of the CUB library.
+ */
+#define CUB_MAJOR_VERSION (CUB_VERSION / 100000)
+
+/*! \def CUB_MINOR_VERSION
+ *  \brief The preprocessor macro \p CUB_MINOR_VERSION encodes the
+ *         minor version number of the CUB library.
+ */
+#define CUB_MINOR_VERSION (CUB_VERSION / 100 % 1000)
+
+/*! \def CUB_SUBMINOR_VERSION
+ *  \brief The preprocessor macro \p CUB_SUBMINOR_VERSION encodes the
+ *         sub-minor version number of the CUB library.
+ */
+#define CUB_SUBMINOR_VERSION (CUB_VERSION % 100)
+
+/*! \def CUB_PATCH_NUMBER
+ *  \brief The preprocessor macro \p CUB_PATCH_NUMBER encodes the
+ *         patch number of the CUB library.
+ */
+#define CUB_PATCH_NUMBER 0
+
+static_assert(CUB_MAJOR_VERSION == CCCL_MAJOR_VERSION);
+static_assert(CUB_MINOR_VERSION == CCCL_MINOR_VERSION);
+static_assert(CUB_SUBMINOR_VERSION == CCCL_PATCH_VERSION);
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__cccl_config b/qwen3_6_scripts/cccl_preload/include/cuda/__cccl_config
new file mode 100644
index 00000000..4ff35e24
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__cccl_config
@@ -0,0 +1,38 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA__CCCL_CONFIG
+#define _CUDA__CCCL_CONFIG
+
+#include <cuda/std/__cccl/architecture.h> // IWYU pragma: export
+#include <cuda/std/__cccl/assert.h> // IWYU pragma: export
+#include <cuda/std/__cccl/attributes.h> // IWYU pragma: export
+#include <cuda/std/__cccl/builtin.h> // IWYU pragma: export
+#include <cuda/std/__cccl/compiler.h> // IWYU pragma: export
+#include <cuda/std/__cccl/cuda_capabilities.h> // IWYU pragma: export
+#include <cuda/std/__cccl/cuda_toolkit.h> // IWYU pragma: export
+#include <cuda/std/__cccl/deprecated.h> // IWYU pragma: export
+#include <cuda/std/__cccl/diagnostic.h> // IWYU pragma: export
+#include <cuda/std/__cccl/dialect.h> // IWYU pragma: export
+#include <cuda/std/__cccl/exceptions.h> // IWYU pragma: export
+#include <cuda/std/__cccl/execution_space.h> // IWYU pragma: export
+#include <cuda/std/__cccl/extended_data_types.h> // IWYU pragma: export
+#include <cuda/std/__cccl/host_std_lib.h> // IWYU pragma: export
+#include <cuda/std/__cccl/os.h> // IWYU pragma: export
+#include <cuda/std/__cccl/preprocessor.h> // IWYU pragma: export
+#include <cuda/std/__cccl/ptx_isa.h> // IWYU pragma: export
+#include <cuda/std/__cccl/rtti.h> // IWYU pragma: export
+#include <cuda/std/__cccl/sequence_access.h> // IWYU pragma: export
+#include <cuda/std/__cccl/system_header.h> // IWYU pragma: export
+#include <cuda/std/__cccl/unreachable.h> // IWYU pragma: export
+#include <cuda/std/__cccl/version.h> // IWYU pragma: export
+#include <cuda/std/__cccl/visibility.h> // IWYU pragma: export
+
+#endif // _CUDA__CCCL_CONFIG
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/ceil_div.h b/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/ceil_div.h
new file mode 100644
index 00000000..139ebf34
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/ceil_div.h
@@ -0,0 +1,123 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___CMATH_CEIL_DIV_H
+#define _CUDA___CMATH_CEIL_DIV_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/min.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/common_type.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_signed.h>
+#include <cuda/std/__type_traits/make_unsigned.h>
+#include <cuda/std/__type_traits/underlying_type.h>
+#include <cuda/std/__utility/to_underlying.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+//! @brief Divides two numbers \p __a and \p __b, rounding up if there is a remainder
+//! @param __a The dividend
+//! @param __b The divisor
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_integral_v<_Tp> _CCCL_AND ::cuda::std::is_integral_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<_Tp, _Up> ceil_div(const _Tp __a, const _Up __b) noexcept
+{
+  _CCCL_ASSERT(__b > _Up{0}, "cuda::ceil_div: 'b' must be positive");
+  if constexpr (::cuda::std::is_signed_v<_Tp>)
+  {
+    _CCCL_ASSERT(__a >= _Tp{0}, "cuda::ceil_div: 'a' must be non negative");
+  }
+  using _Common = ::cuda::std::common_type_t<_Tp, _Up>;
+  using _Prom   = decltype(_Tp{} / _Up{});
+  using _UProm  = ::cuda::std::make_unsigned_t<_Prom>;
+  auto __a1     = static_cast<_UProm>(__a);
+  auto __b1     = static_cast<_UProm>(__b);
+  if constexpr (::cuda::std::is_signed_v<_Prom>)
+  {
+    return static_cast<_Common>((__a1 + __b1 - 1) / __b1);
+  }
+  else
+  {
+    _CCCL_IF_CONSTEVAL_DEFAULT
+    {
+      const auto __res = __a1 / __b1;
+      return static_cast<_Common>(__res + (__res * __b1 != __a1));
+    }
+    else
+    {
+      // the ::min method is faster even if __b is a compile-time constant
+      NV_IF_ELSE_TARGET(NV_IS_DEVICE,
+                        (return static_cast<_Common>(::cuda::std::min(__a1, 1 + ((__a1 - 1) / __b1)));),
+                        (const auto __res = __a1 / __b1; //
+                         return static_cast<_Common>(__res + (__res * __b1 != __a1));))
+    }
+  }
+}
+
+//! @brief Divides two numbers \p __a and \p __b, rounding up if there is a remainder, \p __b is an enum
+//! @param __a The dividend
+//! @param __b The divisor
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_integral_v<_Tp> _CCCL_AND ::cuda::std::is_enum_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<_Tp, ::cuda::std::underlying_type_t<_Up>>
+ceil_div(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::ceil_div(__a, ::cuda::std::to_underlying(__b));
+}
+
+//! @brief Divides two numbers \p __a and \p __b, rounding up if there is a remainder, \p __b is an enum
+//! @param __a The dividend
+//! @param __b The divisor
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_enum_v<_Tp> _CCCL_AND ::cuda::std::is_integral_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<::cuda::std::underlying_type_t<_Tp>, _Up>
+ceil_div(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::ceil_div(::cuda::std::to_underlying(__a), __b);
+}
+
+//! @brief Divides two numbers \p __a and \p __b, rounding up if there is a remainder, \p __b is an enum
+//! @param __a The dividend
+//! @param __b The divisor
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_enum_v<_Tp> _CCCL_AND ::cuda::std::is_enum_v<_Up>)
+[[nodiscard]]
+_CCCL_API constexpr ::cuda::std::common_type_t<::cuda::std::underlying_type_t<_Tp>, ::cuda::std::underlying_type_t<_Up>>
+ceil_div(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::ceil_div(::cuda::std::to_underlying(__a), ::cuda::std::to_underlying(__b));
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___CMATH_CEIL_DIV_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/round_up.h b/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/round_up.h
new file mode 100644
index 00000000..91e8f3f3
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__cmath/round_up.h
@@ -0,0 +1,104 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___CMATH_ROUND_UP_H
+#define _CUDA___CMATH_ROUND_UP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__cmath/ceil_div.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/common_type.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_signed.h>
+#include <cuda/std/__type_traits/make_unsigned.h>
+#include <cuda/std/__utility/to_underlying.h>
+#include <cuda/std/limits>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+//! @brief Round the number \p __a to the next multiple of \p __b
+//! @param __a The input number
+//! @param __b The multiplicand
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_integral_v<_Tp> _CCCL_AND ::cuda::std::is_integral_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<_Tp, _Up> round_up(const _Tp __a, const _Up __b) noexcept
+{
+  _CCCL_ASSERT(__b > _Up{0}, "cuda::round_up: 'b' must be positive");
+  if constexpr (::cuda::std::is_signed_v<_Tp>)
+  {
+    _CCCL_ASSERT(__a >= _Tp{0}, "cuda::round_up: 'a' must be non negative");
+  }
+  using _Common = ::cuda::std::common_type_t<_Tp, _Up>;
+  using _Prom   = decltype(_Tp{} / _Up{});
+  auto __c      = ::cuda::ceil_div(static_cast<_Prom>(__a), static_cast<_Prom>(__b));
+  _CCCL_ASSERT(static_cast<_Common>(__c) <= ::cuda::std::numeric_limits<_Common>::max() / static_cast<_Common>(__b),
+               "cuda::round_up: result overflow");
+  return static_cast<_Common>(static_cast<_Prom>(__c) * static_cast<_Prom>(__b));
+}
+
+//! @brief Round the number \p __a to the next multiple of \p __b
+//! @param __a The input number
+//! @param __b The multiplicand
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_integral_v<_Tp> _CCCL_AND ::cuda::std::is_enum_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<_Tp, ::cuda::std::underlying_type_t<_Up>>
+round_up(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::round_up(__a, ::cuda::std::to_underlying(__b));
+}
+
+//! @brief Round the number \p __a to the next multiple of \p __b
+//! @param __a The input number
+//! @param __b The multiplicand
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_enum_v<_Tp> _CCCL_AND ::cuda::std::is_integral_v<_Up>)
+[[nodiscard]] _CCCL_API constexpr ::cuda::std::common_type_t<::cuda::std::underlying_type_t<_Tp>, _Up>
+round_up(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::round_up(::cuda::std::to_underlying(__a), __b);
+}
+
+//! @brief Round the number \p __a to the next multiple of \p __b
+//! @param __a The input number
+//! @param __b The multiplicand
+//! @pre \p __a must be non-negative
+//! @pre \p __b must be positive
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES(::cuda::std::is_enum_v<_Tp> _CCCL_AND ::cuda::std::is_enum_v<_Up>)
+[[nodiscard]]
+_CCCL_API constexpr ::cuda::std::common_type_t<::cuda::std::underlying_type_t<_Tp>, ::cuda::std::underlying_type_t<_Up>>
+round_up(const _Tp __a, const _Up __b) noexcept
+{
+  return ::cuda::round_up(::cuda::std::to_underlying(__a), ::cuda::std::to_underlying(__b));
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___CMATH_ROUND_UP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__device/compute_capability.h b/qwen3_6_scripts/cccl_preload/include/cuda/__device/compute_capability.h
new file mode 100644
index 00000000..fcf7407d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__device/compute_capability.h
@@ -0,0 +1,272 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___DEVICE_COMPUTE_CAPABILITY_H
+#define _CUDA___DEVICE_COMPUTE_CAPABILITY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__fwd/devices.h>
+#include <cuda/std/__fwd/format.h>
+#include <cuda/std/__type_traits/always_false.h>
+#include <cuda/std/__utility/to_underlying.h>
+#include <cuda/std/array>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+//! @brief Type representing the CUDA compute capability.
+class compute_capability
+{
+public:
+  int __cc_{}; //!< The stored compute capability in format 10 * major + minor.
+
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability() noexcept = default;
+
+  //! @brief Constructs the object from compute capability \c __cc. The expected format is 10 * major + minor.
+  //!
+  //! @param __cc Compute capability.
+  _CCCL_HOST_DEVICE_API explicit constexpr compute_capability(int __cc) noexcept
+      : __cc_{__cc}
+  {}
+
+  //! @brief Constructs the object by combining the \c __major and \c __minor compute capability.
+  //!
+  //! @param __major The major compute capability.
+  //! @param __minor The minor compute capability. Must be less than 10.
+  _CCCL_HOST_DEVICE_API constexpr compute_capability(int __major, int __minor) noexcept
+      : __cc_{10 * __major + __minor}
+  {
+    _CCCL_ASSERT(__minor < 10, "invalid minor compute capability");
+  }
+
+  //! @brief Constructs the object from the architecture id.
+  //!
+  //! @param __arch_id The architecture id.
+  _CCCL_HOST_DEVICE_API explicit constexpr compute_capability(arch_id __arch_id) noexcept
+  {
+    const auto __val = ::cuda::std::to_underlying(__arch_id);
+    if (__val > __arch_specific_id_multiplier)
+    {
+      __cc_ = __val / __arch_specific_id_multiplier;
+    }
+    else
+    {
+      __cc_ = __val;
+    }
+  }
+
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability(const compute_capability&) noexcept = default;
+
+  _CCCL_HIDE_FROM_ABI constexpr compute_capability& operator=(const compute_capability& __other) noexcept = default;
+
+  //! @brief Gets the stored compute capability.
+  //!
+  //! @return The stored compute capability in format 10 * major + minor.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API constexpr int get() const noexcept
+  {
+    return __cc_;
+  }
+
+  //! @brief Gets the major compute capability.
+  //!
+  //! @return Major compute capability.
+  //!
+  //! @deprecated This symbol is deprecated because it collides with major(...) macro defined in <sys/sysmacros.h> and
+  //! will be removed in next major release. Use cc.major_cap() instead.
+  [[nodiscard]]
+  CCCL_DEPRECATED_BECAUSE("This symbol is deprecated because it collides with major(...) macro defined in "
+                          "<sys/sysmacros.h> and will be removed in next major release. Use cc.major_cap() instead.")
+  _CCCL_HOST_DEVICE_API constexpr int major() const noexcept
+  {
+    return major_cap();
+  }
+
+  //! @brief Gets the major compute capability.
+  //!
+  //! @return Major compute capability.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API constexpr int major_cap() const noexcept
+  {
+    return __cc_ / 10;
+  }
+
+  //! @brief Gets the minor compute capability.
+  //!
+  //! @return Minor compute capability. The value is always less than 10.
+  //!
+  //! @deprecated This symbol is deprecated because it collides with minor(...) macro defined in <sys/sysmacros.h> and
+  //! will be removed in next major release. Use cc.minor_cap() instead.
+  [[nodiscard]]
+  CCCL_DEPRECATED_BECAUSE("This symbol is deprecated because it collides with minor(...) macro defined in "
+                          "<sys/sysmacros.h> and will be removed in next major release. Use cc.minor_cap() instead.")
+  _CCCL_HOST_DEVICE_API constexpr int minor() const noexcept
+  {
+    return minor_cap();
+  }
+
+  //! @brief Gets the minor compute capability.
+  //!
+  //! @return Minor compute capability. The value is always less than 10.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API constexpr int minor_cap() const noexcept
+  {
+    return __cc_ % 10;
+  }
+
+  //! @brief Conversion operator to \c int.
+  //!
+  //! @return The stored compute capability in format 10 * major + minor.
+  _CCCL_HOST_DEVICE_API explicit constexpr operator int() const noexcept
+  {
+    return __cc_;
+  }
+
+  //! @brief Equality operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator==(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ == __rhs.__cc_;
+  }
+
+  //! @brief Inequality operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator!=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ != __rhs.__cc_;
+  }
+
+  //! @brief Less than operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator<(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ < __rhs.__cc_;
+  }
+
+  //! @brief Less than or equal to operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator<=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ <= __rhs.__cc_;
+  }
+
+  //! @brief Greater than operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator>(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ > __rhs.__cc_;
+  }
+
+  //! @brief Greater than or equal to operator.
+  [[nodiscard]] _CCCL_HOST_DEVICE_API friend constexpr bool
+  operator>=(compute_capability __lhs, compute_capability __rhs) noexcept
+  {
+    return __lhs.__cc_ >= __rhs.__cc_;
+  }
+};
+
+template <int... _Vs>
+[[nodiscard]] _CCCL_HOST_DEVICE_API _CCCL_CONSTEVAL auto __make_all_compute_capabilities() noexcept
+{
+  return ::cuda::std::array{compute_capability{_Vs}...};
+}
+
+[[nodiscard]] _CCCL_HOST_DEVICE_API _CCCL_CONSTEVAL auto __all_compute_capabilities() noexcept
+{
+  return ::cuda::__make_all_compute_capabilities<_CCCL_KNOWN_CUDA_ARCH_LIST>();
+}
+
+#if _CCCL_CUDA_COMPILATION()
+template <int... _Vs>
+[[nodiscard]] _CCCL_HOST_DEVICE_API _CCCL_CONSTEVAL auto __make_cc_list() noexcept
+{
+#  if defined(__CUDA_ARCH_LIST__)
+  return ::cuda::std::array{compute_capability{_Vs / 10}...};
+#  elif defined(NV_TARGET_SM_INTEGER_LIST)
+  return ::cuda::std::array{compute_capability{_Vs}...};
+#  else // ^^^ has arch list ^^^ / vvv no arch list vvv
+  static_assert(::cuda::std::__always_false_v<decltype(sizeof...(_Vs))>,
+                "This function can be instantiated only when __CUDA_ARCH_LIST__ or NV_TARGET_SM_INTEGER_LIST are "
+                "defined");
+#  endif // ^^^ no arch list ^^^
+}
+
+[[nodiscard]] _CCCL_HOST_DEVICE_API _CCCL_CONSTEVAL auto __target_compute_capabilities() noexcept
+{
+#  if defined(__CUDA_ARCH_LIST__)
+  return ::cuda::__make_cc_list<__CUDA_ARCH_LIST__>();
+#  elif defined(NV_TARGET_SM_INTEGER_LIST)
+  return ::cuda::__make_cc_list<NV_TARGET_SM_INTEGER_LIST>();
+#  else // ^^^ has arch list ^^^ / vvv no arch list vvv
+  // Fallback to a list of all compute capabilities.
+  return ::cuda::__all_compute_capabilities();
+#  endif // ^^^ no arch list ^^^
+}
+#endif // _CCCL_CUDA_COMPILATION()
+
+_CCCL_END_NAMESPACE_CUDA
+
+#if __cpp_lib_format >= 201907L
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _CharT>
+struct formatter<::cuda::compute_capability, _CharT> : private formatter<int, _CharT>
+{
+  template <class _ParseCtx>
+  _CCCL_HOST_API constexpr auto parse(_ParseCtx& __ctx)
+  {
+    return __ctx.begin();
+  }
+
+  template <class _FmtCtx>
+  _CCCL_HOST_API auto format(const ::cuda::compute_capability& __cc, _FmtCtx& __ctx) const
+  {
+    return formatter<int, _CharT>::format(__cc.get(), __ctx);
+  }
+};
+
+_CCCL_END_NAMESPACE_STD
+#endif // __cpp_lib_format >= 201907L
+
+// todo: specialize cuda::std::formatter for cuda::compute_capability
+
+#if _CCCL_CUDA_COMPILATION()
+
+_CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
+
+//! @brief Returns the \c cuda::compute_capability that is currently being compiled.
+//!
+//! @note This API cannot be used in constexpr context when compiling with nvc++ in CUDA mode.
+[[nodiscard]] _CCCL_DEVICE_API inline _CCCL_TARGET_CONSTEXPR ::cuda::compute_capability
+current_compute_capability() noexcept
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC)
+  return ::cuda::compute_capability{__builtin_current_device_sm()};
+#  elif _CCCL_DEVICE_COMPILATION()
+  return ::cuda::compute_capability{__CUDA_ARCH__ / 10};
+#  else // ^^^ _CCCL_DEVICE_COMPILATION() ^^^ / vvv !_CCCL_DEVICE_COMPILATION() vvv
+  return {};
+#  endif // ^^^ !_CCCL_DEVICE_COMPILATION() ^^^
+}
+
+_CCCL_END_NAMESPACE_CUDA_DEVICE
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___DEVICE_COMPUTE_CAPABILITY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/complex.h b/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/complex.h
new file mode 100644
index 00000000..bc299c1c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/complex.h
@@ -0,0 +1,48 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___FWD_COMPLEX_H
+#define _CUDA___FWD_COMPLEX_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+template <class _Tp>
+class _CCCL_TYPE_VISIBILITY_DEFAULT complex;
+
+// __is_cuda_complex_v
+
+template <class _Tp>
+inline constexpr bool __is_cuda_complex_v = false;
+template <class _Tp>
+inline constexpr bool __is_cuda_complex_v<const _Tp> = __is_cuda_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_complex_v<volatile _Tp> = __is_cuda_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_complex_v<const volatile _Tp> = __is_cuda_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_complex_v<complex<_Tp>> = true;
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___FWD_COMPLEX_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/devices.h b/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/devices.h
new file mode 100644
index 00000000..3158b66e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__fwd/devices.h
@@ -0,0 +1,47 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___FWD_DEVICES_H
+#define _CUDA___FWD_DEVICES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/span.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+#if _CCCL_HAS_CTK()
+class __physical_device;
+class device_ref;
+template <::cudaDeviceAttr _Attr>
+struct __dev_attr;
+#endif // _CCCL_HAS_CTK()
+
+struct arch_traits_t;
+class compute_capability;
+enum class arch_id : int;
+
+inline constexpr int __arch_specific_id_multiplier = 100000;
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___FWD_DEVICES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__memory/address_space.h b/qwen3_6_scripts/cccl_preload/include/cuda/__memory/address_space.h
new file mode 100644
index 00000000..383b71e3
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__memory/address_space.h
@@ -0,0 +1,259 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___MEMORY_ADDRESS_SPACE_H
+#define _CUDA___MEMORY_ADDRESS_SPACE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_CUDA_COMPILATION()
+
+#  include <cuda/std/__memory/addressof.h>
+#  include <cuda/std/__utility/to_underlying.h>
+
+#  include <nv/target>
+
+#  include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
+
+//! @brief Address space enumeration for CUDA device code.
+//!
+//!        See https://docs.nvidia.com/cuda/parallel-thread-execution/#state-spaces for more details.
+enum class address_space
+{
+  global, //!< Global state space
+  shared, //!< Shared state space
+  constant, //!< Constant state space
+  local, //!< Local state space
+  grid_constant, //!< Kernel function parameter in the parameter state space
+  cluster_shared, //!< Cluster shared window within the shared state space
+  __max,
+};
+
+[[nodiscard]] _CCCL_DEVICE_API constexpr bool __cccl_is_valid_address_space(address_space __space) noexcept
+{
+  const auto __v = ::cuda::std::to_underlying(__space);
+  return __v >= 0 && __v < ::cuda::std::to_underlying(address_space::__max);
+}
+
+[[nodiscard]] _CCCL_DEVICE_API inline bool __is_smem_valid_ptr(const void* __ptr) noexcept
+{
+  NV_IF_TARGET(NV_PROVIDES_SM_90, (return __ptr != nullptr;), (return true;));
+}
+
+//! @brief Checks if the given pointer is from the specified address state space.
+//! @param __ptr The address to check.
+//! @param __space The address state space to check against.
+//! @return `true` if the pointer is from the specified address space, `false` otherwise.
+[[nodiscard]] _CCCL_DEVICE_API inline bool __internal_is_address_from(const void* __ptr, address_space __space) noexcept
+{
+  _CCCL_ASSERT(::cuda::device::__cccl_is_valid_address_space(__space), "invalid address space");
+  // NVCC and NVRTC < 12.3 have problems tracking the address space of pointers, fallback to inline PTX for them
+  switch (__space)
+  {
+    case address_space::global: {
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3)
+      unsigned __ret;
+      asm volatile(
+        "{\n\t"
+        "  .reg .pred p;\n\t"
+        "  isspacep.global p, %1;\n\t"
+        "  selp.u32 %0, 1, 0, p;\n\t"
+        "}\n\t"
+        : "=r"(__ret)
+        : "l"(__ptr));
+      return static_cast<bool>(__ret);
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^ /
+        // vvv !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) vvv
+      bool __p = static_cast<bool>(::__isGlobal(__ptr));
+      if (__p)
+      {
+        _CCCL_ASSUME(__p);
+      }
+      return __p;
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^
+    }
+    case address_space::constant: {
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3)
+      unsigned __ret;
+      asm volatile(
+        "{\n\t"
+        "  .reg .pred p;\n\t"
+        "  isspacep.const p, %1;\n\t"
+        "  selp.u32 %0, 1, 0, p;\n\t"
+        "}\n\t"
+        : "=r"(__ret)
+        : "l"(__ptr));
+      return static_cast<bool>(__ret);
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^ /
+        // vvv !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) vvv
+      bool __p = static_cast<bool>(::__isConstant(__ptr));
+      if (__p)
+      {
+        _CCCL_ASSUME(__p);
+      }
+      return __p;
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^
+    }
+    case address_space::local: {
+      // __isLocal is buggy until CUDA 13.1, see nvbug 5254298
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 13, 1) || _CCCL_CUDA_COMPILER(NVRTC, <, 13, 1)
+      unsigned __ret;
+      asm volatile(
+        "{\n\t"
+        "  .reg .pred p;\n\t"
+        "  isspacep.local p, %1;\n\t"
+        "  selp.u32 %0, 1, 0, p;\n\t"
+        "}\n\t"
+        : "=r"(__ret)
+        : "l"(__ptr));
+      return static_cast<bool>(__ret);
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVCC, <, 13, 1) || _CCCL_CUDA_COMPILER(NVRTC, <, 13, 1) ^^^ /
+        // vvv !_CCCL_CUDA_COMPILER(NVCC) && !_CCCL_CUDA_COMPILER(NVRTC) vvv
+      bool __p = static_cast<bool>(::__isLocal(__ptr));
+      if (__p)
+      {
+        _CCCL_ASSUME(__p);
+      }
+      return __p;
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVCC) && !_CCCL_CUDA_COMPILER(NVRTC) ^^^
+    }
+    case address_space::grid_constant: {
+#  if _CCCL_CUDA_COMPILER(NVCC, >=, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, >=, 12, 3)
+      NV_IF_ELSE_TARGET(
+        NV_PROVIDES_SM_70,
+        (bool __p = static_cast<bool>(::__isGridConstant(__ptr)); //
+         if (__p) //
+         { //
+           _CCCL_ASSUME(__p); //
+         } //
+         return __p;),
+        (return false;))
+#  else // ^^^ has functional __isGridConstant() ^^^ / vvv no functional __isGridConstant() vvv
+      NV_IF_ELSE_TARGET(
+        NV_PROVIDES_SM_70,
+        (unsigned __ret; //
+         asm volatile("{\n\t"
+                      "  .reg .pred p;\n\t"
+                      "  isspacep.param p, %1;\n\t"
+                      "  selp.u32 %0, 1, 0, p;\n\t"
+                      "}\n\t" : "=r"(__ret) : "l"(__ptr));
+         return static_cast<bool>(__ret);),
+        (return false;))
+#  endif // ^^^ no functional __isGridConstant() ^^^
+    }
+    case address_space::cluster_shared: {
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3)
+      NV_IF_ELSE_TARGET(
+        NV_PROVIDES_SM_90,
+        (unsigned __ret; //
+         asm volatile("{\n\t"
+                      "  .reg .pred p;\n\t"
+                      "  isspacep.shared::cluster p, %1;\n\t"
+                      "  selp.u32 %0, 1, 0, p;\n\t"
+                      "}\n\t" : "=r"(__ret) : "l"(__ptr));
+         return static_cast<bool>(__ret);),
+        ([[fallthrough]]; /* to `case shared:` */))
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^ /
+        // vvv !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) vvv
+      NV_IF_ELSE_TARGET(
+        NV_PROVIDES_SM_90,
+        (bool __p = static_cast<bool>(::__isClusterShared(__ptr)); //
+         if (__p) //
+         { //
+           _CCCL_ASSUME(__p); //
+         } //
+         return __p;),
+        ([[fallthrough]]; /* to `case shared:` */))
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^
+    }
+    case address_space::shared: {
+      // smem can start at address 0x0 before sm_90
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3)
+      unsigned __ret;
+      asm volatile(
+        "{\n\t"
+        "  .reg .pred p;\n\t"
+        "  isspacep.shared p, %1;\n\t"
+        "  selp.u32 %0, 1, 0, p;\n\t"
+        "}\n\t"
+        : "=r"(__ret)
+        : "l"(__ptr));
+      return static_cast<bool>(__ret);
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVCC, <, 12, 3) || _CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^ /
+        // vvv !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) vvv
+      bool __p = static_cast<bool>(::__isShared(__ptr));
+      if (__p)
+      {
+        _CCCL_ASSUME(__p);
+      }
+      return __p;
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVCC, <, 12, 3) && !_CCCL_CUDA_COMPILER(NVRTC, <, 12, 3) ^^^
+    }
+    default:
+      return false;
+  }
+}
+
+//! @brief Checks if the given pointer is from the specified address state space.
+//! @param __ptr The address to check.
+//! @param __space The address state space to check against.
+//! @return `true` if the pointer is from the specified address space, `false` otherwise.
+[[nodiscard]] _CCCL_DEVICE_API inline bool is_address_from(const void* __ptr, address_space __space) noexcept
+{
+  // The debug assertions intentionally differ but compile out in release builds.
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if (__space == address_space::shared)
+  {
+    _CCCL_ASSERT(::cuda::device::__is_smem_valid_ptr(__ptr), "invalid pointer");
+  }
+  else
+  {
+    _CCCL_ASSERT(__ptr != nullptr, "invalid pointer");
+  }
+  // NOLINTEND(bugprone-branch-clone)
+  return ::cuda::device::__internal_is_address_from(__ptr, __space);
+}
+
+//! @brief Checks if the given pointer is from the specified address state space.
+//! @param __ptr The address to check.
+//! @param __space The address state space to check against.
+//! @return `true` if the pointer is from the specified address space, `false` otherwise.
+[[nodiscard]] _CCCL_DEVICE_API inline bool is_address_from(const volatile void* __ptr, address_space __space) noexcept
+{
+  return ::cuda::device::is_address_from(const_cast<const void*>(__ptr), __space);
+}
+
+//! @brief Checks if the given object is from the specified address state space.
+//! @param __obj The object to check.
+//! @param __space The address state space to check against.
+//! @return `true` if the object is from the specified address space, `false` otherwise.
+template <class _Tp>
+[[nodiscard]] _CCCL_DEVICE_API inline bool is_object_from(_Tp& __obj, address_space __space) noexcept
+{
+  return ::cuda::device::is_address_from(::cuda::std::addressof(__obj), __space);
+}
+
+_CCCL_END_NAMESPACE_CUDA_DEVICE
+
+#  include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+#endif // _CUDA___MEMORY_ADDRESS_SPACE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__memory/check_address.h b/qwen3_6_scripts/cccl_preload/include/cuda/__memory/check_address.h
new file mode 100644
index 00000000..cd223c26
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__memory/check_address.h
@@ -0,0 +1,111 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___MEMORY_IS_VALID_ADDRESS
+#define _CUDA___MEMORY_IS_VALID_ADDRESS
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+#include <cuda/std/cstdint>
+#if _CCCL_CUDA_COMPILATION()
+#  include <cuda/__memory/address_space.h>
+#  include <cuda/__ptx/instructions/get_sreg.h>
+#endif // _CCCL_CUDA_COMPILATION()
+
+#include <nv/target>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CUDA_COMPILATION()
+
+_CCCL_BEGIN_NAMESPACE_CUDA_DEVICE
+
+[[nodiscard]] _CCCL_DEVICE_API inline bool
+__is_smem_valid_address_range(const void* __ptr, ::cuda::std::size_t __n) noexcept
+{
+  if (!::cuda::device::__is_smem_valid_ptr(__ptr))
+  {
+    return false;
+  }
+  if (!::cuda::device::__internal_is_address_from(__ptr, ::cuda::device::address_space::shared))
+  {
+    return false;
+  }
+  // if __ptr is a shared memory pointer, __ptr + __n must also be a valid shared memory pointer
+  if (!::cuda::device::__internal_is_address_from(
+        reinterpret_cast<const char*>(__ptr) + __n, ::cuda::device::address_space::shared))
+  {
+    return false;
+  }
+  return (__n <= ::cuda::ptx::get_sreg_total_smem_size());
+}
+
+_CCCL_END_NAMESPACE_CUDA_DEVICE
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+[[nodiscard]] _CCCL_API inline bool __is_valid_address_range(const void* __ptr, ::cuda::std::size_t __n) noexcept
+{
+  if (__n == 0)
+  {
+    return false;
+  }
+
+  // use (~::cuda::std::uintptr_t{0}) instead of cuda::std::numeric_limits<cuda::std::uintptr_t>::max() to avoid
+  // circular dependency because:
+  //    numeric_limits -> bit_cast -> cstring -> check_address
+  // <cuda/std/__utility/cmp.h> also includes cuda/std/limits
+  const auto __limit = (~::cuda::std::uintptr_t{0}) - static_cast<::cuda::std::uintptr_t>(__n);
+
+  if (reinterpret_cast<::cuda::std::uintptr_t>(__ptr) > __limit)
+  {
+    return false;
+  }
+  NV_IF_TARGET(NV_IS_DEVICE, ({
+                 if (::cuda::device::__internal_is_address_from(__ptr, ::cuda::device::address_space::shared)
+                     && !::cuda::device::__is_smem_valid_address_range(__ptr, __n))
+                 {
+                   return false;
+                 }
+               }));
+  return (__ptr != nullptr);
+}
+
+[[nodiscard]] _CCCL_API inline bool __is_valid_address(const void* __ptr) noexcept
+{
+  return ::cuda::__is_valid_address_range(__ptr, 0);
+}
+
+[[nodiscard]] _CCCL_API inline bool
+__are_ptrs_overlapping(const void* __ptr_lhs, const void* __ptr_rhs, ::cuda::std::size_t __n) noexcept
+{
+  const auto __ptr1_start = static_cast<const char*>(__ptr_lhs);
+  const auto __ptr2_start = static_cast<const char*>(__ptr_rhs);
+  const auto __ptr1_end   = __ptr1_start + __n;
+  const auto __ptr2_end   = __ptr2_start + __n;
+  return __ptr1_start < __ptr2_end && __ptr2_start < __ptr1_end;
+}
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA___MEMORY_IS_VALID_ADDRESS
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx.h b/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx.h
new file mode 100644
index 00000000..5e7780c9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx.h
@@ -0,0 +1,150 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___NVTX_NVTX_H
+#define _CUDA___NVTX_NVTX_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#ifdef _CCCL_DOXYGEN_INVOKED // Only parse this during doxygen passes:
+//! When this macro is defined, no NVTX ranges are emitted by CCCL
+#  define CCCL_DISABLE_NVTX
+#endif // _CCCL_DOXYGEN_INVOKED
+
+#define _CCCL_HAS_NVTX3() 0
+
+// Enable the functionality of this header if:
+// * The NVTX3 C API is available in CTK
+// * NVTX is not explicitly disabled (via CCCL_DISABLE_NVTX or NVTX_DISABLE)
+// * the compiler is not nvc++ (NVTX3 uses module as an identifier, which trips up NVHPC, fixed in CTK >= 13.0)
+// * the compiler is not NVRTC
+#if __has_include(<nvtx3/nvToolsExt.h>) && !defined(CCCL_DISABLE_NVTX) && !defined(NVTX_DISABLE) \
+                      && (!_CCCL_COMPILER(NVHPC) || _CCCL_CTK_AT_LEAST(13, 0))                       \
+                      && !_CCCL_COMPILER(NVRTC)
+
+// Since NVTX 3.2, the NVTX headers can declare themselves as system headers by declaring the following macro:
+#  ifdef NVTX_AS_SYSTEM_HEADER
+#    define NVTX_AS_SYSTEM_HEADER_DEFINED_BY_USER
+#  else // NVTX_AS_SYSTEM_HEADER
+#    define NVTX_AS_SYSTEM_HEADER
+#  endif // NVTX_AS_SYSTEM_HEADER
+
+// Include our NVTX3 C++ wrapper if not available from the CTK or not provided by the user
+// Note: NVTX3 is available in the CTK since 12.9, so we can drop our copy once this is the minimum supported version
+#  if __has_include(<nvtx3/nvtx3.hpp>)
+#    include <nvtx3/nvtx3.hpp>
+#  else // __has_include(<nvtx3/nvtx3.hpp>)
+#    include <cuda/__nvtx/nvtx3.h>
+#  endif // __has_include(<nvtx3/nvtx3.hpp>)
+
+#  ifndef NVTX_AS_SYSTEM_HEADER_DEFINED_BY_USER
+#    undef NVTX_AS_SYSTEM_HEADER
+#  endif // NVTX_AS_SYSTEM_HEADER_DEFINED_BY_USER
+#  undef NVTX_AS_SYSTEM_HEADER_DEFINED_BY_USER
+
+// We expect the NVTX3 V1 C++ API to be available when nvtx3.hpp is available. This should work, because newer versions
+// of NVTX3 will continue to declare previous API versions. See also:
+// https://github.com/NVIDIA/NVTX/blob/release-v3/c/include/nvtx3/nvtx3.hpp#L2835-L2841.
+#  ifdef NVTX3_CPP_DEFINITIONS_V1_0
+#    undef _CCCL_HAS_NVTX3
+#    define _CCCL_HAS_NVTX3() 1
+#  else // NVTX3_CPP_DEFINITIONS_V1_0
+// If this happens NVTX3 changed in a way we did not anticipate, and we need to get in touch with them
+#    if _CCCL_COMPILER(MSVC)
+#      pragma message( \
+        "warning: nvtx3.h is available but does not define the V1 API. This is odd. Please open a GitHub issue at: https://github.com/NVIDIA/cccl/issues.")
+#    else
+#      warning nvtx3.h is available but does not define the V1 API. This is odd. Please open a GitHub issue at: https://github.com/NVIDIA/cccl/issues.
+#    endif
+#  endif // NVTX3_CPP_DEFINITIONS_V1_0
+#endif // __has_include(<nvtx3/nvToolsExt.h>) && !defined(CCCL_DISABLE_NVTX) && !defined(NVTX_DISABLE) &&
+       // (!_CCCL_COMPILER(NVHPC)) && !_CCCL_COMPILER(NVRTC)
+
+#if _CCCL_HAS_NVTX3()
+#  include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+struct __nvtx_cccl_domain
+{
+  static constexpr const char* name{"CCCL"};
+};
+
+using __nvtx_cccl_range = ::nvtx3::v1::scoped_range_in<__nvtx_cccl_domain>;
+
+// this type ensures that no NVTX range code is emitted in device code
+struct __nvtx_cccl_optional_range_host_only
+{
+  bool __engaged = false;
+  alignas(__nvtx_cccl_range) unsigned char __storage[sizeof(__nvtx_cccl_range)];
+
+  __nvtx_cccl_optional_range_host_only() = default;
+
+  _CCCL_HOST_API void __start(const ::nvtx3::v1::event_attributes& __attributes)
+  {
+    ::new (__storage) __nvtx_cccl_range(__attributes);
+    __engaged = true;
+  }
+
+  _CCCL_API ~__nvtx_cccl_optional_range_host_only()
+  {
+    NV_IF_TARGET(NV_IS_HOST, ({
+                   if (__engaged)
+                   {
+                     reinterpret_cast<__nvtx_cccl_range*>(__storage)->~__nvtx_cccl_range();
+                   }
+                 }));
+  }
+};
+_CCCL_END_NAMESPACE_CUDA
+
+// Hook for the NestedNVTXRangeGuard from the unit tests
+#  ifndef _CCCL_BEFORE_NVTX_RANGE_SCOPE
+#    define _CCCL_BEFORE_NVTX_RANGE_SCOPE(name)
+#  endif // !CCCL_DETAIL_BEFORE_NVTX_RANGE_SCOPE
+
+#  if _CCCL_HOST_COMPILATION()
+// Conditionally inserts a NVTX range starting here until the end of the current function scope in host code. Does
+// nothing in device code.
+// The __nvtx_cccl_optional_range_host_only type (a simplified optional<T>) is needed to defer the construction of the
+// NVTX range and message string registration (static variables) into a region running only on the host, while
+// preserving the semantic scope where the range is declared.
+#    define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name)                                                             \
+      _CCCL_BEFORE_NVTX_RANGE_SCOPE(name)                                                                          \
+      ::cuda::__nvtx_cccl_optional_range_host_only __cuda_nvtx3_range;                                             \
+      NV_IF_TARGET(                                                                                                \
+        NV_IS_HOST, ({                                                                                             \
+          static const ::nvtx3::v1::registered_string_in<::cuda::__nvtx_cccl_domain> __cuda_nvtx3_func_name{name}; \
+          static const ::nvtx3::v1::event_attributes __cuda_nvtx3_func_attr{__cuda_nvtx3_func_name};               \
+          if (condition)                                                                                           \
+          {                                                                                                        \
+            __cuda_nvtx3_range.__start(__cuda_nvtx3_func_attr);                                                    \
+          }                                                                                                        \
+        }))
+#  else // ^^^ _CCCL_HOST_COMPILATION() ^^^ / vvv !_CCCL_HOST_COMPILATION() vvv
+#    define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name)
+#  endif // ^^^ !_CCCL_HOST_COMPILATION() ^^^
+
+#  define _CCCL_NVTX_RANGE_SCOPE(name) _CCCL_NVTX_RANGE_SCOPE_IF(true, name)
+
+#  include <cuda/std/__cccl/epilogue.h>
+
+#else // _CCCL_HAS_NVTX3()
+#  define _CCCL_NVTX_RANGE_SCOPE_IF(condition, name)
+#  define _CCCL_NVTX_RANGE_SCOPE(name)
+#endif // _CCCL_HAS_NVTX3()
+
+#endif // _CUDA___NVTX_NVTX_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx3.h b/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx3.h
new file mode 100644
index 00000000..6e49c0f4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__nvtx/nvtx3.h
@@ -0,0 +1,2977 @@
+/*
+ * SPDX-FileCopyrightText: Copyright (c) 2020-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+ * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * Licensed under the Apache License v2.0 with LLVM Exceptions.
+ * See https://nvidia.github.io/NVTX/LICENSE.txt for license information.
+ */
+
+/* Temporary helper #defines, #undef'ed at end of header */
+#define NVTX3_CPP_VERSION_MAJOR 1
+#define NVTX3_CPP_VERSION_MINOR 0
+
+/* This section handles the decision of whether to provide unversioned symbols.
+ * If NVTX3_CPP_REQUIRE_EXPLICIT_VERSION is #defined, unversioned symbols are
+ * not provided, and explicit-version symbols such as nvtx3::v1::scoped_range
+ * and NVTX3_V1_FUNC_RANGE must be used.  By default, the first #include of this
+ * header will define the unversioned symbols such as nvtx3::scoped_range and
+ * NVTX3_FUNC_RANGE.  Subsequently including a different major version of this
+ * header without #defining NVTX3_CPP_REQUIRE_EXPLICIT_VERSION triggers an error
+ * since the symbols would conflict.  Subsequently including of a different
+ * minor version within the same major version is allowed. Functionality of
+ * minor versions is cumulative, regardless of include order.
+ *
+ * Since NVTX3_CPP_REQUIRE_EXPLICIT_VERSION allows all combinations of versions
+ * to coexist without problems within a translation unit, the recommended best
+ * practice for instrumenting header-based libraries with NVTX C++ Wrappers is
+ * is to #define NVTX3_CPP_REQUIRE_EXPLICIT_VERSION before including nvtx3.hpp,
+ * #undef it afterward, and only use explicit-version symbols.  This is not
+ * necessary in common cases, such as instrumenting a standalone application, or
+ * static/shared libraries in .cpp files or headers private to those projects.
+ */
+/* clang-format off */
+#if !defined(NVTX3_CPP_REQUIRE_EXPLICIT_VERSION)
+  /* Define macro used by all definitions in this header to indicate the
+   * unversioned symbols should be defined in addition to the versioned ones.
+   */
+  #define NVTX3_INLINE_THIS_VERSION
+
+  #if !defined(NVTX3_CPP_INLINED_VERSION_MAJOR)
+    /* First occurrence of this header in the translation unit.  Define macros
+     * indicating which version shall be used for unversioned symbols.
+     */
+
+    /**
+     * @brief Semantic major version number for NVTX C++ wrappers of unversioned symbols
+     *
+     * Breaking changes may occur between major versions, and different major versions
+     * cannot provide unversioned symbols in the same translation unit (.cpp file).
+     *
+     * Note: If NVTX3_CPP_REQUIRE_EXPLICIT_VERSION is defined, this macro is not defined.
+     *
+     * Not to be confused with the version number of the NVTX core library.
+     */
+    #define NVTX3_CPP_INLINED_VERSION_MAJOR 1  // NVTX3_CPP_VERSION_MAJOR
+
+    /**
+     * @brief Semantic minor version number for NVTX C++ wrappers of unversioned symbols
+     *
+     * No breaking changes occur between minor versions -- minor version changes within
+     * a major version are purely additive.
+     *
+     * Note: If NVTX3_CPP_REQUIRE_EXPLICIT_VERSION is defined, this macro is not defined.
+     *
+     * Not to be confused with the version number of the NVTX core library.
+     */
+    #define NVTX3_CPP_INLINED_VERSION_MINOR 0  // NVTX3_CPP_VERSION_MINOR
+  #elif NVTX3_CPP_INLINED_VERSION_MAJOR != NVTX3_CPP_VERSION_MAJOR
+    /* Unsupported case -- cannot define unversioned symbols for different major versions
+     * in the same translation unit.
+     */
+    #error \
+      "Two different major versions of the NVTX C++ Wrappers are being included in a single .cpp file, with unversioned symbols enabled in both.  Only one major version can enable unversioned symbols in a .cpp file.  To disable unversioned symbols, #define NVTX3_CPP_REQUIRE_EXPLICIT_VERSION before #including nvtx3.hpp, and use the explicit-version symbols instead -- this is the preferred way to use nvtx3.hpp from a header file."
+  #elif (NVTX3_CPP_INLINED_VERSION_MAJOR == NVTX3_CPP_VERSION_MAJOR) && \
+    (NVTX3_CPP_INLINED_VERSION_MINOR < NVTX3_CPP_VERSION_MINOR)
+    /* An older minor version of the same major version already defined unversioned
+     * symbols.  The new features provided in this header will be inlined
+     * redefine the minor version macro to this header's version.
+     */
+    #undef NVTX3_CPP_INLINED_VERSION_MINOR
+    #define NVTX3_CPP_INLINED_VERSION_MINOR 0  // NVTX3_CPP_VERSION_MINOR
+    // else, already have this version or newer, nothing to do
+  #endif
+#endif
+/* clang-format on */
+
+/**
+ * @file nvtx3.hpp
+ *
+ * @brief Provides C++ constructs making the NVTX library safer and easier to
+ * use with zero overhead.
+ */
+
+/**
+ * \mainpage
+ * \tableofcontents
+ *
+ * \section QUICK_START Quick Start
+ *
+ * To add NVTX ranges to your code, use the `nvtx3::scoped_range` RAII object. A
+ * range begins when the object is created, and ends when the object is
+ * destroyed.
+ *
+ * \code{.cpp}
+ * #include "nvtx3.hpp"
+ * void some_function() {
+ *    // Begins a NVTX range with the message "some_function"
+ *    // The range ends when some_function() returns and `r` is destroyed
+ *    nvtx3::scoped_range r{"some_function"};
+ *
+ *    for(int i = 0; i < 6; ++i) {
+ *       nvtx3::scoped_range loop{"loop range"};
+ *       std::this_thread::sleep_for(std::chrono::seconds{1});
+ *    }
+ * } // Range ends when `r` is destroyed
+ * \endcode
+ *
+ * The example code above generates the following timeline view in Nsight
+ * Systems:
+ *
+ * \image html
+ * https://raw.githubusercontent.com/NVIDIA/NVTX/release-v3/docs/images/example_range.png
+ *
+ * Alternatively, use the \ref MACROS like `NVTX3_FUNC_RANGE()` to add
+ * ranges to your code that automatically use the name of the enclosing function
+ * as the range's message.
+ *
+ * \code{.cpp}
+ * #include "nvtx3.hpp"
+ * void some_function() {
+ *    // Creates a range with a message "some_function" that ends when the
+ *    // enclosing function returns
+ *    NVTX3_FUNC_RANGE();
+ *    ...
+ * }
+ * \endcode
+ *
+ *
+ * \section Overview
+ *
+ * The NVTX library provides a set of functions for users to annotate their code
+ * to aid in performance profiling and optimization. These annotations provide
+ * information to tools like Nsight Systems to improve visualization of
+ * application timelines.
+ *
+ * \ref RANGES are one of the most commonly used NVTX constructs for annotating
+ * a span of time. For example, imagine a user wanted to see every time a
+ * function, `my_function`, is called and how long it takes to execute. This can
+ * be accomplished with an NVTX range created on the entry to the function and
+ * terminated on return from `my_function` using the push/pop C APIs:
+ *
+ * \code{.cpp}
+ * void my_function(...) {
+ *    nvtxRangePushA("my_function"); // Begins NVTX range
+ *    // do work
+ *    nvtxRangePop(); // Ends NVTX range
+ * }
+ * \endcode
+ *
+ * One of the challenges with using the NVTX C API is that it requires manually
+ * terminating the end of the range with `nvtxRangePop`. This can be challenging
+ * if `my_function()` has multiple returns or can throw exceptions as it
+ * requires calling `nvtxRangePop()` before all possible return points.
+ *
+ * NVTX C++ solves this inconvenience through the "RAII" technique by providing
+ * a `nvtx3::scoped_range` class that begins a range at construction and ends
+ * the range on destruction. The above example then becomes:
+ *
+ * \code{.cpp}
+ * void my_function(...) {
+ *    nvtx3::scoped_range r{"my_function"}; // Begins NVTX range
+ *    // do work
+ * } // Range ends on exit from `my_function` when `r` is destroyed
+ * \endcode
+ *
+ * The range object `r` is deterministically destroyed whenever `my_function`
+ * returns---ending the NVTX range without manual intervention. For more
+ * information, see \ref RANGES and `nvtx3::scoped_range_in`.
+ *
+ * Another inconvenience of the NVTX C APIs are the several constructs where the
+ * user is expected to initialize an object at the beginning of an application
+ * and reuse that object throughout the lifetime of the application. For example
+ * see domains, categories, and registered messages.
+ *
+ * Example:
+ * \code{.cpp}
+ * nvtxDomainHandle_t D = nvtxDomainCreateA("my domain");
+ * // Reuse `D` throughout the rest of the application
+ * \endcode
+ *
+ * This can be problematic if the user application or library does not have an
+ * explicit initialization function called before all other functions to
+ * ensure that these long-lived objects are initialized before being used.
+ *
+ * NVTX C++ makes use of the "construct on first use" technique to alleviate
+ * this inconvenience. In short, a function local static object is constructed
+ * upon the first invocation of a function and returns a reference to that
+ * object on all future invocations. See the documentation for `nvtx3::domain`,
+ * `nvtx3::named_category`, `nvtx3::registered_string`, and
+ * https://isocpp.org/wiki/faq/ctors#static-init-order-on-first-use for more
+ * information.
+ *
+ * Using construct on first use, the above example becomes:
+ * \code{.cpp}
+ * struct my_domain{ static constexpr char const* name{"my domain"}; };
+ *
+ * // The first invocation of `domain::get` for the type `my_domain` will
+ * // construct a `nvtx3::domain` object and return a reference to it. Future
+ * // invocations simply return a reference.
+ * nvtx3::domain const& D = nvtx3::domain::get<my_domain>();
+ * \endcode
+ * For more information about NVTX and how it can be used, see
+ * https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvtx and
+ * https://devblogs.nvidia.com/cuda-pro-tip-generate-custom-application-profile-timelines-nvtx/
+ * for more information.
+ *
+ * \section RANGES Ranges
+ *
+ * Ranges are used to describe a span of time during the execution of an
+ * application. Common examples are using ranges to annotate the time it takes
+ * to execute a function or an iteration of a loop.
+ *
+ * NVTX C++ uses RAII to automate the generation of ranges that are tied to the
+ * lifetime of objects. Similar to `std::lock_guard` in the C++ Standard
+ * Template Library.
+ *
+ * \subsection scoped_range Scoped Range
+ *
+ * `nvtx3::scoped_range_in` is a class that begins a range upon construction
+ * and ends the range at destruction. This is one of the most commonly used
+ * constructs in NVTX C++ and is useful for annotating spans of time on a
+ * particular thread. These ranges can be nested to arbitrary depths.
+ *
+ * `nvtx3::scoped_range` is an alias for a `nvtx3::scoped_range_in` in the
+ * global NVTX domain. For more information about Domains, see \ref DOMAINS.
+ *
+ * Various attributes of a range can be configured constructing a
+ * `nvtx3::scoped_range_in` with a `nvtx3::event_attributes` object. For
+ * more information, see \ref ATTRIBUTES.
+ *
+ * Example:
+ *
+ * \code{.cpp}
+ * void some_function() {
+ *    // Creates a range for the duration of `some_function`
+ *    nvtx3::scoped_range r{};
+ *
+ *    while(true) {
+ *       // Creates a range for every loop iteration
+ *       // `loop_range` is nested inside `r`
+ *       nvtx3::scoped_range loop_range{};
+ *    }
+ * }
+ * \endcode
+ *
+ * \subsection unique_range Unique Range
+ *
+ * `nvtx3::unique_range` is similar to `nvtx3::scoped_range`, with a few key differences:
+ * - `unique_range` objects can be destroyed in any order whereas `scoped_range` objects must be
+ *    destroyed in exact reverse creation order
+ * - `unique_range` can start and end on different threads
+ * - `unique_range` is movable
+ * - `unique_range` objects can be constructed as heap objects
+ *
+ * There is extra overhead associated with `unique_range` constructs and therefore use of
+ * `nvtx3::scoped_range_in` should be preferred.
+ *
+ * \section MARKS Marks
+ *
+ * `nvtx3::mark` annotates an instantaneous point in time with a "marker".
+ *
+ * Unlike a "range" which has a beginning and an end, a marker is a single event
+ * in an application, such as detecting a problem:
+ *
+ * \code{.cpp}
+ * bool success = do_operation(...);
+ * if (!success) {
+ *    nvtx3::mark("operation failed!");
+ * }
+ * \endcode
+ *
+ * \section DOMAINS Domains
+ *
+ * Similar to C++ namespaces, domains allow for scoping NVTX events. By default,
+ * all NVTX events belong to the "global" domain. Libraries and applications
+ * should scope their events to use a custom domain to differentiate where the
+ * events originate from.
+ *
+ * It is common for a library or application to have only a single domain and
+ * for the name of that domain to be known at compile time. Therefore, Domains
+ * in NVTX C++ are represented by _tag types_.
+ *
+ * For example, to define a custom domain, simply define a new concrete type
+ * (a `class` or `struct`) with a `static` member called `name` that contains
+ * the desired name of the domain.
+ *
+ * \code{.cpp}
+ * struct my_domain{ static constexpr char const* name{"my domain"}; };
+ * \endcode
+ *
+ * For any NVTX C++ construct that can be scoped to a domain, the type
+ * `my_domain` can be passed as an explicit template argument to scope it to
+ * the custom domain.
+ *
+ * The tag type `nvtx3::domain::global` represents the global NVTX domain.
+ *
+ * \code{.cpp}
+ * // By default, `scoped_range_in` belongs to the global domain
+ * nvtx3::scoped_range_in<> r0{};
+ *
+ * // Alias for a `scoped_range_in` in the global domain
+ * nvtx3::scoped_range r1{};
+ *
+ * // `r` belongs to the custom domain
+ * nvtx3::scoped_range_in<my_domain> r{};
+ * \endcode
+ *
+ * When using a custom domain, it is recommended to define type aliases for NVTX
+ * constructs in the custom domain.
+ * \code{.cpp}
+ * using my_scoped_range = nvtx3::scoped_range_in<my_domain>;
+ * using my_registered_string = nvtx3::registered_string_in<my_domain>;
+ * using my_named_category = nvtx3::named_category_in<my_domain>;
+ * \endcode
+ *
+ * See `nvtx3::domain` for more information.
+ *
+ * \section ATTRIBUTES Event Attributes
+ *
+ * NVTX events can be customized with various attributes to provide additional
+ * information (such as a custom message) or to control visualization of the
+ * event (such as the color used). These attributes can be specified per-event
+ * via arguments to a `nvtx3::event_attributes` object.
+ *
+ * NVTX events can be customized via four "attributes":
+ * - \ref COLOR : color used to visualize the event in tools.
+ * - \ref MESSAGES :  Custom message string.
+ * - \ref PAYLOAD :  User-defined numerical value.
+ * - \ref CATEGORY : Intra-domain grouping.
+ *
+ * It is possible to construct a `nvtx3::event_attributes` from any number of
+ * attribute objects (nvtx3::color, nvtx3::message, nvtx3::payload,
+ * nvtx3::category) in any order. If an attribute is not specified, a tool
+ * specific default value is used. See `nvtx3::event_attributes` for more
+ * information.
+ *
+ * \code{.cpp}
+ * // Set message, same as passing nvtx3::message{"message"}
+ * nvtx3::event_attributes attr{"message"};
+ *
+ * // Set message and color
+ * nvtx3::event_attributes attr{"message", nvtx3::rgb{127, 255, 0}};
+ *
+ * // Set message, color, payload, category
+ * nvtx3::event_attributes attr{"message",
+ *                              nvtx3::rgb{127, 255, 0},
+ *                              nvtx3::payload{42},
+ *                              nvtx3::category{1}};
+ *
+ * // Same as above -- can use any order of arguments
+ * nvtx3::event_attributes attr{nvtx3::payload{42},
+ *                              nvtx3::category{1},
+ *                              "message",
+ *                              nvtx3::rgb{127, 255, 0}};
+ *
+ * // Multiple arguments of the same type are allowed, but only the first is
+ * // used -- in this example, payload is set to 42:
+ * nvtx3::event_attributes attr{ nvtx3::payload{42}, nvtx3::payload{7} };
+ *
+ * // Using the nvtx3 namespace in a local scope makes the syntax more succinct:
+ * using namespace nvtx3;
+ * event_attributes attr{"message", rgb{127, 255, 0}, payload{42}, category{1}};
+ * \endcode
+ *
+ * \subsection MESSAGES message
+ *
+ * `nvtx3::message` sets the message string for an NVTX event.
+ *
+ * Example:
+ * \code{.cpp}
+ * // Create an `event_attributes` with the message "my message"
+ * nvtx3::event_attributes attr{nvtx3::message{"my message"}};
+ *
+ * // strings and string literals implicitly assumed to be a `nvtx3::message`
+ * nvtx3::event_attributes attr{"my message"};
+ * \endcode
+ *
+ * \subsubsection REGISTERED_MESSAGE Registered Messages
+ *
+ * Associating a `nvtx3::message` with an event requires copying the contents of
+ * the message every time the message is used, i.e., copying the entire message
+ * string. This may cause non-trivial overhead in performance sensitive code.
+ *
+ * To eliminate this overhead, NVTX allows registering a message string,
+ * yielding a "handle" that is inexpensive to copy that may be used in place of
+ * a message string. When visualizing the events, tools such as Nsight Systems
+ * will take care of mapping the message handle to its string.
+ *
+ * A message should be registered once and the handle reused throughout the rest
+ * of the application. This can be done by either explicitly creating static
+ * `nvtx3::registered_string` objects, or using the
+ * `nvtx3::registered_string::get` construct on first use helper (recommended).
+ *
+ * Similar to \ref DOMAINS, `nvtx3::registered_string::get` requires defining a
+ * custom tag type with a static `message` member whose value will be the
+ * contents of the registered string.
+ *
+ * Example:
+ * \code{.cpp}
+ * // Explicitly constructed, static `registered_string` in my_domain:
+ * static registered_string_in<my_domain> static_message{"my message"};
+ *
+ * // Or use construct on first use:
+ * // Define a tag type with a `message` member string to register
+ * struct my_message{ static constexpr char const* message{ "my message" }; };
+ *
+ * // Uses construct on first use to register the contents of
+ * // `my_message::message`
+ * auto& msg = nvtx3::registered_string_in<my_domain>::get<my_message>();
+ * \endcode
+ *
+ * \subsection COLOR color
+ *
+ * Associating a `nvtx3::color` with an event allows controlling how the event
+ * is visualized in a tool such as Nsight Systems. This is a convenient way to
+ * visually differentiate among different events.
+ *
+ * \code{.cpp}
+ * // Define a color via rgb color values
+ * nvtx3::color c{nvtx3::rgb{127, 255, 0}};
+ * nvtx3::event_attributes attr{c};
+ *
+ * // rgb color values can be passed directly to an `event_attributes`
+ * nvtx3::event_attributes attr1{nvtx3::rgb{127,255,0}};
+ * \endcode
+ *
+ * \subsection CATEGORY category
+ *
+ * A `nvtx3::category` is simply an integer id that allows for fine-grain
+ * grouping of NVTX events. For example, one might use separate categories for
+ * IO, memory allocation, compute, etc.
+ *
+ * \code{.cpp}
+ * nvtx3::event_attributes{nvtx3::category{1}};
+ * \endcode
+ *
+ * \subsubsection NAMED_CATEGORIES Named Categories
+ *
+ * Associates a `name` string with a category `id` to help differentiate among
+ * categories.
+ *
+ * For any given category id `Id`, a `named_category{Id, "name"}` should only
+ * be constructed once and reused throughout an application. This can be done by
+ * either explicitly creating static `nvtx3::named_category` objects, or using
+ * the `nvtx3::named_category::get` construct on first use helper (recommended).
+ *
+ * Similar to \ref DOMAINS, `nvtx3::named_category::get` requires defining a
+ * custom tag type with static `name` and `id` members.
+ *
+ * \code{.cpp}
+ * // Explicitly constructed, static `named_category` in my_domain:
+ * static nvtx3::named_category_in<my_domain> static_category{42, "my category"};
+ *
+ * // Or use construct on first use:
+ * // Define a tag type with `name` and `id` members
+ * struct my_category {
+ *    static constexpr char const* name{"my category"}; // category name
+ *    static constexpr uint32_t id{42}; // category id
+ * };
+ *
+ * // Use construct on first use to name the category id `42`
+ * // with name "my category":
+ * auto& cat = named_category_in<my_domain>::get<my_category>();
+ *
+ * // Range `r` associated with category id `42`
+ * nvtx3::event_attributes attr{cat};
+ * \endcode
+ *
+ * \subsection PAYLOAD payload
+ *
+ * Allows associating a user-defined numerical value with an event.
+ *
+ * \code{.cpp}
+ * // Constructs a payload from the `int32_t` value 42
+ * nvtx3:: event_attributes attr{nvtx3::payload{42}};
+ * \endcode
+ *
+ *
+ * \section EXAMPLE Example
+ *
+ * Putting it all together:
+ * \code{.cpp}
+ * // Define a custom domain tag type
+ * struct my_domain{ static constexpr char const* name{"my domain"}; };
+ *
+ * // Define a named category tag type
+ * struct my_category{
+ *    static constexpr char const* name{"my category"};
+ *    static constexpr uint32_t id{42};
+ * };
+ *
+ * // Define a registered string tag type
+ * struct my_message{ static constexpr char const* message{"my message"}; };
+ *
+ * // For convenience, use aliases for domain scoped objects
+ * using my_scoped_range = nvtx3::scoped_range_in<my_domain>;
+ * using my_registered_string = nvtx3::registered_string_in<my_domain>;
+ * using my_named_category = nvtx3::named_category_in<my_domain>;
+ *
+ * // Default values for all attributes
+ * nvtx3::event_attributes attr{};
+ * my_scoped_range r0{attr};
+ *
+ * // Custom (unregistered) message, and unnamed category
+ * nvtx3::event_attributes attr1{"message", nvtx3::category{2}};
+ * my_scoped_range r1{attr1};
+ *
+ * // Alternatively, pass arguments of `event_attributes` constructor directly
+ * // to `my_scoped_range`
+ * my_scoped_range r2{"message", nvtx3::category{2}};
+ *
+ * // construct on first use a registered string
+ * auto& msg = my_registered_string::get<my_message>();
+ *
+ * // construct on first use a named category
+ * auto& cat = my_named_category::get<my_category>();
+ *
+ * // Use registered string and named category with a custom payload
+ * my_scoped_range r3{msg, cat, nvtx3::payload{42}};
+ *
+ * // Any number of arguments in any order
+ * my_scoped_range r{nvtx3::rgb{127, 255,0}, msg};
+ *
+ * \endcode
+ * \section MACROS Convenience Macros
+ *
+ * Oftentimes users want to quickly and easily add NVTX ranges to their library
+ * or application to aid in profiling and optimization.
+ *
+ * A convenient way to do this is to use the \ref NVTX3_FUNC_RANGE and
+ * \ref NVTX3_FUNC_RANGE_IN macros. These macros take care of constructing an
+ * `nvtx3::scoped_range_in` with the name of the enclosing function as the
+ * range's message.
+ *
+ * \code{.cpp}
+ * void some_function() {
+ *    // Automatically generates an NVTX range for the duration of the function
+ *    // using "some_function" as the event's message.
+ *    NVTX3_FUNC_RANGE();
+ * }
+ * \endcode
+ *
+ */
+
+/* Temporary helper #defines, removed with #undef at end of header */
+
+/* Some compilers do not correctly support SFINAE, which is used in this API
+ * to detect common usage errors and provide clearer error messages (by using
+ * static_assert) than the compiler would produce otherwise.  These compilers
+ * will generate errors while compiling this file such as:
+ *
+ *  error: 'name' is not a member of 'nvtx3::v1::domain::global'
+ *
+ * The following compiler versions are known to have this problem, and so are
+ * set by default to disable the SFINAE-based checks:
+ *
+ * - All MSVC versions prior to VS2017 Update 7 (15.7)
+ * - GCC 8.1-8.3 (the problem was fixed in GCC 8.4)
+ *
+ * If you find your compiler hits this problem, you can work around it by
+ * defining NVTX3_USE_CHECKED_OVERLOADS_FOR_GET to 0 before including this
+ * header, or you can add a check for your compiler version to this #if.
+ * Also, please report the issue on the NVTX GitHub page.
+ */
+#if !defined(NVTX3_USE_CHECKED_OVERLOADS_FOR_GET)
+#  if defined(_MSC_VER) && _MSC_VER < 1914 || defined(__GNUC__) && __GNUC__ == 8 && __GNUC_MINOR__ < 4
+#    define NVTX3_USE_CHECKED_OVERLOADS_FOR_GET 0
+#  else
+#    define NVTX3_USE_CHECKED_OVERLOADS_FOR_GET 1
+#  endif
+#  define NVTX3_USE_CHECKED_OVERLOADS_FOR_GET_DEFINED_HERE
+#endif
+
+/* Within this header, nvtx3::NVTX3_VERSION_NAMESPACE resolves to nvtx3::vX,
+ * where "X" is the major version number. */
+#define NVTX3_CONCAT(A, B)           A##B
+#define NVTX3_NAMESPACE_FOR(VERSION) NVTX3_CONCAT(v, VERSION)
+#define NVTX3_VERSION_NAMESPACE      NVTX3_NAMESPACE_FOR(NVTX3_CPP_VERSION_MAJOR)
+
+/* Avoid duplicating #if defined(NVTX3_INLINE_THIS_VERSION) for namespaces
+ * in each minor version by making a macro to use unconditionally, which
+ * resolves to "inline" or nothing as appropriate. */
+#if defined(NVTX3_INLINE_THIS_VERSION)
+#  define NVTX3_INLINE_IF_REQUESTED inline
+#else
+#  define NVTX3_INLINE_IF_REQUESTED
+#endif
+
+/* Enables the use of constexpr when support for C++14 constexpr is present.
+ *
+ * Initialization of a class member that is a union to a specific union member
+ * can only be done in the body of a constructor, not in a member initializer
+ * list.  A constexpr constructor must have an empty body until C++14, so there
+ * is no way to make an initializer of a member union constexpr in C++11.  This
+ * macro allows making functions constexpr in C++14 or newer, but non-constexpr
+ * in C++11 compilation.  It is used here on constructors that initialize their
+ * member unions.
+ */
+#if __cpp_constexpr >= 201304L
+#  define NVTX3_CONSTEXPR_IF_CPP14 constexpr
+#else
+#  define NVTX3_CONSTEXPR_IF_CPP14
+#endif
+
+// Macro wrappers for C++ attributes
+#if !defined(__has_cpp_attribute)
+#  define __has_cpp_attribute(x) 0
+#endif
+#if __has_cpp_attribute(maybe_unused)
+#  define NVTX3_MAYBE_UNUSED [[maybe_unused]]
+#else
+#  define NVTX3_MAYBE_UNUSED
+#endif
+#if __has_cpp_attribute(nodiscard)
+#  define NVTX3_NO_DISCARD [[nodiscard]]
+#else
+#  define NVTX3_NO_DISCARD
+#endif
+
+/* Use a macro for static asserts, which defaults to static_assert, but that
+ * testing tools can replace with a logging function.  For example:
+ * #define NVTX3_STATIC_ASSERT(c, m) \
+ *   do { if (!(c)) printf("static_assert would fail: %s\n", m); } while (0)
+ */
+#if !defined(NVTX3_STATIC_ASSERT)
+#  define NVTX3_STATIC_ASSERT(condition, message) static_assert(condition, message)
+#  define NVTX3_STATIC_ASSERT_DEFINED_HERE
+#endif
+
+/* Implementation sections, enclosed in guard macros for each minor version */
+
+#ifndef NVTX3_CPP_DEFINITIONS_V1_0
+#  define NVTX3_CPP_DEFINITIONS_V1_0
+
+#  include <cuda/std/__host_stdlib/memory>
+
+#  include <cstddef>
+#  include <string>
+#  include <type_traits>
+#  include <utility>
+
+#  include <nvtx3/nvToolsExt.h>
+
+namespace nvtx3
+{
+NVTX3_INLINE_IF_REQUESTED namespace NVTX3_VERSION_NAMESPACE
+{
+  namespace detail
+  {
+  template <typename Unused>
+  struct always_false : std::false_type
+  {};
+
+  template <typename T, typename = void>
+  struct has_name : std::false_type
+  {};
+  template <typename T>
+  struct has_name<T, decltype((void) T::name, void())> : std::true_type
+  {};
+
+  template <typename T, typename = void>
+  struct has_id : std::false_type
+  {};
+  template <typename T>
+  struct has_id<T, decltype((void) T::id, void())> : std::true_type
+  {};
+
+  template <typename T, typename = void>
+  struct has_message : std::false_type
+  {};
+  template <typename T>
+  struct has_message<T, decltype((void) T::message, void())> : std::true_type
+  {};
+
+  template <typename T, typename = void>
+  struct is_c_string : std::false_type
+  {};
+  template <typename T>
+  struct is_c_string<T,
+                     typename std::enable_if<std::is_convertible<T, char const*>::value
+                                             || std::is_convertible<T, wchar_t const*>::value>::type> : std::true_type
+  {};
+
+  template <typename T>
+  using is_uint32 = std::is_same<typename std::decay<T>::type, uint32_t>;
+  } // namespace detail
+
+  /**
+   * @brief `domain`s allow for grouping NVTX events into a single scope to
+   * differentiate them from events in other `domain`s.
+   *
+   * By default, all NVTX constructs are placed in the "global" NVTX domain.
+   *
+   * A custom `domain` may be used in order to differentiate a library's or
+   * application's NVTX events from other events.
+   *
+   * `domain`s are expected to be long-lived and unique to a library or
+   * application. As such, it is assumed a domain's name is known at compile
+   * time. Therefore, all NVTX constructs that can be associated with a domain
+   * require the domain to be specified via a *type* `D` passed as an
+   * explicit template parameter.
+   *
+   * The type `domain::global` may be used to indicate that the global NVTX
+   * domain should be used.
+   *
+   * None of the C++ NVTX constructs require the user to manually construct a
+   * `domain` object. Instead, if a custom domain is desired, the user is
+   * expected to define a type `D` that contains a member
+   * `D::name` which resolves to either a `char const*` or `wchar_t
+   * const*`. The value of `D::name` is used to name and uniquely
+   * identify the custom domain.
+   *
+   * Upon the first use of an NVTX construct associated with the type
+   * `D`, the "construct on first use" pattern is used to construct a
+   * function local static `domain` object. All future NVTX constructs
+   * associated with `D` will use a reference to the previously
+   * constructed `domain` object. See `domain::get`.
+   *
+   * Example:
+   * \code{.cpp}
+   * // The type `my_domain` defines a `name` member used to name and identify
+   * // the `domain` object identified by `my_domain`.
+   * struct my_domain{ static constexpr char const* name{"my_domain"}; };
+   *
+   * // The NVTX range `r` will be grouped with all other NVTX constructs
+   * // associated with  `my_domain`.
+   * nvtx3::scoped_range_in<my_domain> r{};
+   *
+   * // An alias can be created for a `scoped_range_in` in the custom domain
+   * using my_scoped_range = nvtx3::scoped_range_in<my_domain>;
+   * my_scoped_range my_range{};
+   *
+   * // `domain::global` indicates that the global NVTX domain is used
+   * nvtx3::scoped_range_in<domain::global> r2{};
+   *
+   * // For convenience, `nvtx3::scoped_range` is an alias for a range in the
+   * // global domain
+   * nvtx3::scoped_range r3{};
+   * \endcode
+   */
+  class domain
+  {
+  public:
+    domain(domain const&)            = delete;
+    domain& operator=(domain const&) = delete;
+    domain(domain&&)                 = delete;
+    domain& operator=(domain&&)      = delete;
+
+    /**
+     * @brief Tag type for the "global" NVTX domain.
+     *
+     * This type may be passed as a template argument to any function/class
+     * expecting a type to identify a domain to indicate that the global domain
+     * should be used.
+     *
+     * All NVTX events in the global domain across all libraries and
+     * applications will be grouped together.
+     *
+     */
+    struct global
+    {};
+
+#  if NVTX3_USE_CHECKED_OVERLOADS_FOR_GET
+    /**
+     * @brief Returns reference to an instance of a function local static
+     * `domain` object.
+     *
+     * Uses the "construct on first use" idiom to safely ensure the `domain`
+     * object is initialized exactly once upon first invocation of
+     * `domain::get<D>()`. All following invocations will return a
+     * reference to the previously constructed `domain` object. See
+     * https://isocpp.org/wiki/faq/ctors#static-init-order-on-first-use
+     *
+     * None of the constructs in this header require the user to directly invoke
+     * `domain::get`. It is automatically invoked when constructing objects like
+     * a `scoped_range_in` or `category`. Advanced users may wish to use
+     * `domain::get` for the convenience of the "construct on first use" idiom
+     * when using domains with their own use of the NVTX C API.
+     *
+     * This function is thread-safe as of C++11. If two or more threads call
+     * `domain::get<D>` concurrently, exactly one of them is guaranteed
+     * to construct the `domain` object and the other(s) will receive a
+     * reference to the object after it is fully constructed.
+     *
+     * The domain's name is specified via the type `D` pass as an
+     * explicit template parameter. `D` is required to contain a
+     * member `D::name` that resolves to either a `char const*` or
+     * `wchar_t const*`. The value of `D::name` is used to name and
+     * uniquely identify the `domain`.
+     *
+     * Example:
+     * \code{.cpp}
+     * // The type `my_domain` defines a `name` member used to name and identify
+     * // the `domain` object identified by `my_domain`.
+     * struct my_domain{ static constexpr char const* name{"my domain"}; };
+     *
+     * auto& D1 = domain::get<my_domain>(); // First invocation constructs a
+     *                                      // `domain` with the name "my domain"
+     *
+     * auto& D2 = domain::get<my_domain>(); // Quickly returns reference to
+     *                                      // previously constructed `domain`.
+     * \endcode
+     *
+     * @tparam D Type that contains a `D::name` member used to
+     * name the `domain` object.
+     * @return Reference to the `domain` corresponding to the type `D`.
+     */
+    template <typename D = global, typename std::enable_if<detail::is_c_string<decltype(D::name)>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static domain const& get() noexcept
+    {
+      static domain const d(D::name);
+      return d;
+    }
+
+    /**
+     * @brief Overload of `domain::get` to provide a clear compile error when
+     * `D` has a `name` member that is not directly convertible to either
+     * `char const*` or `wchar_t const*`.
+     */
+    template <typename D                                                                         = global,
+              typename std::enable_if<!detail::is_c_string<decltype(D::name)>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static domain const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::always_false<D>::value,
+                          "Type used to identify an NVTX domain must contain a static constexpr member "
+                          "called 'name' of type const char* or const wchar_t* -- 'name' member is not "
+                          "convertible to either of those types");
+      static domain const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+
+    /**
+     * @brief Overload of `domain::get` to provide a clear compile error when
+     * `D` does not have a `name` member.
+     */
+    template <typename D = global, typename std::enable_if<!detail::has_name<D>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static domain const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::always_false<D>::value,
+                          "Type used to identify an NVTX domain must contain a static constexpr member "
+                          "called 'name' of type const char* or const wchar_t* -- 'name' member is missing");
+      static domain const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+#  else
+    template <typename D = global>
+    NVTX3_NO_DISCARD static domain const& get() noexcept
+    {
+      static domain const d(D::name);
+      return d;
+    }
+#  endif
+
+    /**
+     * @brief Conversion operator to `nvtxDomainHandle_t`.
+     *
+     * Allows transparently passing a domain object into an API expecting a
+     * native `nvtxDomainHandle_t` object.
+     */
+    operator nvtxDomainHandle_t() const noexcept
+    {
+      return _domain;
+    }
+
+  private:
+    /**
+     * @brief Construct a new domain with the specified `name`.
+     *
+     * This constructor is private as it is intended that `domain` objects only
+     * be created through the `domain::get` function.
+     *
+     * @param name A unique name identifying the domain
+     */
+    explicit domain(char const* name) noexcept
+        : _domain{nvtxDomainCreateA(name)}
+    {}
+
+    /**
+     * @brief Construct a new domain with the specified `name`.
+     *
+     * This constructor is private as it is intended that `domain` objects only
+     * be created through the `domain::get` function.
+     *
+     * @param name A unique name identifying the domain
+     */
+    explicit domain(wchar_t const* name) noexcept
+        : _domain{nvtxDomainCreateW(name)}
+    {}
+
+    /**
+     * @brief Construct a new domain with the specified `name`.
+     *
+     * This constructor is private as it is intended that `domain` objects only
+     * be created through the `domain::get` function.
+     *
+     * @param name A unique name identifying the domain
+     */
+    explicit domain(std::string const& name) noexcept
+        : domain{name.c_str()}
+    {}
+
+    /**
+     * @brief Construct a new domain with the specified `name`.
+     *
+     * This constructor is private as it is intended that `domain` objects only
+     * be created through the `domain::get` function.
+     *
+     * @param name A unique name identifying the domain
+     */
+    explicit domain(std::wstring const& name) noexcept
+        : domain{name.c_str()}
+    {}
+
+    /**
+     * @brief Default constructor creates a `domain` representing the
+     * "global" NVTX domain.
+     *
+     * All events not associated with a custom `domain` are grouped in the
+     * "global" NVTX domain.
+     *
+     */
+    constexpr domain() noexcept {}
+
+    /**
+     * @brief Intentionally avoid calling nvtxDomainDestroy on the `domain` object.
+     *
+     * No currently-available tools attempt to free domain resources when the
+     * nvtxDomainDestroy function is called, due to the thread-safety and
+     * efficiency challenges of freeing thread-local storage for other threads.
+     * Since libraries may be disallowed from introducing static destructors,
+     * and destroying the domain is likely to have no effect, the destructor
+     * for `domain` intentionally chooses to not destroy the domain.
+     *
+     * In a situation where domain destruction is necessary, either manually
+     * call nvtxDomainDestroy on the domain's handle, or make a class that
+     * derives from `domain` and calls nvtxDomainDestroy in its destructor.
+     */
+    ~domain() = default;
+
+  private:
+    nvtxDomainHandle_t const _domain{}; ///< The `domain`s NVTX handle
+  };
+
+  /**
+   * @brief Returns reference to the `domain` object that represents the global
+   * NVTX domain.
+   *
+   * This specialization for `domain::global` returns a default constructed,
+   * `domain` object for use when the "global" domain is desired.
+   *
+   * All NVTX events in the global domain across all libraries and applications
+   * will be grouped together.
+   *
+   * @return Reference to the `domain` corresponding to the global NVTX domain.
+   *
+   */
+  template <>
+  NVTX3_NO_DISCARD inline domain const& domain::get<domain::global>() noexcept
+  {
+    static domain const d{};
+    return d;
+  }
+
+  /**
+   * @brief Indicates the values of the red, green, and blue color channels for
+   * an RGB color to use as an event attribute (assumes no transparency).
+   *
+   */
+  struct rgb
+  {
+    /// Type used for component values
+    using component_type = uint8_t;
+
+    /**
+     * @brief Construct a rgb with red, green, and blue channels
+     * specified by `red_`, `green_`, and `blue_`, respectively.
+     *
+     * Valid values are in the range `[0,255]`.
+     *
+     * @param red_ Value of the red channel
+     * @param green_ Value of the green channel
+     * @param blue_ Value of the blue channel
+     */
+    constexpr rgb(component_type red_, component_type green_, component_type blue_) noexcept
+        : red{red_}
+        , green{green_}
+        , blue{blue_}
+    {}
+
+    component_type red{}; ///< Red channel value
+    component_type green{}; ///< Green channel value
+    component_type blue{}; ///< Blue channel value
+  };
+
+  /**
+   * @brief Indicates the value of the alpha, red, green, and blue color
+   * channels for an ARGB color to use as an event attribute.
+   *
+   */
+  struct argb final : rgb
+  {
+    /**
+     * @brief Construct an argb with alpha, red, green, and blue channels
+     * specified by `alpha_`, `red_`, `green_`, and `blue_`, respectively.
+     *
+     * Valid values are in the range `[0,255]`.
+     *
+     * @param alpha_  Value of the alpha channel (opacity)
+     * @param red_  Value of the red channel
+     * @param green_  Value of the green channel
+     * @param blue_  Value of the blue channel
+     *
+     */
+    constexpr argb(component_type alpha_, component_type red_, component_type green_, component_type blue_) noexcept
+        : rgb{red_, green_, blue_}
+        , alpha{alpha_}
+    {}
+
+    component_type alpha{}; ///< Alpha channel value
+  };
+
+  /**
+   * @brief Represents a custom color that can be associated with an NVTX event
+   * via its `event_attributes`.
+   *
+   * Specifying colors for NVTX events is a convenient way to visually
+   * differentiate among different events in a visualization tool such as Nsight
+   * Systems.
+   *
+   */
+  class color
+  {
+  public:
+    /// Type used for the color's value
+    using value_type = uint32_t;
+
+    /**
+     * @brief Constructs a `color` using the value provided by `hex_code`.
+     *
+     * `hex_code` is expected to be a 4 byte argb hex code.
+     *
+     * The most significant byte indicates the value of the alpha channel
+     * (opacity) (0-255)
+     *
+     * The next byte indicates the value of the red channel (0-255)
+     *
+     * The next byte indicates the value of the green channel (0-255)
+     *
+     * The least significant byte indicates the value of the blue channel
+     * (0-255)
+     *
+     * @param hex_code The hex code used to construct the `color`
+     */
+    constexpr explicit color(value_type hex_code) noexcept
+        : _value{hex_code}
+    {}
+
+    /**
+     * @brief Construct a `color` using the alpha, red, green, blue components
+     * in `argb`.
+     *
+     * @param argb_ The alpha, red, green, blue components of the desired `color`
+     */
+    constexpr color(argb argb_) noexcept
+        : color{from_bytes_msb_to_lsb(argb_.alpha, argb_.red, argb_.green, argb_.blue)}
+    {}
+
+    /**
+     * @brief Construct a `color` using the red, green, blue components in
+     * `rgb`.
+     *
+     * Uses maximum value for the alpha channel (opacity) of the `color`.
+     *
+     * @param rgb_ The red, green, blue components of the desired `color`
+     */
+    constexpr color(rgb rgb_) noexcept
+        : color{from_bytes_msb_to_lsb(0xFF, rgb_.red, rgb_.green, rgb_.blue)}
+    {}
+
+    /**
+     * @brief Returns the `color`s argb hex code
+     *
+     */
+    constexpr value_type get_value() const noexcept
+    {
+      return _value;
+    }
+
+    /**
+     * @brief Return the NVTX color type of the color.
+     *
+     */
+    constexpr nvtxColorType_t get_type() const noexcept
+    {
+      return _type;
+    }
+
+    color()                        = delete;
+    ~color()                       = default;
+    color(color const&)            = default;
+    color& operator=(color const&) = default;
+    color(color&&)                 = default;
+    color& operator=(color&&)      = default;
+
+  private:
+    /**
+     * @brief Constructs an unsigned, 4B integer from the component bytes in
+     * most to least significant byte order.
+     *
+     */
+    constexpr static value_type
+    from_bytes_msb_to_lsb(uint8_t byte3, uint8_t byte2, uint8_t byte1, uint8_t byte0) noexcept
+    {
+      return uint32_t{byte3} << 24 | uint32_t{byte2} << 16 | uint32_t{byte1} << 8 | uint32_t{byte0};
+    }
+
+    value_type _value{}; ///< color's argb color code
+    nvtxColorType_t _type{NVTX_COLOR_ARGB}; ///< NVTX color type code
+  };
+
+  /**
+   * @brief Object for intra-domain grouping of NVTX events.
+   *
+   * A `category` is simply an integer id that allows for fine-grain grouping of
+   * NVTX events. For example, one might use separate categories for IO, memory
+   * allocation, compute, etc.
+   *
+   * Example:
+   * \code{.cpp}
+   * nvtx3::category cat1{1};
+   *
+   * // Range `r1` belongs to the category identified by the value `1`.
+   * nvtx3::scoped_range r1{cat1};
+   *
+   * // Range `r2` belongs to the same category as `r1`
+   * nvtx3::scoped_range r2{nvtx3::category{1}};
+   * \endcode
+   *
+   * To associate a name string with a category id, see `named_category`.
+   *
+   */
+  class category
+  {
+  public:
+    /// Type used for `category`s integer id.
+    using id_type = uint32_t;
+
+    /**
+     * @brief Construct a `category` with the specified `id`.
+     *
+     * The `category` will be unnamed and identified only by its `id` value.
+     *
+     * All `category`s in a domain sharing the same `id` are equivalent.
+     *
+     * @param[in] id The `category`'s identifying value
+     */
+    constexpr explicit category(id_type id) noexcept
+        : id_{id}
+    {}
+
+    /**
+     * @brief Returns the id of the category.
+     *
+     */
+    constexpr id_type get_id() const noexcept
+    {
+      return id_;
+    }
+
+    category()                           = delete;
+    ~category()                          = default;
+    category(category const&)            = default;
+    category& operator=(category const&) = default;
+    category(category&&)                 = default;
+    category& operator=(category&&)      = default;
+
+  private:
+    id_type id_{}; ///< category's unique identifier
+  };
+
+  /**
+   * @brief A `category` with an associated name string.
+   *
+   * Associates a `name` string with a category `id` to help differentiate among
+   * categories.
+   *
+   * For any given category id `Id`, a `named_category(Id, "name")` should only
+   * be constructed once and reused throughout an application. This can be done
+   * by either explicitly creating static `named_category` objects, or using the
+   * `named_category::get` construct on first use helper (recommended).
+   *
+   * Creating two or more `named_category` objects with the same value for `id`
+   * in the same domain results in undefined behavior.
+   *
+   * Similarly, behavior is undefined when a `named_category` and `category`
+   * share the same value of `id`.
+   *
+   * Example:
+   * \code{.cpp}
+   * // Explicitly constructed, static `named_category` in global domain:
+   * static nvtx3::named_category static_category{42, "my category"};
+   *
+   * // Range `r` associated with category id `42`
+   * nvtx3::scoped_range r{static_category};
+   *
+   * // OR use construct on first use:
+   *
+   * // Define a type with `name` and `id` members
+   * struct my_category {
+   *    static constexpr char const* name{"my category"}; // category name
+   *    static constexpr uint32_t id{42}; // category id
+   * };
+   *
+   * // Use construct on first use to name the category id `42`
+   * // with name "my category"
+   * auto& cat = named_category_in<my_domain>::get<my_category>();
+   *
+   * // Range `r` associated with category id `42`
+   * nvtx3::scoped_range r{cat};
+   * \endcode
+   *
+   * `named_category_in<D>`'s association of a name to a category id is local to
+   * the domain specified by the type `D`. An id may have a different name in
+   * another domain.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain` to
+   * which the `named_category_in` belongs. Else, `domain::global` to indicate
+   * that the global NVTX domain should be used.
+   */
+  template <typename D = domain::global>
+  class named_category_in final : public category
+  {
+  public:
+#  if NVTX3_USE_CHECKED_OVERLOADS_FOR_GET
+    /**
+     * @brief Returns a global instance of a `named_category_in` as a
+     * function-local static.
+     *
+     * Creates a `named_category_in<D>` with name and id specified by the contents
+     * of a type `C`. `C::name` determines the name and `C::id` determines the
+     * category id.
+     *
+     * This function is useful for constructing a named `category` exactly once
+     * and reusing the same instance throughout an application.
+     *
+     * Example:
+     * \code{.cpp}
+     * // Define a type with `name` and `id` members
+     * struct my_category {
+     *    static constexpr char const* name{"my category"}; // category name
+     *    static constexpr uint32_t id{42}; // category id
+     * };
+     *
+     * // Use construct on first use to name the category id `42`
+     * // with name "my category"
+     * auto& cat = named_category_in<my_domain>::get<my_category>();
+     *
+     * // Range `r` associated with category id `42`
+     * nvtx3::scoped_range r{cat};
+     * \endcode
+     *
+     * Uses the "construct on first use" idiom to safely ensure the `category`
+     * object is initialized exactly once. See
+     * https://isocpp.org/wiki/faq/ctors#static-init-order-on-first-use
+     *
+     * @tparam C Type containing a member `C::name` that resolves to either a
+     * `char const*` or `wchar_t const*` and `C::id`.
+     */
+    template <
+      typename C,
+      typename std::enable_if<detail::is_c_string<decltype(C::name)>::value && detail::is_uint32<decltype(C::id)>::value,
+                              int>::type = 0>
+    static named_category_in const& get() noexcept
+    {
+      static named_category_in const cat(C::id, C::name);
+      return cat;
+    }
+
+    /**
+     * @brief Overload of `named_category_in::get` to provide a clear compile error
+     * when `C` has the required `name` and `id` members, but they are not the
+     * required types.  `name` must be directly convertible to `char const*` or
+     * `wchar_t const*`, and `id` must be `uint32_t`.
+     */
+    template <typename C,
+              typename std::enable_if<!detail::is_c_string<decltype(C::name)>::value
+                                        || !detail::is_uint32<decltype(C::id)>::value,
+                                      int>::type = 0>
+    NVTX3_NO_DISCARD static named_category_in const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::is_c_string<decltype(C::name)>::value,
+                          "Type used to name an NVTX category must contain a static constexpr member "
+                          "called 'name' of type const char* or const wchar_t* -- 'name' member is not "
+                          "convertible to either of those types");
+      NVTX3_STATIC_ASSERT(detail::is_uint32<decltype(C::id)>::value,
+                          "Type used to name an NVTX category must contain a static constexpr member "
+                          "called 'id' of type uint32_t -- 'id' member is the wrong type");
+      static named_category_in const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+
+    /**
+     * @brief Overload of `named_category_in::get` to provide a clear compile error
+     * when `C` does not have the required `name` and `id` members.
+     */
+    template <typename C,
+              typename std::enable_if<!detail::has_name<C>::value || !detail::has_id<C>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static named_category_in const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::has_name<C>::value,
+                          "Type used to name an NVTX category must contain a static constexpr member "
+                          "called 'name' of type const char* or const wchar_t* -- 'name' member is missing");
+      NVTX3_STATIC_ASSERT(detail::has_id<C>::value,
+                          "Type used to name an NVTX category must contain a static constexpr member "
+                          "called 'id' of type uint32_t -- 'id' member is missing");
+      static named_category_in const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+#  else
+    template <typename C>
+    NVTX3_NO_DISCARD static named_category_in const& get() noexcept
+    {
+      static named_category_in const cat(C::id, C::name);
+      return cat;
+    }
+#  endif
+
+  private:
+    // Default constructor is only used internally for static_assert(false) cases.
+    named_category_in() noexcept
+        : category{0}
+    {}
+
+  public:
+    /**
+     * @brief Construct a `named_category_in` with the specified `id` and `name`.
+     *
+     * The name `name` will be registered with `id`.
+     *
+     * Every unique value of `id` should only be named once.
+     *
+     * @param[in] id The category id to name
+     * @param[in] name The name to associated with `id`
+     */
+    named_category_in(id_type id, char const* name) noexcept
+        : category{id}
+    {
+#  ifndef NVTX_DISABLE
+      nvtxDomainNameCategoryA(domain::get<D>(), get_id(), name);
+#  else
+      (void) id;
+      (void) name;
+#  endif
+    }
+
+    /**
+     * @brief Construct a `named_category_in` with the specified `id` and `name`.
+     *
+     * The name `name` will be registered with `id`.
+     *
+     * Every unique value of `id` should only be named once.
+     *
+     * @param[in] id The category id to name
+     * @param[in] name The name to associated with `id`
+     */
+    named_category_in(id_type id, wchar_t const* name) noexcept
+        : category{id}
+    {
+#  ifndef NVTX_DISABLE
+      nvtxDomainNameCategoryW(domain::get<D>(), get_id(), name);
+#  else
+      (void) id;
+      (void) name;
+#  endif
+    }
+  };
+
+  /**
+   * @brief Alias for a `named_category_in` in the global NVTX domain.
+   *
+   */
+  using named_category = named_category_in<domain::global>;
+
+  /**
+   * @brief A message registered with NVTX.
+   *
+   * Normally, associating a `message` with an NVTX event requires copying the
+   * contents of the message string. This may cause non-trivial overhead in
+   * highly performance sensitive regions of code.
+   *
+   * message registration is an optimization to lower the overhead of
+   * associating a message with an NVTX event. Registering a message yields a
+   * handle that is inexpensive to copy that may be used in place of a message
+   * string.
+   *
+   * A particular message should only be registered once and the handle
+   * reused throughout the rest of the application. This can be done by either
+   * explicitly creating static `registered_string_in` objects, or using the
+   * `registered_string_in::get` construct on first use helper (recommended).
+   *
+   * Example:
+   * \code{.cpp}
+   * // Explicitly constructed, static `registered_string` in my_domain:
+   * static registered_string_in<my_domain> static_message{"message"};
+   *
+   * // "message" is associated with the range `r`
+   * nvtx3::scoped_range r{static_message};
+   *
+   * // Or use construct on first use:
+   *
+   * // Define a type with a `message` member that defines the contents of the
+   * // registered string
+   * struct my_message{ static constexpr char const* message{ "my message" }; };
+   *
+   * // Uses construct on first use to register the contents of
+   * // `my_message::message`
+   * auto& msg = registered_string_in<my_domain>::get<my_message>();
+   *
+   * // "my message" is associated with the range `r`
+   * nvtx3::scoped_range r{msg};
+   * \endcode
+   *
+   * `registered_string_in`s are local to a particular domain specified via
+   * the type `D`.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain` to
+   * which the `registered_string_in` belongs. Else, `domain::global` to indicate
+   * that the global NVTX domain should be used.
+   */
+  template <typename D = domain::global>
+  class registered_string_in
+  {
+  public:
+#  if NVTX3_USE_CHECKED_OVERLOADS_FOR_GET
+    /**
+     * @brief Returns a global instance of a `registered_string_in` as a function
+     * local static.
+     *
+     * Provides a convenient way to register a message with NVTX without having
+     * to explicitly register the message.
+     *
+     * Upon first invocation, constructs a `registered_string_in` whose contents
+     * are specified by `message::message`.
+     *
+     * All future invocations will return a reference to the object constructed
+     * in the first invocation.
+     *
+     * Example:
+     * \code{.cpp}
+     * // Define a type with a `message` member that defines the contents of the
+     * // registered string
+     * struct my_message{ static constexpr char const* message{ "my message" };
+     * };
+     *
+     * // Uses construct on first use to register the contents of
+     * // `my_message::message`
+     * auto& msg = registered_string_in<my_domain>::get<my_message>();
+     *
+     * // "my message" is associated with the range `r`
+     * nvtx3::scoped_range r{msg};
+     * \endcode
+     *
+     * @tparam M Type required to contain a member `M::message` that
+     * resolves to either a `char const*` or `wchar_t const*` used as the
+     * registered string's contents.
+     * @return Reference to a `registered_string_in` associated with the type `M`.
+     */
+    template <typename M, typename std::enable_if<detail::is_c_string<decltype(M::message)>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static registered_string_in const& get() noexcept
+    {
+      static registered_string_in const regstr(M::message);
+      return regstr;
+    }
+
+    /**
+     * @brief Overload of `registered_string_in::get` to provide a clear compile error
+     * when `M` has a `message` member that is not directly convertible to either
+     * `char const*` or `wchar_t const*`.
+     */
+    template <typename M, typename std::enable_if<!detail::is_c_string<decltype(M::message)>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static registered_string_in const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::always_false<M>::value,
+                          "Type used to register an NVTX string must contain a static constexpr member "
+                          "called 'message' of type const char* or const wchar_t* -- 'message' member is "
+                          "not convertible to either of those types");
+      static registered_string_in const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+
+    /**
+     * @brief Overload of `registered_string_in::get` to provide a clear compile error when
+     * `M` does not have a `message` member.
+     */
+    template <typename M, typename std::enable_if<!detail::has_message<M>::value, int>::type = 0>
+    NVTX3_NO_DISCARD static registered_string_in const& get() noexcept
+    {
+      NVTX3_STATIC_ASSERT(detail::always_false<M>::value,
+                          "Type used to register an NVTX string must contain a static constexpr member "
+                          "called 'message' of type const char* or const wchar_t* -- 'message' member "
+                          "is missing");
+      static registered_string_in const unused;
+      return unused; // Function must compile for static_assert to be triggered
+    }
+#  else
+    template <typename M>
+    NVTX3_NO_DISCARD static registered_string_in const& get() noexcept
+    {
+      static registered_string_in const regstr(M::message);
+      return regstr;
+    }
+#  endif
+
+    /**
+     * @brief Constructs a `registered_string_in` from the specified `msg` string.
+     *
+     * Registers `msg` with NVTX and associates a handle with the registered
+     * message.
+     *
+     * A particular message should should only be registered once and the handle
+     * reused throughout the rest of the application.
+     *
+     * @param msg The contents of the message
+     */
+    explicit registered_string_in(char const* msg) noexcept
+        : handle_{nvtxDomainRegisterStringA(domain::get<D>(), msg)}
+    {}
+
+    /**
+     * @brief Constructs a `registered_string_in` from the specified `msg` string.
+     *
+     * Registers `msg` with NVTX and associates a handle with the registered
+     * message.
+     *
+     * A particular message should should only be registered once and the handle
+     * reused throughout the rest of the application.
+     *
+     * @param msg The contents of the message
+     */
+    explicit registered_string_in(std::string const& msg) noexcept
+        : registered_string_in{msg.c_str()}
+    {}
+
+    /**
+     * @brief Constructs a `registered_string_in` from the specified `msg` string.
+     *
+     * Registers `msg` with NVTX and associates a handle with the registered
+     * message.
+     *
+     * A particular message should should only be registered once and the handle
+     * reused throughout the rest of the application.
+     *
+     * @param msg The contents of the message
+     */
+    explicit registered_string_in(wchar_t const* msg) noexcept
+        : handle_{nvtxDomainRegisterStringW(domain::get<D>(), msg)}
+    {}
+
+    /**
+     * @brief Constructs a `registered_string_in` from the specified `msg` string.
+     *
+     * Registers `msg` with NVTX and associates a handle with the registered
+     * message.
+     *
+     * A particular message should only be registered once and the handle
+     * reused throughout the rest of the application.
+     *
+     * @param msg The contents of the message
+     */
+    explicit registered_string_in(std::wstring const& msg) noexcept
+        : registered_string_in{msg.c_str()}
+    {}
+
+    /**
+     * @brief Returns the registered string's handle
+     *
+     */
+    nvtxStringHandle_t get_handle() const noexcept
+    {
+      return handle_;
+    }
+
+  private:
+    // Default constructor is only used internally for static_assert(false) cases.
+    registered_string_in() noexcept {}
+
+  public:
+    ~registered_string_in()                                      = default;
+    registered_string_in(registered_string_in const&)            = default;
+    registered_string_in& operator=(registered_string_in const&) = default;
+    registered_string_in(registered_string_in&&)                 = default;
+    registered_string_in& operator=(registered_string_in&&)      = default;
+
+  private:
+    nvtxStringHandle_t handle_{}; ///< The handle returned from
+                                  ///< registering the message with NVTX
+  };
+
+  /**
+   * @brief Alias for a `registered_string_in` in the global NVTX domain.
+   *
+   */
+  using registered_string = registered_string_in<domain::global>;
+
+  /**
+   * @brief Allows associating a message string with an NVTX event via
+   * its `EventAttribute`s.
+   *
+   * Associating a `message` with an NVTX event through its `event_attributes`
+   * allows for naming events to easily differentiate them from other events.
+   *
+   * Every time an NVTX event is created with an associated `message`, the
+   * contents of the message string must be copied.  This may cause non-trivial
+   * overhead in highly performance sensitive sections of code. Use of a
+   * `nvtx3::registered_string` is recommended in these situations.
+   *
+   * Example:
+   * \code{.cpp}
+   * // Creates an `event_attributes` with message "message 0"
+   * nvtx3::event_attributes attr0{nvtx3::message{"message 0"}};
+   *
+   * // `range0` contains message "message 0"
+   * nvtx3::scoped_range range0{attr0};
+   *
+   * // `std::string` and string literals are implicitly assumed to be
+   * // the contents of an `nvtx3::message`
+   * // Creates an `event_attributes` with message "message 1"
+   * nvtx3::event_attributes attr1{"message 1"};
+   *
+   * // `range1` contains message "message 1"
+   * nvtx3::scoped_range range1{attr1};
+   *
+   * // `range2` contains message "message 2"
+   * nvtx3::scoped_range range2{nvtx3::message{"message 2"}};
+   *
+   * // `std::string` and string literals are implicitly assumed to be
+   * // the contents of an `nvtx3::message`
+   * // `range3` contains message "message 3"
+   * nvtx3::scoped_range range3{"message 3"};
+   * \endcode
+   */
+  class message
+  {
+  public:
+    using value_type = nvtxMessageValue_t;
+
+    /**
+     * @brief Construct a `message` whose contents are specified by `msg`.
+     *
+     * @param msg The contents of the message
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 message(char const* msg) noexcept
+        : type_{NVTX_MESSAGE_TYPE_ASCII}
+    {
+      value_.ascii = msg;
+    }
+
+    /**
+     * @brief Construct a `message` whose contents are specified by `msg`.
+     *
+     * @param msg The contents of the message
+     */
+    message(std::string const& msg) noexcept
+        : message{msg.c_str()}
+    {}
+
+    /**
+     * @brief Disallow construction for `std::string` r-value
+     *
+     * `message` is a non-owning type and therefore cannot take ownership of an
+     * r-value. Therefore, constructing from an r-value is disallowed to prevent
+     * a dangling pointer.
+     *
+     */
+    message(std::string&&) = delete;
+
+    /**
+     * @brief Construct a `message` whose contents are specified by `msg`.
+     *
+     * @param msg The contents of the message
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 message(wchar_t const* msg) noexcept
+        : type_{NVTX_MESSAGE_TYPE_UNICODE}
+    {
+      value_.unicode = msg;
+    }
+
+    /**
+     * @brief Construct a `message` whose contents are specified by `msg`.
+     *
+     * @param msg The contents of the message
+     */
+    message(std::wstring const& msg) noexcept
+        : message{msg.c_str()}
+    {}
+
+    /**
+     * @brief Disallow construction for `std::wstring` r-value
+     *
+     * `message` is a non-owning type and therefore cannot take ownership of an
+     * r-value. Therefore, constructing from an r-value is disallowed to prevent
+     * a dangling pointer.
+     *
+     */
+    message(std::wstring&&) = delete;
+
+    /**
+     * @brief Construct a `message` from a `registered_string_in`.
+     *
+     * @tparam D Type containing `name` member used to identify the `domain`
+     * to which the `registered_string_in` belongs. Else, `domain::global` to
+     * indicate that the global NVTX domain should be used.
+     * @param msg The message that has already been registered with NVTX.
+     */
+    template <typename D>
+    NVTX3_CONSTEXPR_IF_CPP14 message(registered_string_in<D> const& msg) noexcept
+        : type_{NVTX_MESSAGE_TYPE_REGISTERED}
+    {
+      value_.registered = msg.get_handle();
+    }
+
+    /**
+     * @brief Construct a `message` from NVTX C API type and value.
+     *
+     * @param type nvtxMessageType_t enum value indicating type of the payload
+     * @param value nvtxMessageValue_t union containing message
+     */
+    constexpr message(nvtxMessageType_t const& type, nvtxMessageValue_t const& value) noexcept
+        : type_{type}
+        , value_(value)
+    {}
+
+    /**
+     * @brief Construct a `message` from NVTX C API registered string handle.
+     *
+     * @param handle nvtxStringHandle_t value of registered string handle
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 message(nvtxStringHandle_t handle) noexcept
+        : type_{NVTX_MESSAGE_TYPE_REGISTERED}
+    {
+      value_.registered = handle;
+    }
+
+    /**
+     * @brief Return the union holding the value of the message.
+     *
+     */
+    constexpr value_type get_value() const noexcept
+    {
+      return value_;
+    }
+
+    /**
+     * @brief Return the type information about the value the union holds.
+     *
+     */
+    constexpr nvtxMessageType_t get_type() const noexcept
+    {
+      return type_;
+    }
+
+  private:
+    nvtxMessageType_t type_{}; ///< message type
+    nvtxMessageValue_t value_{}; ///< message contents
+  };
+
+  /**
+   * @brief A numerical value that can be associated with an NVTX event via
+   * its `event_attributes`.
+   *
+   * Example:
+   * \code{.cpp}
+   * // Constructs a payload from the int32_t value 42
+   * nvtx3:: event_attributes attr{nvtx3::payload{42}};
+   *
+   * // `range0` will have an int32_t payload of 42
+   * nvtx3::scoped_range range0{attr};
+   *
+   * // range1 has double payload of 3.14
+   * nvtx3::scoped_range range1{nvtx3::payload{3.14}};
+   * \endcode
+   */
+  class payload
+  {
+  public:
+    using value_type = typename nvtxEventAttributes_v2::payload_t;
+
+    /**
+     * @brief Construct a `payload` from a signed, 8 byte integer.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(int64_t value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_INT64}
+        , value_{}
+    {
+      value_.llValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from a signed, 4 byte integer.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(int32_t value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_INT32}
+        , value_{}
+    {
+      value_.iValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from an unsigned, 8 byte integer.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(uint64_t value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_UNSIGNED_INT64}
+        , value_{}
+    {
+      value_.ullValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from an unsigned, 4 byte integer.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(uint32_t value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_UNSIGNED_INT32}
+        , value_{}
+    {
+      value_.uiValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from a single-precision floating point
+     * value.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(float value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_FLOAT}
+        , value_{}
+    {
+      value_.fValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from a double-precision floating point
+     * value.
+     *
+     * @param value Value to use as contents of the payload
+     */
+    NVTX3_CONSTEXPR_IF_CPP14 explicit payload(double value) noexcept
+        : type_{NVTX_PAYLOAD_TYPE_DOUBLE}
+        , value_{}
+    {
+      value_.dValue = value;
+    }
+
+    /**
+     * @brief Construct a `payload` from NVTX C API type and value.
+     *
+     * @param type nvtxPayloadType_t enum value indicating type of the payload
+     * @param value nvtxEventAttributes_t::payload_t union containing payload
+     */
+    constexpr payload(nvtxPayloadType_t const& type, value_type const& value) noexcept
+        : type_{type}
+        , value_(value)
+    {}
+
+    /**
+     * @brief Return the union holding the value of the payload
+     *
+     */
+    constexpr value_type get_value() const noexcept
+    {
+      return value_;
+    }
+
+    /**
+     * @brief Return the information about the type the union holds.
+     *
+     */
+    constexpr nvtxPayloadType_t get_type() const noexcept
+    {
+      return type_;
+    }
+
+  private:
+    nvtxPayloadType_t type_; ///< Type of the payload value
+    value_type value_; ///< Union holding the payload value
+  };
+
+  /**
+   * @brief Describes the attributes of a NVTX event.
+   *
+   * NVTX events can be customized via four "attributes":
+   *
+   * - color:    color used to visualize the event in tools such as Nsight
+   *             Systems. See `color`.
+   * - message:  Custom message string. See `message`.
+   * - payload:  User-defined numerical value. See `payload`.
+   * - category: Intra-domain grouping. See `category`.
+   *
+   * These component attributes are specified via an `event_attributes` object.
+   * See `nvtx3::color`, `nvtx3::message`, `nvtx3::payload`, and
+   * `nvtx3::category` for how these individual attributes are constructed.
+   *
+   * While it is possible to specify all four attributes, it is common to want
+   * to only specify a subset of attributes and use default values for the
+   * others. For convenience, `event_attributes` can be constructed from any
+   * number of attribute components in any order.
+   *
+   * Example:
+   * \code{.cpp}
+   * // Set message, same as using nvtx3::message{"message"}
+   * event_attributes attr{"message"};
+   *
+   * // Set message and color
+   * event_attributes attr{"message", nvtx3::rgb{127, 255, 0}};
+   *
+   * // Set message, color, payload, category
+   * event_attributes attr{"message",
+   *                       nvtx3::rgb{127, 255, 0},
+   *                       nvtx3::payload{42},
+   *                       nvtx3::category{1}};
+   *
+   * // Same as above -- can use any order of arguments
+   * event_attributes attr{nvtx3::payload{42},
+   *                       nvtx3::category{1},
+   *                       "message",
+   *                       nvtx3::rgb{127, 255, 0}};
+   *
+   * // Multiple arguments of the same type are allowed, but only the first is
+   * // used -- in this example, payload is set to 42:
+   * event_attributes attr{ nvtx3::payload{42}, nvtx3::payload{7} };
+   *
+   * // Range `r` will be customized according the attributes in `attr`
+   * nvtx3::scoped_range r{attr};
+   *
+   * // For convenience, `event_attributes` constructor arguments may be passed
+   * // to the `scoped_range_in` constructor -- they are forwarded to the
+   * // `event_attributes` constructor
+   * nvtx3::scoped_range r{nvtx3::payload{42}, nvtx3::category{1}, "message"};
+   *
+   * // Using the nvtx3 namespace in a local scope makes the syntax more succinct:
+   * using namespace nvtx3;
+   * scoped_range r{payload{42}, category{1}, "message"};
+   * \endcode
+   *
+   */
+  class event_attributes
+  {
+  public:
+    using value_type = nvtxEventAttributes_t;
+
+    /**
+     * @brief Default constructor creates an `event_attributes` with no
+     * category, color, payload, nor message.
+     */
+    constexpr event_attributes() noexcept
+        : attributes_{
+            NVTX_VERSION, // version
+            sizeof(nvtxEventAttributes_t), // size
+            0, // category
+            NVTX_COLOR_UNKNOWN, // color type
+            0, // color value
+            NVTX_PAYLOAD_UNKNOWN, // payload type
+            0, // reserved 4B
+            {0}, // payload value (union)
+            NVTX_MESSAGE_UNKNOWN, // message type
+            {0} // message value (union)
+          }
+    {}
+
+    /**
+     * @brief Variadic constructor where the first argument is a `category`.
+     *
+     * Sets the value of the `EventAttribute`s category based on `c` and
+     * forwards the remaining variadic parameter pack to the next constructor.
+     *
+     */
+    template <typename... Args>
+    NVTX3_CONSTEXPR_IF_CPP14 explicit event_attributes(category const& c, Args const&... args) noexcept
+        : event_attributes(args...)
+    {
+      attributes_.category = c.get_id();
+    }
+
+    /**
+     * @brief Variadic constructor where the first argument is a `color`.
+     *
+     * Sets the value of the `EventAttribute`s color based on `c` and forwards
+     * the remaining variadic parameter pack to the next constructor.
+     *
+     */
+    template <typename... Args>
+    NVTX3_CONSTEXPR_IF_CPP14 explicit event_attributes(color const& c, Args const&... args) noexcept
+        : event_attributes(args...)
+    {
+      attributes_.color     = c.get_value();
+      attributes_.colorType = c.get_type();
+    }
+
+    /**
+     * @brief Variadic constructor where the first argument is a `payload`.
+     *
+     * Sets the value of the `EventAttribute`s payload based on `p` and forwards
+     * the remaining variadic parameter pack to the next constructor.
+     *
+     */
+    template <typename... Args>
+    NVTX3_CONSTEXPR_IF_CPP14 explicit event_attributes(payload const& p, Args const&... args) noexcept
+        : event_attributes(args...)
+    {
+      attributes_.payload     = p.get_value();
+      attributes_.payloadType = p.get_type();
+    }
+
+    /**
+     * @brief Variadic constructor where the first argument is a `message`.
+     *
+     * Sets the value of the `EventAttribute`s message based on `m` and forwards
+     * the remaining variadic parameter pack to the next constructor.
+     *
+     */
+    template <typename... Args>
+    NVTX3_CONSTEXPR_IF_CPP14 explicit event_attributes(message const& m, Args const&... args) noexcept
+        : event_attributes(args...)
+    {
+      attributes_.message     = m.get_value();
+      attributes_.messageType = m.get_type();
+    }
+
+    ~event_attributes()                                  = default;
+    event_attributes(event_attributes const&)            = default;
+    event_attributes& operator=(event_attributes const&) = default;
+    event_attributes(event_attributes&&)                 = default;
+    event_attributes& operator=(event_attributes&&)      = default;
+
+    /**
+     * @brief Get raw pointer to underlying NVTX attributes object.
+     *
+     */
+    constexpr value_type const* get() const noexcept
+    {
+      return &attributes_;
+    }
+
+  private:
+    value_type attributes_{}; ///< The NVTX attributes structure
+  };
+
+  /**
+   * @brief A RAII object for creating a NVTX range local to a thread within a
+   * domain.
+   *
+   * When constructed, begins a nested NVTX range on the calling thread in the
+   * specified domain. Upon destruction, ends the NVTX range.
+   *
+   * Behavior is undefined if a `scoped_range_in` object is
+   * created/destroyed on different threads.
+   *
+   * `scoped_range_in` is neither movable nor copyable.
+   *
+   * `scoped_range_in`s may be nested within other ranges.
+   *
+   * The domain of the range is specified by the template type parameter `D`.
+   * By default, the `domain::global` is used, which scopes the range to the
+   * global NVTX domain. The convenience alias `scoped_range` is provided for
+   * ranges scoped to the global domain.
+   *
+   * A custom domain can be defined by creating a type, `D`, with a static
+   * member `D::name` whose value is used to name the domain associated with
+   * `D`. `D::name` must resolve to either `char const*` or `wchar_t const*`
+   *
+   * Example:
+   * \code{.cpp}
+   * // Define a type `my_domain` with a member `name` used to name the domain
+   * // associated with the type `my_domain`.
+   * struct my_domain{
+   *    static constexpr char const* name{"my domain"};
+   * };
+   * \endcode
+   *
+   * Usage:
+   * \code{.cpp}
+   * nvtx3::scoped_range_in<my_domain> r1{"range 1"}; // Range in my domain
+   *
+   * // Three equivalent ways to make a range in the global domain:
+   * nvtx3::scoped_range_in<nvtx3::domain::global> r2{"range 2"};
+   * nvtx3::scoped_range_in<> r3{"range 3"};
+   * nvtx3::scoped_range r4{"range 4"};
+   *
+   * // Create an alias to succinctly make ranges in my domain:
+   * using my_scoped_range = nvtx3::scoped_range_in<my_domain>;
+   *
+   * my_scoped_range r3{"range 3"};
+   * \endcode
+   */
+  template <class D = domain::global>
+  class NVTX3_MAYBE_UNUSED scoped_range_in
+  {
+  public:
+    /**
+     * @brief Construct a `scoped_range_in` with the specified
+     * `event_attributes`
+     *
+     * Example:
+     * \code{cpp}
+     * nvtx3::event_attributes attr{"msg", nvtx3::rgb{127,255,0}};
+     * nvtx3::scoped_range range{attr}; // Creates a range with message contents
+     *                                  // "msg" and green color
+     * \endcode
+     *
+     * @param[in] attr `event_attributes` that describes the desired attributes
+     * of the range.
+     */
+    explicit scoped_range_in(event_attributes const& attr) noexcept
+    {
+#  ifndef NVTX_DISABLE
+      nvtxDomainRangePushEx(domain::get<D>(), attr.get());
+#  else
+      (void) attr;
+#  endif
+    }
+
+    /**
+     * @brief Constructs a `scoped_range_in` from the constructor arguments
+     * of an `event_attributes`.
+     *
+     * Forwards the arguments `args...` to construct an
+     * `event_attributes` object. The `event_attributes` object is then
+     * associated with the `scoped_range_in`.
+     *
+     * For more detail, see `event_attributes` documentation.
+     *
+     * Example:
+     * \code{cpp}
+     * // Creates a range with message "message" and green color
+     * nvtx3::scoped_range r{"message", nvtx3::rgb{127,255,0}};
+     * \endcode
+     *
+     * @param[in] args Arguments to used to construct an `event_attributes` associated with this
+     * range.
+     *
+     */
+    template <typename... Args>
+    explicit scoped_range_in(Args const&... args) noexcept
+        : scoped_range_in{event_attributes{args...}}
+    {}
+
+    /**
+     * @brief Default constructor creates a `scoped_range_in` with no
+     * message, color, payload, nor category.
+     *
+     */
+    scoped_range_in() noexcept
+        : scoped_range_in{event_attributes{}}
+    {}
+
+    /**
+     * @brief Delete `operator new` to disallow heap allocated objects.
+     *
+     * `scoped_range_in` must follow RAII semantics to guarantee proper push/pop semantics.
+     *
+     */
+    void* operator new(std::size_t) = delete;
+
+    scoped_range_in(scoped_range_in const&)            = delete;
+    scoped_range_in& operator=(scoped_range_in const&) = delete;
+    scoped_range_in(scoped_range_in&&)                 = delete;
+    scoped_range_in& operator=(scoped_range_in&&)      = delete;
+
+    /**
+     * @brief Destroy the scoped_range_in, ending the NVTX range event.
+     */
+    ~scoped_range_in() noexcept
+    {
+#  ifndef NVTX_DISABLE
+      nvtxDomainRangePop(domain::get<D>());
+#  endif
+    }
+  };
+
+  /**
+   * @brief Alias for a `scoped_range_in` in the global NVTX domain.
+   *
+   */
+  using scoped_range = scoped_range_in<domain::global>;
+
+  namespace detail
+  {
+  /// @cond internal
+  template <typename D = domain::global>
+  class NVTX3_MAYBE_UNUSED optional_scoped_range_in
+  {
+  public:
+    optional_scoped_range_in() = default;
+
+    void begin(event_attributes const& attr) noexcept
+    {
+#  ifndef NVTX_DISABLE
+      // This class is not meant to be part of the public NVTX C++ API and should
+      // only be used in the `NVTX3_FUNC_RANGE_IF` and `NVTX3_FUNC_RANGE_IF_IN`
+      // macros. However, to prevent developers from misusing this class, make
+      // sure to not start multiple ranges.
+      if (initialized)
+      {
+        return;
+      }
+
+      nvtxDomainRangePushEx(domain::get<D>(), attr.get());
+      initialized = true;
+#  endif
+    }
+
+    ~optional_scoped_range_in() noexcept
+    {
+#  ifndef NVTX_DISABLE
+      if (initialized)
+      {
+        nvtxDomainRangePop(domain::get<D>());
+      }
+#  endif
+    }
+
+    void* operator new(std::size_t)                                      = delete;
+    optional_scoped_range_in(optional_scoped_range_in const&)            = delete;
+    optional_scoped_range_in& operator=(optional_scoped_range_in const&) = delete;
+    optional_scoped_range_in(optional_scoped_range_in&&)                 = delete;
+    optional_scoped_range_in& operator=(optional_scoped_range_in&&)      = delete;
+
+  private:
+#  ifndef NVTX_DISABLE
+    bool initialized = false;
+#  endif
+  };
+  /// @endcond
+  } // namespace detail
+
+  /**
+   * @brief Handle used for correlating explicit range start and end events.
+   *
+   * A handle is "null" if it does not correspond to any range.
+   *
+   */
+  struct range_handle
+  {
+    /// Type used for the handle's value
+    using value_type = nvtxRangeId_t;
+
+    /**
+     * @brief Construct a `range_handle` from the given id.
+     *
+     */
+    constexpr explicit range_handle(value_type id) noexcept
+        : _range_id{id}
+    {}
+
+    /**
+     * @brief Constructs a null range handle.
+     *
+     * A null range_handle corresponds to no range. Calling `end_range` on a
+     * null handle is undefined behavior when a tool is active.
+     *
+     */
+    constexpr range_handle() noexcept = default;
+
+    /**
+     * @brief Checks whether this handle is null
+     *
+     * Provides contextual conversion to `bool`.
+     *
+     * \code{cpp}
+     * range_handle handle{};
+     * if (handle) {...}
+     * \endcode
+     *
+     */
+    constexpr explicit operator bool() const noexcept
+    {
+      return get_value() != null_range_id;
+    }
+
+    /**
+     * @brief Implicit conversion from `nullptr` constructs a null handle.
+     *
+     * Satisfies the "NullablePointer" requirement to make `range_handle` comparable with `nullptr`.
+     *
+     */
+    constexpr range_handle(std::nullptr_t) noexcept {}
+
+    /**
+     * @brief Returns the `range_handle`'s value
+     *
+     * @return value_type The handle's value
+     */
+    constexpr value_type get_value() const noexcept
+    {
+      return _range_id;
+    }
+
+  private:
+    /// Sentinel value for a null handle that corresponds to no range
+    static constexpr value_type null_range_id = nvtxRangeId_t{0};
+
+    value_type _range_id{null_range_id}; ///< The underlying NVTX range id
+  };
+
+  /**
+   * @brief Compares two range_handles for equality
+   *
+   * @param lhs The first range_handle to compare
+   * @param rhs The second range_handle to compare
+   */
+  inline constexpr bool operator==(range_handle lhs, range_handle rhs) noexcept
+  {
+    return lhs.get_value() == rhs.get_value();
+  }
+
+  /**
+   * @brief Compares two range_handles for inequality
+   *
+   * @param lhs The first range_handle to compare
+   * @param rhs The second range_handle to compare
+   */
+  inline constexpr bool operator!=(range_handle lhs, range_handle rhs) noexcept
+  {
+    return !(lhs == rhs);
+  }
+
+  /**
+   * @brief Manually begin an NVTX range.
+   *
+   * Explicitly begins an NVTX range and returns a unique handle. To end the
+   * range, pass the handle to `end_range_in<D>()`.
+   *
+   * `nvtx3::start_range(...)` is equivalent to `nvtx3::start_range_in<>(...)` and
+   * `nvtx3::start_range_in<nvtx3::domain::global>(...)`.
+   *
+   * `start_range_in/end_range_in` are the most explicit and lowest level APIs
+   * provided for creating ranges.  Use of `nvtx3::unique_range_in` should be
+   * preferred unless one is unable to tie the range to the lifetime of an object.
+   *
+   * Example:
+   * \code{.cpp}
+   * nvtx3::event_attributes attr{"msg", nvtx3::rgb{127,255,0}};
+   * // Manually begin a range
+   * nvtx3::range_handle h = nvtx3::start_range_in<my_domain>(attr);
+   * ...
+   * nvtx3::end_range_in<my_domain>(h); // End the range
+   * \endcode
+   *
+   * @tparam D Type containing `name` member used to identify the `domain`
+   * to which the range belongs. Else, `domain::global` to indicate that the
+   * global NVTX domain should be used.
+   * @param[in] attr `event_attributes` that describes the desired attributes
+   * of the range.
+   * @return Unique handle to be passed to `end_range_in` to end the range.
+   */
+  template <typename D = domain::global>
+  NVTX3_NO_DISCARD inline range_handle start_range_in(event_attributes const& attr) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    return range_handle{nvtxDomainRangeStartEx(domain::get<D>(), attr.get())};
+#  else
+    (void) attr;
+    return {};
+#  endif
+  }
+
+  /**
+   * @brief Manually begin an NVTX range.
+   *
+   * Explicitly begins an NVTX range and returns a unique handle. To end the
+   * range, pass the handle to `end_range_in<D>()`.
+   *
+   * `nvtx3::start_range(...)` is equivalent to `nvtx3::start_range_in<>(...)` and
+   * `nvtx3::start_range_in<nvtx3::domain::global>(...)`.
+   *
+   * `start_range_in/end_range_in` are the most explicit and lowest level APIs
+   * provided for creating ranges.  Use of `nvtx3::unique_range_in` should be
+   * preferred unless one is unable to tie the range to the lifetime of an object.
+   *
+   * This overload uses `args...` to construct an  `event_attributes` to
+   * associate with the range.  For more detail, see `event_attributes`.
+   *
+   * Example:
+   * \code{cpp}
+   * // Manually begin a range
+   * nvtx3::range_handle h = nvtx3::start_range_in<D>("msg", nvtx3::rgb{127,255,0});
+   * ...
+   * nvtx3::end_range_in<D>(h); // Ends the range
+   * \endcode
+   *
+   * @tparam D Type containing `name` member used to identify the `domain`
+   * to which the range belongs. Else, `domain::global` to indicate that the
+   * global NVTX domain should be used.
+   * @param[in] args Variadic parameter pack of the arguments for an `event_attributes`.
+   * @return Unique handle to be passed to `end_range` to end the range.
+   */
+  template <typename D = domain::global, typename... Args>
+  NVTX3_NO_DISCARD inline range_handle start_range_in(Args const&... args) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    return start_range_in<D>(event_attributes{args...});
+#  else
+    return {};
+#  endif
+  }
+
+  /**
+   * @brief Manually begin an NVTX range in the global domain.
+   *
+   * Explicitly begins an NVTX range and returns a unique handle. To end the
+   * range, pass the handle to `end_range()`.
+   *
+   * `nvtx3::start_range(...)` is equivalent to `nvtx3::start_range_in<>(...)` and
+   * `nvtx3::start_range_in<nvtx3::domain::global>(...)`.
+   *
+   * `start_range/end_range` are the most explicit and lowest level APIs
+   * provided for creating ranges.  Use of `nvtx3::unique_range` should be
+   * preferred unless one is unable to tie the range to the lifetime of an object.
+   *
+   * Example:
+   * \code{.cpp}
+   * nvtx3::event_attributes attr{"msg", nvtx3::rgb{127,255,0}};
+   * // Manually begin a range
+   * nvtx3::range_handle h = nvtx3::start_range(attr);
+   * ...
+   * nvtx3::end_range(h); // End the range
+   * \endcode
+   *
+   * @param[in] attr `event_attributes` that describes the desired attributes
+   * of the range.
+   * @return Unique handle to be passed to `end_range_in` to end the range.
+   */
+  NVTX3_NO_DISCARD inline range_handle start_range(event_attributes const& attr) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    return start_range_in<domain::global>(attr);
+#  else
+    (void) attr;
+    return {};
+#  endif
+  }
+
+  /**
+   * @brief Manually begin an NVTX range in the global domain.
+   *
+   * Explicitly begins an NVTX range and returns a unique handle. To end the
+   * range, pass the handle to `end_range_in<D>()`.
+   *
+   * `nvtx3::start_range(...)` is equivalent to `nvtx3::start_range_in<>(...)` and
+   * `nvtx3::start_range_in<nvtx3::domain::global>(...)`.
+   *
+   * `start_range_in/end_range_in` are the most explicit and lowest level APIs
+   * provided for creating ranges.  Use of `nvtx3::unique_range_in` should be
+   * preferred unless one is unable to tie the range to the lifetime of an object.
+   *
+   * This overload uses `args...` to construct an  `event_attributes` to
+   * associate with the range.  For more detail, see `event_attributes`.
+   *
+   * Example:
+   * \code{cpp}
+   * // Manually begin a range
+   * nvtx3::range_handle h = nvtx3::start_range("msg", nvtx3::rgb{127,255,0});
+   * ...
+   * nvtx3::end_range(h); // Ends the range
+   * \endcode
+   *
+   * @param[in] args Variadic parameter pack of the arguments for an `event_attributes`.
+   * @return Unique handle to be passed to `end_range` to end the range.
+   */
+  template <typename... Args>
+  NVTX3_NO_DISCARD inline range_handle start_range(Args const&... args) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    return start_range_in<domain::global>(args...);
+#  else
+    return {};
+#  endif
+  }
+
+  /**
+   * @brief Manually end the range associated with the handle `r` in domain `D`.
+   *
+   * Explicitly ends the NVTX range indicated by the handle `r` returned from a
+   * prior call to `start_range_in<D>`. The range may end on a different thread
+   * from where it began.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain` to
+   * which the range belongs. Else, `domain::global` to indicate that the global
+   * NVTX domain should be used.
+   * @param r Handle to a range started by a prior call to `start_range_in`.
+   *
+   * @warning The domain type specified as template parameter to this function
+   * must be the same that was specified on the associated `start_range_in` call.
+   */
+  template <typename D = domain::global>
+  inline void end_range_in(range_handle r) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    nvtxDomainRangeEnd(domain::get<D>(), r.get_value());
+#  else
+    (void) r;
+#  endif
+  }
+
+  /**
+   * @brief Manually end the range associated with the handle `r` in the global
+   * domain.
+   *
+   * Explicitly ends the NVTX range indicated by the handle `r` returned from a
+   * prior call to `start_range`. The range may end on a different thread from
+   * where it began.
+   *
+   * @param r Handle to a range started by a prior call to `start_range`.
+   *
+   * @warning The domain type specified as template parameter to this function
+   * must be the same that was specified on the associated `start_range` call.
+   */
+  inline void end_range(range_handle r) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    end_range_in<domain::global>(r);
+#  else
+    (void) r;
+#  endif
+  }
+
+  /**
+   * @brief A RAII object for creating a NVTX range within a domain that can
+   * be created and destroyed on different threads.
+   *
+   * When constructed, begins a NVTX range in the specified domain. Upon
+   * destruction, ends the NVTX range.
+   *
+   * Similar to `nvtx3::scoped_range_in`, with a few key differences:
+   * - `unique_range` objects can be destroyed in an order whereas `scoped_range` objects must be
+   *    destroyed in exact reverse creation order
+   * - `unique_range` can start and end on different threads
+   * - `unique_range` is movable
+   * - `unique_range` objects can be constructed as heap objects
+   *
+   * There is extra overhead associated with `unique_range` constructs and therefore use of
+   * `nvtx3::scoped_range_in` should be preferred.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain`
+   * to which the `unique_range_in` belongs. Else, `domain::global` to
+   * indicate that the global NVTX domain should be used.
+   */
+  template <typename D = domain::global>
+  class NVTX3_MAYBE_UNUSED unique_range_in
+  {
+  public:
+    /**
+     * @brief Construct a new unique_range_in object with the specified event attributes
+     *
+     * Example:
+     * \code{cpp}
+     * nvtx3::event_attributes attr{"msg", nvtx3::rgb{127,255,0}};
+     * nvtx3::unique_range_in<my_domain> range{attr}; // Creates a range with message contents
+     *                                            // "msg" and green color
+     * \endcode
+     *
+     * @param[in] attr `event_attributes` that describes the desired attributes
+     * of the range.
+     */
+    explicit unique_range_in(event_attributes const& attr) noexcept
+        : handle_{start_range_in<D>(attr)}
+    {}
+
+    /**
+     * @brief Constructs a `unique_range_in` from the constructor arguments
+     * of an `event_attributes`.
+     *
+     * Forwards the arguments `args...` to construct an
+     * `event_attributes` object. The `event_attributes` object is then
+     * associated with the `unique_range_in`.
+     *
+     * For more detail, see `event_attributes` documentation.
+     *
+     * Example:
+     * \code{.cpp}
+     * // Creates a range with message "message" and green color
+     * nvtx3::unique_range_in<> r{"message", nvtx3::rgb{127,255,0}};
+     * \endcode
+     *
+     * @param[in] args Variadic parameter pack of arguments to construct an `event_attributes`
+     * associated with this range.
+     */
+    template <typename... Args>
+    explicit unique_range_in(Args const&... args) noexcept
+        : unique_range_in{event_attributes{args...}}
+    {}
+
+    /**
+     * @brief Default constructor creates a `unique_range_in` with no
+     * message, color, payload, nor category.
+     *
+     */
+    constexpr unique_range_in() noexcept
+        : unique_range_in{event_attributes{}}
+    {}
+
+    /**
+     * @brief Destroy the `unique_range_in` ending the range.
+     *
+     */
+    ~unique_range_in() noexcept = default;
+
+    /**
+     * @brief Move constructor allows taking ownership of the NVTX range from
+     * another `unique_range_in`.
+     *
+     * @param other The range to take ownership of
+     */
+    unique_range_in(unique_range_in&& other) noexcept = default;
+
+    /**
+     * @brief Move assignment operator allows taking ownership of an NVTX range
+     * from another `unique_range_in`.
+     *
+     * @param other The range to take ownership of
+     */
+    unique_range_in& operator=(unique_range_in&& other) noexcept = default;
+
+    /// Copy construction is not allowed to prevent multiple objects from owning
+    /// the same range handle
+    unique_range_in(unique_range_in const&) = delete;
+
+    /// Copy assignment is not allowed to prevent multiple objects from owning the
+    /// same range handle
+    unique_range_in& operator=(unique_range_in const&) = delete;
+
+  private:
+    struct end_range_handle
+    {
+      using pointer = range_handle; /// Override the pointer type of the unique_ptr
+      void operator()(range_handle h) const noexcept
+      {
+        end_range_in<D>(h);
+      }
+    };
+
+    /// Range handle used to correlate the start/end of the range
+    std::unique_ptr<range_handle, end_range_handle> handle_;
+  };
+
+  /**
+   * @brief Alias for a `unique_range_in` in the global NVTX domain.
+   *
+   */
+  using unique_range = unique_range_in<domain::global>;
+
+  /**
+   * @brief Annotates an instantaneous point in time with a "marker", using the
+   * attributes specified by `attr`.
+   *
+   * Unlike a "range" which has a beginning and an end, a marker is a single event
+   * in an application, such as detecting a problem:
+   *
+   * \code{.cpp}
+   * bool success = do_operation(...);
+   * if (!success) {
+   *    nvtx3::event_attributes attr{"operation failed!", nvtx3::rgb{255,0,0}};
+   *    nvtx3::mark_in<my_domain>(attr);
+   * }
+   * \endcode
+   *
+   * Note that nvtx3::mark_in<D> is a function, not a class like scoped_range_in<D>.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain`
+   * to which the `unique_range_in` belongs. Else, `domain::global` to
+   * indicate that the global NVTX domain should be used.
+   * @param[in] attr `event_attributes` that describes the desired attributes
+   * of the mark.
+   */
+  template <typename D = domain::global>
+  inline void mark_in(event_attributes const& attr) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    nvtxDomainMarkEx(domain::get<D>(), attr.get());
+#  else
+    (void) (attr);
+#  endif
+  }
+
+  /**
+   * @brief Annotates an instantaneous point in time with a "marker", using the
+   * arguments to construct an `event_attributes`.
+   *
+   * Unlike a "range" which has a beginning and an end, a marker is a single event
+   * in an application, such as detecting a problem:
+   *
+   * \code{.cpp}
+   * bool success = do_operation(...);
+   * if (!success) {
+   *    nvtx3::mark_in<my_domain>("operation failed!", nvtx3::rgb{255,0,0});
+   * }
+   * \endcode
+   *
+   * Note that nvtx3::mark_in<D> is a function, not a class like scoped_range_in<D>.
+   *
+   * Forwards the arguments `args...` to construct an `event_attributes` object.
+   * The attributes are then associated with the marker. For more detail, see
+   * the `event_attributes` documentation.
+   *
+   * @tparam D Type containing `name` member used to identify the `domain`
+   * to which the `unique_range_in` belongs. Else `domain::global` to
+   * indicate that the global NVTX domain should be used.
+   * @param[in] args Variadic parameter pack of arguments to construct an `event_attributes`
+   * associated with this range.
+   *
+   */
+  template <typename D = domain::global, typename... Args>
+  inline void mark_in(Args const&... args) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    mark_in<D>(event_attributes{args...});
+#  endif
+  }
+
+  /**
+   * @brief Annotates an instantaneous point in time with a "marker", using the
+   * attributes specified by `attr`, in the global domain.
+   *
+   * Unlike a "range" which has a beginning and an end, a marker is a single event
+   * in an application, such as detecting a problem:
+   *
+   * \code{.cpp}
+   * bool success = do_operation(...);
+   * if (!success) {
+   *    nvtx3::event_attributes attr{"operation failed!", nvtx3::rgb{255,0,0}};
+   *    nvtx3::mark(attr);
+   * }
+   * \endcode
+   *
+   * Note that nvtx3::mark is a function, not a class like scoped_range.
+   *
+   * @param[in] attr `event_attributes` that describes the desired attributes
+   * of the mark.
+   */
+  inline void mark(event_attributes const& attr) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    mark_in<domain::global>(attr);
+#  endif
+  }
+
+  /**
+   * @brief Annotates an instantaneous point in time with a "marker", using the
+   * arguments to construct an `event_attributes`, in the global domain.
+   *
+   * Unlike a "range" which has a beginning and an end, a marker is a single event
+   * in an application, such as detecting a problem:
+   *
+   * \code{.cpp}
+   * bool success = do_operation(...);
+   * if (!success) {
+   *    nvtx3::mark("operation failed!", nvtx3::rgb{255,0,0});
+   * }
+   * \endcode
+   *
+   * Note that nvtx3::mark is a function, not a class like scoped_range.
+   *
+   * Forwards the arguments `args...` to construct an `event_attributes` object.
+   * The attributes are then associated with the marker. For more detail, see
+   * the `event_attributes` documentation.
+   *
+   * @param[in] args Variadic parameter pack of arguments to construct an
+   * `event_attributes` associated with this range.
+   *
+   */
+  template <typename... Args>
+  inline void mark(Args const&... args) noexcept
+  {
+#  ifndef NVTX_DISABLE
+    mark_in<domain::global>(args...);
+#  endif
+  }
+
+} // namespace NVTX3_VERSION_NAMESPACE
+} // namespace nvtx3
+
+#  ifndef NVTX_DISABLE
+/**
+ * @brief Convenience macro for generating a range in the specified `domain`
+ * from the lifetime of a function
+ *
+ * This macro is useful for generating an NVTX range in `domain` from
+ * the entry point of a function to its exit. It is intended to be the first
+ * line of the function.
+ *
+ * Constructs a static `registered_string_in` using the name of the immediately
+ * enclosing function returned by `__func__` and constructs a
+ * `nvtx3::scoped_range` using the registered function name as the range's
+ * message.
+ *
+ * Example:
+ * \code{.cpp}
+ * struct my_domain{static constexpr char const* name{"my_domain"};};
+ *
+ * void foo(...) {
+ *    NVTX3_FUNC_RANGE_IN(my_domain); // Range begins on entry to foo()
+ *    // do stuff
+ *    ...
+ * } // Range ends on return from foo()
+ * \endcode
+ *
+ * @param[in] D Type containing `name` member used to identify the
+ * `domain` to which the `registered_string_in` belongs. Else,
+ * `domain::global` to  indicate that the global NVTX domain should be used.
+ */
+#    define NVTX3_V1_FUNC_RANGE_IN(D)                                                  \
+      static ::nvtx3::v1::registered_string_in<D> const nvtx3_func_name__{__func__};   \
+      static ::nvtx3::v1::event_attributes const nvtx3_func_attr__{nvtx3_func_name__}; \
+      ::nvtx3::v1::scoped_range_in<D> const nvtx3_range__{nvtx3_func_attr__};
+
+/**
+ * @brief Convenience macro for generating a range in the specified `domain`
+ * from the lifetime of a function if the given boolean expression evaluates
+ * to true.
+ *
+ * Similar to `NVTX3_V1_FUNC_RANGE_IN(D)`, the only difference being that
+ * `NVTX3_V1_FUNC_RANGE_IF_IN(D, C)` only generates a range if the given boolean
+ * expression evaluates to true.
+ *
+ * @param[in] D Type containing `name` member used to identify the
+ * `domain` to which the `registered_string_in` belongs. Else,
+ * `domain::global` to indicate that the global NVTX domain should be used.
+ *
+ * @param[in] C Boolean expression used to determine if a range should be
+ * generated.
+ */
+#    define NVTX3_V1_FUNC_RANGE_IF_IN(D, C)                                              \
+      ::nvtx3::v1::detail::optional_scoped_range_in<D> optional_nvtx3_range__;           \
+      if (C)                                                                             \
+      {                                                                                  \
+        static ::nvtx3::v1::registered_string_in<D> const nvtx3_func_name__{__func__};   \
+        static ::nvtx3::v1::event_attributes const nvtx3_func_attr__{nvtx3_func_name__}; \
+        optional_nvtx3_range__.begin(nvtx3_func_attr__);                                 \
+      }
+#  else
+#    define NVTX3_V1_FUNC_RANGE_IN(D)
+#    define NVTX3_V1_FUNC_RANGE_IF_IN(D, C)
+#  endif // NVTX_DISABLE
+
+/**
+ * @brief Convenience macro for generating a range in the global domain from the
+ * lifetime of a function.
+ *
+ * This macro is useful for generating an NVTX range in the global domain from
+ * the entry point of a function to its exit. It is intended to be the first
+ * line of the function.
+ *
+ * Constructs a static `registered_string_in` using the name of the immediately
+ * enclosing function returned by `__func__` and constructs a
+ * `nvtx3::scoped_range` using the registered function name as the range's
+ * message.
+ *
+ * Example:
+ * \code{.cpp}
+ * void foo(...) {
+ *    NVTX3_FUNC_RANGE(); // Range begins on entry to foo()
+ *    // do stuff
+ *    ...
+ * } // Range ends on return from foo()
+ * \endcode
+ */
+#  define NVTX3_V1_FUNC_RANGE() NVTX3_V1_FUNC_RANGE_IN(::nvtx3::v1::domain::global)
+
+/**
+ * @brief Convenience macro for generating a range in the global domain from the
+ * lifetime of a function if the given boolean expression evaluates to true.
+ *
+ * Similar to `NVTX3_V1_FUNC_RANGE()`, the only difference being that
+ * `NVTX3_V1_FUNC_RANGE_IF(C)` only generates a range if the given boolean
+ * expression evaluates to true.
+ *
+ * @param[in] C Boolean expression used to determine if a range should be
+ * generated.
+ */
+#  define NVTX3_V1_FUNC_RANGE_IF(C) NVTX3_V1_FUNC_RANGE_IF_IN(::nvtx3::v1::domain::global, C)
+
+/* When inlining this version, versioned macros must have unversioned aliases.
+ * For each NVTX3_Vx_ #define, make an NVTX3_ alias of it here.*/
+#  if defined(NVTX3_INLINE_THIS_VERSION)
+/* clang format off */
+#    define NVTX3_FUNC_RANGE       NVTX3_V1_FUNC_RANGE
+#    define NVTX3_FUNC_RANGE_IF    NVTX3_V1_FUNC_RANGE_IF
+#    define NVTX3_FUNC_RANGE_IN    NVTX3_V1_FUNC_RANGE_IN
+#    define NVTX3_FUNC_RANGE_IF_IN NVTX3_V1_FUNC_RANGE_IF_IN
+/* clang format on */
+#  endif
+
+#endif // NVTX3_CPP_DEFINITIONS_V1_0
+
+/* Add functionality for new minor versions here, by copying the above section enclosed
+ * in #ifndef NVTX3_CPP_DEFINITIONS_Vx_y, and incrementing the minor version.  This code
+ * is an example of how additions for version 1.2 would look, indented for clarity.  Note
+ * that the versioned symbols and macros are always provided, and the unversioned symbols
+ * are only provided if NVTX3_INLINE_THIS_VERSION was defined at the top of this header.
+ *
+ * \code{.cpp}
+ * #ifndef NVTX3_CPP_DEFINITIONS_V1_2
+ * #define NVTX3_CPP_DEFINITIONS_V1_2
+ *     namespace nvtx3 {
+ *         NVTX3_INLINE_IF_REQUESTED namespace NVTX3_VERSION_NAMESPACE {
+ *             class new_class {};
+ *             inline void new_function() {}
+ *         }
+ *     }
+ *
+ *     // Macros must have the major version in their names:
+ *     #define NVTX3_V1_NEW_MACRO_A() ...
+ *     #define NVTX3_V1_NEW_MACRO_B() ...
+ *
+ *     // If inlining, make aliases for the macros with the version number omitted
+ *     #if defined(NVTX3_INLINE_THIS_VERSION)
+ *         #define NVTX3_NEW_MACRO_A NVTX3_V1_NEW_MACRO_A
+ *         #define NVTX3_NEW_MACRO_B NVTX3_V1_NEW_MACRO_B
+ *     #endif
+ * #endif // NVTX3_CPP_DEFINITIONS_V1_2
+ * \endcode
+ */
+
+/* Undefine all temporarily-defined unversioned macros, which would conflict with
+ * subsequent includes of different versions of this header. */
+#undef NVTX3_CPP_VERSION_MAJOR
+#undef NVTX3_CPP_VERSION_MINOR
+#undef NVTX3_CONCAT
+#undef NVTX3_NAMESPACE_FOR
+#undef NVTX3_VERSION_NAMESPACE
+#undef NVTX3_INLINE_IF_REQUESTED
+#undef NVTX3_CONSTEXPR_IF_CPP14
+#undef NVTX3_MAYBE_UNUSED
+#undef NVTX3_NO_DISCARD
+
+#if defined(NVTX3_INLINE_THIS_VERSION)
+#  undef NVTX3_INLINE_THIS_VERSION
+#endif
+
+#if defined(NVTX3_USE_CHECKED_OVERLOADS_FOR_GET_DEFINED_HERE)
+#  undef NVTX3_USE_CHECKED_OVERLOADS_FOR_GET_DEFINED_HERE
+#  undef NVTX3_USE_CHECKED_OVERLOADS_FOR_GET
+#endif
+
+#if defined(NVTX3_STATIC_ASSERT_DEFINED_HERE)
+#  undef NVTX3_STATIC_ASSERT_DEFINED_HERE
+#  undef NVTX3_STATIC_ASSERT
+#endif
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/generated/get_sreg.h b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/generated/get_sreg.h
new file mode 100644
index 00000000..d0441d2f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/generated/get_sreg.h
@@ -0,0 +1,949 @@
+// This file was automatically generated. Do not edit.
+
+#ifndef _CUDA_PTX_GENERATED_GET_SREG_H_
+#define _CUDA_PTX_GENERATED_GET_SREG_H_
+
+/*
+// mov.u32 sreg_value, %%tid.x; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_tid_x();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_tid_x()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%tid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%tid.y; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_tid_y();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_tid_y()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%tid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%tid.z; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_tid_z();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_tid_z()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%tid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ntid.x; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ntid_x();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ntid_x()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%ntid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ntid.y; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ntid_y();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ntid_y()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%ntid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ntid.z; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ntid_z();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ntid_z()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%ntid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%laneid; // PTX ISA 13
+template <typename = void>
+__device__ static inline uint32_t get_sreg_laneid();
+*/
+#if __cccl_ptx_isa >= 130
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_laneid()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%laneid;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 130
+
+/*
+// mov.u32 sreg_value, %%warpid; // PTX ISA 13
+template <typename = void>
+__device__ static inline uint32_t get_sreg_warpid();
+*/
+#if __cccl_ptx_isa >= 130
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_warpid()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%warpid;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 130
+
+/*
+// mov.u32 sreg_value, %%nwarpid; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nwarpid();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_nwarpid_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nwarpid()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%nwarpid;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_nwarpid_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ctaid.x; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ctaid_x();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ctaid_x()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%ctaid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ctaid.y; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ctaid_y();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ctaid_y()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%ctaid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%ctaid.z; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_ctaid_z();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_ctaid_z()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%ctaid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%nctaid.x; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nctaid_x();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nctaid_x()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nctaid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%nctaid.y; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nctaid_y();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nctaid_y()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nctaid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%nctaid.z; // PTX ISA 20
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nctaid_z();
+*/
+#if __cccl_ptx_isa >= 200
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nctaid_z()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nctaid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%smid; // PTX ISA 13
+template <typename = void>
+__device__ static inline uint32_t get_sreg_smid();
+*/
+#if __cccl_ptx_isa >= 130
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_smid()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%smid;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 130
+
+/*
+// mov.u32 sreg_value, %%nsmid; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nsmid();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_nsmid_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nsmid()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%nsmid;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_nsmid_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u64 sreg_value, %%gridid; // PTX ISA 30
+template <typename = void>
+__device__ static inline uint64_t get_sreg_gridid();
+*/
+#if __cccl_ptx_isa >= 300
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint64_t get_sreg_gridid()
+{
+  ::cuda::std::uint64_t __sreg_value;
+  asm("mov.u64 %0, %%gridid;" : "=l"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 300
+
+/*
+// mov.pred sreg_value, %%is_explicit_cluster; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline bool get_sreg_is_explicit_cluster();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_is_explicit_cluster_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline bool get_sreg_is_explicit_cluster()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("{\n\t .reg .pred P_OUT; \n\t"
+      "mov.pred P_OUT, %%is_explicit_cluster;\n\t"
+      "selp.b32 %0, 1, 0, P_OUT; \n"
+      "}"
+      : "=r"(__sreg_value)
+      :
+      :);
+  return static_cast<bool>(__sreg_value);
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_is_explicit_cluster_is_not_supported_before_SM_90__();
+  return false;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%clusterid.x; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_clusterid_x();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_clusterid_x_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_clusterid_x()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%clusterid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_clusterid_x_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%clusterid.y; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_clusterid_y();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_clusterid_y_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_clusterid_y()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%clusterid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_clusterid_y_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%clusterid.z; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_clusterid_z();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_clusterid_z_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_clusterid_z()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%clusterid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_clusterid_z_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%nclusterid.x; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nclusterid_x();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_nclusterid_x_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nclusterid_x()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nclusterid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_nclusterid_x_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%nclusterid.y; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nclusterid_y();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_nclusterid_y_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nclusterid_y()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nclusterid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_nclusterid_y_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%nclusterid.z; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_nclusterid_z();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_nclusterid_z_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_nclusterid_z()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%nclusterid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_nclusterid_z_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_ctaid.x; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_ctaid_x();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_ctaid_x_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_ctaid_x()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_ctaid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_ctaid_x_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_ctaid.y; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_ctaid_y();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_ctaid_y_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_ctaid_y()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_ctaid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_ctaid_y_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_ctaid.z; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_ctaid_z();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_ctaid_z_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_ctaid_z()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_ctaid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_ctaid_z_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_nctaid.x; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_nctaid_x();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_nctaid_x_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_nctaid_x()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_nctaid.x;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_nctaid_x_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_nctaid.y; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_nctaid_y();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_nctaid_y_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_nctaid_y()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_nctaid.y;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_nctaid_y_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_nctaid.z; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_nctaid_z();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_nctaid_z_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_nctaid_z()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_nctaid.z;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_nctaid_z_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_ctarank; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_ctarank();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_ctarank_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_ctarank()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_ctarank;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_ctarank_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%cluster_nctarank; // PTX ISA 78, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_cluster_nctarank();
+*/
+#if __cccl_ptx_isa >= 780
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_cluster_nctarank_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_cluster_nctarank()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%cluster_nctarank;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_cluster_nctarank_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 780
+
+/*
+// mov.u32 sreg_value, %%lanemask_eq; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_lanemask_eq();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_lanemask_eq_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_lanemask_eq()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%lanemask_eq;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_lanemask_eq_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%lanemask_le; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_lanemask_le();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_lanemask_le_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_lanemask_le()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%lanemask_le;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_lanemask_le_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%lanemask_lt; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_lanemask_lt();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_lanemask_lt_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_lanemask_lt()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%lanemask_lt;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_lanemask_lt_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%lanemask_ge; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_lanemask_ge();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_lanemask_ge_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_lanemask_ge()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%lanemask_ge;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_lanemask_ge_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%lanemask_gt; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_lanemask_gt();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_lanemask_gt_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_lanemask_gt()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%lanemask_gt;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_lanemask_gt_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u32 sreg_value, %%clock; // PTX ISA 10
+template <typename = void>
+__device__ static inline uint32_t get_sreg_clock();
+*/
+#if __cccl_ptx_isa >= 100
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_clock()
+{
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%clock;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+}
+#endif // __cccl_ptx_isa >= 100
+
+/*
+// mov.u32 sreg_value, %%clock_hi; // PTX ISA 50, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_clock_hi();
+*/
+#if __cccl_ptx_isa >= 500
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_clock_hi_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_clock_hi()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%clock_hi;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_clock_hi_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 500
+
+/*
+// mov.u64 sreg_value, %%clock64; // PTX ISA 20, SM_35
+template <typename = void>
+__device__ static inline uint64_t get_sreg_clock64();
+*/
+#if __cccl_ptx_isa >= 200
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_clock64_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint64_t get_sreg_clock64()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint64_t __sreg_value;
+  asm volatile("mov.u64 %0, %%clock64;" : "=l"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_clock64_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 200
+
+/*
+// mov.u64 sreg_value, %%globaltimer; // PTX ISA 31, SM_35
+template <typename = void>
+__device__ static inline uint64_t get_sreg_globaltimer();
+*/
+#if __cccl_ptx_isa >= 310
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_globaltimer_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint64_t get_sreg_globaltimer()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint64_t __sreg_value;
+  asm volatile("mov.u64 %0, %%globaltimer;" : "=l"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_globaltimer_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 310
+
+/*
+// mov.u32 sreg_value, %%globaltimer_lo; // PTX ISA 31, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_globaltimer_lo();
+*/
+#if __cccl_ptx_isa >= 310
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_globaltimer_lo_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_globaltimer_lo()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%globaltimer_lo;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_globaltimer_lo_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 310
+
+/*
+// mov.u32 sreg_value, %%globaltimer_hi; // PTX ISA 31, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_globaltimer_hi();
+*/
+#if __cccl_ptx_isa >= 310
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_globaltimer_hi_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_globaltimer_hi()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm volatile("mov.u32 %0, %%globaltimer_hi;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_globaltimer_hi_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 310
+
+/*
+// mov.u32 sreg_value, %%total_smem_size; // PTX ISA 41, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_total_smem_size();
+*/
+#if __cccl_ptx_isa >= 410
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_total_smem_size_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_total_smem_size()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%total_smem_size;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_total_smem_size_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 410
+
+/*
+// mov.u32 sreg_value, %%aggr_smem_size; // PTX ISA 81, SM_90
+template <typename = void>
+__device__ static inline uint32_t get_sreg_aggr_smem_size();
+*/
+#if __cccl_ptx_isa >= 810
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_aggr_smem_size_is_not_supported_before_SM_90__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_aggr_smem_size()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 900
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%aggr_smem_size;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_aggr_smem_size_is_not_supported_before_SM_90__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 810
+
+/*
+// mov.u32 sreg_value, %%dynamic_smem_size; // PTX ISA 41, SM_35
+template <typename = void>
+__device__ static inline uint32_t get_sreg_dynamic_smem_size();
+*/
+#if __cccl_ptx_isa >= 410
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_dynamic_smem_size_is_not_supported_before_SM_35__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint32_t get_sreg_dynamic_smem_size()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 350
+  ::cuda::std::uint32_t __sreg_value;
+  asm("mov.u32 %0, %%dynamic_smem_size;" : "=r"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_dynamic_smem_size_is_not_supported_before_SM_35__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 410
+
+/*
+// mov.u64 sreg_value, %%current_graph_exec; // PTX ISA 80, SM_50
+template <typename = void>
+__device__ static inline uint64_t get_sreg_current_graph_exec();
+*/
+#if __cccl_ptx_isa >= 800
+extern "C" _CCCL_DEVICE void __cuda_ptx_get_sreg_current_graph_exec_is_not_supported_before_SM_50__();
+template <typename = void>
+_CCCL_DEVICE static inline ::cuda::std::uint64_t get_sreg_current_graph_exec()
+{
+#  if _CCCL_CUDA_COMPILER(NVHPC) || __CUDA_ARCH__ >= 500
+  ::cuda::std::uint64_t __sreg_value;
+  asm("mov.u64 %0, %%current_graph_exec;" : "=l"(__sreg_value) : :);
+  return __sreg_value;
+#  else
+  // Unsupported architectures will have a linker error with a semi-decent error message
+  __cuda_ptx_get_sreg_current_graph_exec_is_not_supported_before_SM_50__();
+  return 0;
+#  endif
+}
+#endif // __cccl_ptx_isa >= 800
+
+#endif // _CUDA_PTX_GENERATED_GET_SREG_H_
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/get_sreg.h b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/get_sreg.h
new file mode 100644
index 00000000..1ac117c2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/instructions/get_sreg.h
@@ -0,0 +1,43 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_PTX_GET_SREG_H_
+#define _CUDA_PTX_GET_SREG_H_
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__ptx/ptx_dot_variants.h>
+#include <cuda/__ptx/ptx_helper_functions.h>
+#include <cuda/std/cstdint>
+
+#include <nv/target> // __CUDA_MINIMUM_ARCH__ and friends
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_PTX
+
+// 10. Special Registers
+// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#special-registers
+#include <cuda/__ptx/instructions/generated/get_sreg.h>
+
+_CCCL_END_NAMESPACE_CUDA_PTX
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_PTX_GET_SREG_H_
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_dot_variants.h b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_dot_variants.h
new file mode 100644
index 00000000..d923a548
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_dot_variants.h
@@ -0,0 +1,230 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// WARNING: The source of truth for this file is libcuda-ptx. Do not modify without syncing with libcuda-ptx.
+
+#ifndef _CUDA_PTX_DOT_VARIANTS_H_
+#define _CUDA_PTX_DOT_VARIANTS_H_
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+/*
+ * Public integral constant types and values for ".variant"s:
+ *
+ * - .sem:     acquire, release, ..
+ * - .space:   global, shared, constant, ..
+ * - .scope:   cta, cluster, gpu, ..
+ * - .op:      add, min, cas, ..
+ *
+ * For each .variant, the code below defines:
+ * - An enum `dot_variant` with each possible value
+ * - A type template `variant_t<dot_variant>`
+ * - Types `variant_A_t`, ..., `variant_Z_t`
+ * - Constexpr values `variant_A` of type `variant_A_t`
+ *
+ * These types enable specifying fine-grained overloads of a PTX binding. If a
+ * binding can handle multiple variants, then it is defined as:
+ *
+ * template <dot_variant var>
+ * [...] void ptx_binding(variant_t<var> __v) { ... }
+ *
+ * If it only handles a single variant, then it is defined as:
+ *
+ * [...] void ptx_binding(variant_A __v) { ... }
+ *
+ * If two variants have different behaviors or return types (see .space
+ * overloads of mbarrier.arrive.expect_tx for an example), then these can be
+ * provided as separate overloads of the same function:
+ *
+ * [...] void ptx_binding(variant_A __v) { ... }
+ * [...] int ptx_binding(variant_B __v) { ... }
+ *
+ */
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_PTX
+
+// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#operation-types
+enum class dot_sem
+{
+  acq_rel,
+  acquire,
+  relaxed,
+  release,
+  sc,
+  weak
+};
+
+// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#state-spaces
+enum class dot_space
+{
+  global,
+  cluster, // The PTX spelling is shared::cluster
+  shared, // The PTX spelling is shared::cta
+
+  // The following state spaces are unlikely to be used in cuda::ptx in the near
+  // future, so they are not exposed:
+
+  // reg,
+  // sreg,
+  // const_mem, // Using const_mem as `const` is reserved in C++.
+  // local,
+  // param,
+  // tex // deprecated
+};
+
+// https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#scope
+enum class dot_scope
+{
+  cta,
+  cluster,
+  gpu,
+  sys
+};
+
+enum class dot_op
+{
+  add,
+  dec,
+  inc,
+  max,
+  min,
+  and_op, // Using and_op, as `and, or, xor` are reserved in C++.
+  or_op,
+  xor_op,
+  cas,
+  exch
+};
+
+enum class dot_cta_group
+{
+  cta_group_1,
+  cta_group_2
+};
+
+enum class dot_kind
+{
+  f16,
+  f8f6f4,
+  i8,
+  mxf4,
+  mxf4nvf4,
+  mxf8f6f4,
+  tf32
+};
+
+template <dot_sem __sem>
+using sem_t         = ::cuda::std::integral_constant<dot_sem, __sem>;
+using sem_acq_rel_t = sem_t<dot_sem::acq_rel>;
+using sem_acquire_t = sem_t<dot_sem::acquire>;
+using sem_relaxed_t = sem_t<dot_sem::relaxed>;
+using sem_release_t = sem_t<dot_sem::release>;
+using sem_sc_t      = sem_t<dot_sem::sc>;
+using sem_weak_t    = sem_t<dot_sem::weak>;
+
+[[maybe_unused]] static constexpr sem_acq_rel_t sem_acq_rel{};
+[[maybe_unused]] static constexpr sem_acquire_t sem_acquire{};
+[[maybe_unused]] static constexpr sem_relaxed_t sem_relaxed{};
+[[maybe_unused]] static constexpr sem_release_t sem_release{};
+[[maybe_unused]] static constexpr sem_sc_t sem_sc{};
+[[maybe_unused]] static constexpr sem_weak_t sem_weak{};
+
+template <dot_space __spc>
+using space_t         = ::cuda::std::integral_constant<dot_space, __spc>;
+using space_global_t  = space_t<dot_space::global>;
+using space_shared_t  = space_t<dot_space::shared>;
+using space_cluster_t = space_t<dot_space::cluster>;
+
+[[maybe_unused]] static constexpr space_global_t space_global{};
+[[maybe_unused]] static constexpr space_shared_t space_shared{};
+[[maybe_unused]] static constexpr space_cluster_t space_cluster{};
+
+template <dot_scope __scope>
+using scope_t         = ::cuda::std::integral_constant<dot_scope, __scope>;
+using scope_cluster_t = scope_t<dot_scope::cluster>;
+using scope_cta_t     = scope_t<dot_scope::cta>;
+using scope_gpu_t     = scope_t<dot_scope::gpu>;
+using scope_sys_t     = scope_t<dot_scope::sys>;
+
+[[maybe_unused]] static constexpr scope_cluster_t scope_cluster{};
+[[maybe_unused]] static constexpr scope_cta_t scope_cta{};
+[[maybe_unused]] static constexpr scope_gpu_t scope_gpu{};
+[[maybe_unused]] static constexpr scope_sys_t scope_sys{};
+
+template <dot_op __op>
+using op_t        = ::cuda::std::integral_constant<dot_op, __op>;
+using op_add_t    = op_t<dot_op::add>;
+using op_dec_t    = op_t<dot_op::dec>;
+using op_inc_t    = op_t<dot_op::inc>;
+using op_max_t    = op_t<dot_op::max>;
+using op_min_t    = op_t<dot_op::min>;
+using op_and_op_t = op_t<dot_op::and_op>;
+using op_or_op_t  = op_t<dot_op::or_op>;
+using op_xor_op_t = op_t<dot_op::xor_op>;
+using op_cas_t    = op_t<dot_op::cas>;
+using op_exch_t   = op_t<dot_op::exch>;
+
+[[maybe_unused]] static constexpr op_add_t op_add{};
+[[maybe_unused]] static constexpr op_dec_t op_dec{};
+[[maybe_unused]] static constexpr op_inc_t op_inc{};
+[[maybe_unused]] static constexpr op_max_t op_max{};
+[[maybe_unused]] static constexpr op_min_t op_min{};
+[[maybe_unused]] static constexpr op_and_op_t op_and_op{};
+[[maybe_unused]] static constexpr op_or_op_t op_or_op{};
+[[maybe_unused]] static constexpr op_xor_op_t op_xor_op{};
+[[maybe_unused]] static constexpr op_cas_t op_cas{};
+[[maybe_unused]] static constexpr op_exch_t op_exch{};
+
+template <dot_cta_group __cta_group>
+using cta_group_t   = ::cuda::std::integral_constant<dot_cta_group, __cta_group>;
+using cta_group_1_t = cta_group_t<dot_cta_group::cta_group_1>;
+using cta_group_2_t = cta_group_t<dot_cta_group::cta_group_2>;
+
+[[maybe_unused]] static constexpr cta_group_1_t cta_group_1{};
+[[maybe_unused]] static constexpr cta_group_2_t cta_group_2{};
+
+template <dot_kind __kind>
+using kind_t          = ::cuda::std::integral_constant<dot_kind, __kind>;
+using kind_f16_t      = kind_t<dot_kind::f16>;
+using kind_f8f6f4_t   = kind_t<dot_kind::f8f6f4>;
+using kind_i8_t       = kind_t<dot_kind::i8>;
+using kind_mxf4_t     = kind_t<dot_kind::mxf4>;
+using kind_mxf4nvf4_t = kind_t<dot_kind::mxf4nvf4>;
+using kind_mxf8f6f4_t = kind_t<dot_kind::mxf8f6f4>;
+using kind_tf32_t     = kind_t<dot_kind::tf32>;
+
+[[maybe_unused]] static constexpr kind_f16_t kind_f16{};
+[[maybe_unused]] static constexpr kind_f8f6f4_t kind_f8f6f4{};
+[[maybe_unused]] static constexpr kind_i8_t kind_i8{};
+[[maybe_unused]] static constexpr kind_mxf4_t kind_mxf4{};
+[[maybe_unused]] static constexpr kind_mxf4nvf4_t kind_mxf4nvf4{};
+[[maybe_unused]] static constexpr kind_mxf8f6f4_t kind_mxf8f6f4{};
+[[maybe_unused]] static constexpr kind_tf32_t kind_tf32{};
+
+template <int n>
+using n32_t = ::cuda::std::integral_constant<int, n>;
+
+_CCCL_END_NAMESPACE_CUDA_PTX
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_PTX_DOT_VARIANTS_H_
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_helper_functions.h b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_helper_functions.h
new file mode 100644
index 00000000..08f8a7f9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__ptx/ptx_helper_functions.h
@@ -0,0 +1,178 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_PTX_HELPER_FUNCTIONS_H_
+#define _CUDA_PTX_HELPER_FUNCTIONS_H_
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/cstddef>
+#include <cuda/std/cstdint>
+
+#if _CCCL_CUDA_COMPILATION()
+
+#  include <cuda/std/__cccl/prologue.h>
+
+#  if defined(__CUDACC__) || defined(_NVHPC_CUDA) || defined(__CUDACC_RTC__)
+#    define _CUDA_PTX_CUDACC_MAJOR() __CUDACC_VER_MAJOR__
+#  elif defined(__CUDA__) && defined(__clang__)
+#    define _CUDA_PTX_CUDACC_MAJOR() (CUDA_VERSION / 1000)
+#  endif // ^^^ has cuda compiler ^^^
+
+#  if !defined(_LIBCUDA_PTX_ARCH_SPECIFIC)
+#    if defined(__CUDA_ARCH_SPECIFIC__)
+#      define _LIBCUDA_PTX_ARCH_SPECIFIC() __CUDA_ARCH_SPECIFIC__
+#    else
+#      if defined(__CUDA_ARCH_FEAT_SM90_ALL)
+#        define _LIBCUDA_PTX_ARCH_SPECIFIC() 900
+#      elif defined(__CUDA_ARCH_FEAT_SM100_ALL)
+#        define _LIBCUDA_PTX_ARCH_SPECIFIC() 1000
+#      elif defined(__CUDA_ARCH_FEAT_SM103_ALL)
+#        define _LIBCUDA_PTX_ARCH_SPECIFIC() 1030
+#      elif defined(__CUDA_ARCH_FEAT_SM120_ALL)
+#        define _LIBCUDA_PTX_ARCH_SPECIFIC() 1200
+#      else
+#        define _LIBCUDA_PTX_ARCH_SPECIFIC() 0
+#      endif
+#    endif // ^^^ !defined(__CUDA_ARCH_SPECIFIC__)
+#  endif // ^^^ !defined(_LIBCUDA_PTX_ARCH_SPECIFIC)
+
+#  if !defined(__CUDA_HAS_ARCH_FAMILY_SPECIFIC)
+
+#    define __CUDA_HAS_ARCH_FAMILY_SPECIFIC(N) false
+
+#  endif // !defined(__CUDA_HAS_ARCH_FAMILY_SPECIFIC)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_PTX
+
+#  if _CUDA_PTX_CUDACC_MAJOR() < 13
+struct alignas(32) longlong4_32a
+{
+  long long x, y, z, w;
+};
+struct alignas(32) ulonglong4_32a
+{
+  unsigned long long x, y, z, w;
+};
+struct alignas(32) double4_32a
+{
+  double x, y, z, w;
+};
+#  else
+using ::double4_32a;
+using ::longlong4_32a;
+using ::ulonglong4_32a;
+#  endif // _CUDA_PTX_CUDACC_MAJOR() < 13
+
+/*************************************************************
+ *
+ * Conversion from generic pointer -> state space "pointer"
+ *
+ **************************************************************/
+_CCCL_DEVICE_API inline ::cuda::std::uint32_t __as_ptr_smem(const void* __ptr)
+{
+  // Consider adding debug asserts here.
+  return static_cast<::cuda::std::uint32_t>(::__cvta_generic_to_shared(__ptr));
+}
+
+_CCCL_DEVICE_API inline ::cuda::std::uint32_t __as_ptr_dsmem(const void* __ptr)
+{
+  // No difference in implementation to __as_ptr_smem.
+  return __as_ptr_smem(__ptr);
+}
+
+_CCCL_DEVICE_API inline ::cuda::std::uint32_t __as_ptr_remote_dsmem(const void* __ptr)
+{
+  // No difference in implementation to __as_ptr_smem.
+  // Consider adding debug asserts here.
+  return __as_ptr_smem(__ptr);
+}
+
+_CCCL_DEVICE_API inline ::cuda::std::uint64_t __as_ptr_gmem(const void* __ptr)
+{
+  // Consider adding debug asserts here.
+  return static_cast<::cuda::std::uint64_t>(::__cvta_generic_to_global(__ptr));
+}
+
+/*************************************************************
+ *
+ * Conversion from state space "pointer" -> generic pointer
+ *
+ **************************************************************/
+template <typename _Tp>
+_CCCL_DEVICE_API _Tp* __from_ptr_smem(::cuda::std::size_t __ptr)
+{
+  // Consider adding debug asserts here.
+  return reinterpret_cast<_Tp*>(::__cvta_shared_to_generic(__ptr));
+}
+
+template <typename _Tp>
+_CCCL_DEVICE_API _Tp* __from_ptr_dsmem(::cuda::std::size_t __ptr)
+{
+  // Consider adding debug asserts here.
+  return __from_ptr_smem<_Tp>(__ptr);
+}
+
+template <typename _Tp>
+_CCCL_DEVICE_API _Tp* __from_ptr_remote_dsmem(::cuda::std::size_t __ptr)
+{
+  // Consider adding debug asserts here.
+  return __from_ptr_smem<_Tp>(__ptr);
+}
+
+template <typename _Tp>
+_CCCL_DEVICE_API _Tp* __from_ptr_gmem(::cuda::std::size_t __ptr)
+{
+  // Consider adding debug asserts here.
+  return reinterpret_cast<_Tp*>(::__cvta_global_to_generic(__ptr));
+}
+
+/*************************************************************
+ *
+ * Conversion to and from b8 type
+ *
+ **************************************************************/
+
+template <typename _B8>
+_CCCL_DEVICE_API uint32_t __b8_as_u32(_B8 __val)
+{
+  static_assert(sizeof(_B8) == 1);
+  ::cuda::std::uint32_t __u32 = 0;
+  ::memcpy(&__u32, &__val, 1);
+  return __u32;
+}
+
+template <typename _B8>
+_CCCL_DEVICE_API _B8 __u32_as_b8(uint32_t __u32)
+{
+  static_assert(sizeof(_B8) == 1);
+  _B8 b8;
+  ::memcpy(&b8, &__u32, 1);
+  return b8;
+}
+
+_CCCL_END_NAMESPACE_CUDA_PTX
+
+#  include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CCCL_CUDA_COMPILATION()
+
+#endif // _CUDA_PTX_HELPER_FUNCTIONS_H_
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/__type_traits/is_trivially_copyable.h b/qwen3_6_scripts/cccl_preload/include/cuda/__type_traits/is_trivially_copyable.h
new file mode 100644
index 00000000..5b1ab962
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/__type_traits/is_trivially_copyable.h
@@ -0,0 +1,115 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CUDA__TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
+#define __CUDA__TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__fwd/complex.h>
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__fwd/array.h>
+#include <cuda/std/__fwd/complex.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__type_traits/aggregate_members_all_of.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_aggregate.h>
+#include <cuda/std/__type_traits/is_trivially_copyable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+template <typename _Tp, typename = void>
+inline constexpr bool __is_aggregate_trivially_copyable_v = false;
+
+template <typename _Tp>
+inline constexpr bool __is_trivially_copyable_v =
+  ::cuda::std::is_trivially_copyable_v<_Tp> || __is_aggregate_trivially_copyable_v<_Tp>;
+
+#if _CCCL_HAS_NVFP16()
+
+template <>
+inline constexpr bool __is_trivially_copyable_v<::__half> = true;
+
+template <>
+inline constexpr bool __is_trivially_copyable_v<::__half2> = true;
+
+#endif // _CCCL_HAS_NVFP16()
+
+#if _CCCL_HAS_NVBF16()
+
+template <>
+inline constexpr bool __is_trivially_copyable_v<::__nv_bfloat16> = true;
+template <>
+inline constexpr bool __is_trivially_copyable_v<::__nv_bfloat162> = true;
+
+#endif // _CCCL_HAS_NVBF16()
+
+template <typename _Tp>
+inline constexpr bool __is_trivially_copyable_v<_Tp[]> = __is_trivially_copyable_v<_Tp>;
+
+template <typename _Tp, ::cuda::std::size_t _Size>
+inline constexpr bool __is_trivially_copyable_v<_Tp[_Size]> = __is_trivially_copyable_v<_Tp>;
+
+template <typename _Tp, ::cuda::std::size_t _Size>
+inline constexpr bool __is_trivially_copyable_v<::cuda::std::array<_Tp, _Size>> = __is_trivially_copyable_v<_Tp>;
+
+template <typename _T1, typename _T2>
+inline constexpr bool __is_trivially_copyable_v<::cuda::std::pair<_T1, _T2>> =
+  __is_trivially_copyable_v<_T1> && __is_trivially_copyable_v<_T2>;
+
+template <typename... _Ts>
+inline constexpr bool __is_trivially_copyable_v<::cuda::std::tuple<_Ts...>> = (__is_trivially_copyable_v<_Ts> && ...);
+
+template <typename _Tp>
+inline constexpr bool __is_trivially_copyable_v<complex<_Tp>> = true;
+
+template <typename _Tp>
+inline constexpr bool __is_trivially_copyable_v<::cuda::std::complex<_Tp>> = true;
+
+// if all the previous conditions fail, check if the type is an aggregate and all its members are trivially copyable
+template <typename _Tp>
+using __is_trivially_copyable_callable = ::cuda::std::bool_constant<__is_trivially_copyable_v<_Tp>>;
+
+template <typename _Tp>
+inline constexpr bool __is_aggregate_trivially_copyable_v<
+  _Tp,
+  ::cuda::std::enable_if_t<::cuda::std::is_aggregate_v<_Tp> && !::cuda::std::is_trivially_copyable_v<_Tp>>> =
+  ::cuda::std::__aggregate_all_of_v<__is_trivially_copyable_callable, _Tp>;
+
+//----------------------------------------------------------------------------------------------------------------------
+// public traits
+
+template <typename _Tp>
+inline constexpr bool is_trivially_copyable_v = __is_trivially_copyable_v<_Tp>;
+
+template <typename _Tp>
+inline constexpr bool is_trivially_copyable_v<const _Tp> = is_trivially_copyable_v<_Tp>;
+
+// defined as alias so users cannot specialize it (they should specialize the variable template instead)
+template <typename _Tp>
+using is_trivially_copyable = ::cuda::std::bool_constant<is_trivially_copyable_v<_Tp>>;
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA__TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/clamp.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/clamp.h
new file mode 100644
index 00000000..260b9223
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/clamp.h
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_CLAMP_H
+#define _CUDA_STD___ALGORITHM_CLAMP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp, class _Compare>
+[[nodiscard]] _CCCL_API constexpr const _Tp&
+clamp(const _Tp& __v _CCCL_LIFETIMEBOUND,
+      const _Tp& __lo _CCCL_LIFETIMEBOUND,
+      const _Tp& __hi _CCCL_LIFETIMEBOUND,
+      _Compare __comp)
+{
+  _CCCL_ASSERT(!__comp(__hi, __lo), "Bad bounds passed to cuda::std::clamp");
+  return __comp(__v, __lo) ? __lo : __comp(__hi, __v) ? __hi : __v;
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr const _Tp&
+clamp(const _Tp& __v _CCCL_LIFETIMEBOUND, const _Tp& __lo _CCCL_LIFETIMEBOUND, const _Tp& __hi _CCCL_LIFETIMEBOUND)
+{
+  _CCCL_ASSERT(!(__hi < __lo), "Bad bounds passed to cuda::std::clamp");
+  return __v < __lo ? __lo : __hi < __v ? __hi : __v;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_CLAMP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp.h
new file mode 100644
index 00000000..140b1633
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp.h
@@ -0,0 +1,58 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_COMP_H
+#define _CUDA_STD___ALGORITHM_COMP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#if defined(_LIBCUDACXX_HAS_STRING)
+#  include <cuda/std/__type_traits/predicate_traits.h>
+#endif // _LIBCUDACXX_HAS_STRING
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct __equal_to
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API constexpr bool operator()(const _T1& __lhs, const _T2& __rhs) const
+    noexcept(noexcept(__lhs == __rhs))
+  {
+    return __lhs == __rhs;
+  }
+};
+
+struct __less
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Tp, class _Up>
+  [[nodiscard]] _CCCL_API constexpr bool operator()(const _Tp& __lhs, const _Up& __rhs) const
+    noexcept(noexcept(__lhs < __rhs))
+  {
+    return __lhs < __rhs;
+  }
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_COMP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp_ref_type.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp_ref_type.h
new file mode 100644
index 00000000..b892fb6b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/comp_ref_type.h
@@ -0,0 +1,85 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_COMP_REF_TYPE_H
+#define _CUDA_STD___ALGORITHM_COMP_REF_TYPE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Compare>
+struct __debug_less
+{
+  _Compare& __comp_;
+  _CCCL_API constexpr __debug_less(_Compare& __c)
+      : __comp_(__c)
+  {}
+
+  template <class _Tp, class _Up>
+  [[nodiscard]] _CCCL_API constexpr bool operator()(const _Tp& __x, const _Up& __y)
+  {
+    bool __r = __comp_(__x, __y);
+    if (__r)
+    {
+      __do_compare_assert(0, __y, __x);
+    }
+    return __r;
+  }
+
+  template <class _Tp, class _Up>
+  [[nodiscard]] _CCCL_API constexpr bool operator()(_Tp& __x, _Up& __y)
+  {
+    bool __r = __comp_(__x, __y);
+    if (__r)
+    {
+      __do_compare_assert(0, __y, __x);
+    }
+    return __r;
+  }
+
+  template <class _LHS, class _RHS>
+  _CCCL_API constexpr decltype((void) declval<_Compare&>()(declval<_LHS&>(), declval<_RHS&>()))
+  __do_compare_assert(int, [[maybe_unused]] _LHS& __l, [[maybe_unused]] _RHS& __r)
+  {
+    _CCCL_ASSERT(!__comp_(__l, __r), "Comparator does not induce a strict weak ordering");
+  }
+
+  template <class _LHS, class _RHS>
+  _CCCL_API constexpr void __do_compare_assert(long, _LHS&, _RHS&)
+  {}
+};
+
+// Pass the comparator by lvalue reference. Or in debug mode, using a
+// debugging wrapper that stores a reference.
+#ifdef _CCCL_ENABLE_DEBUG_MODE
+template <class _Comp>
+using __comp_ref_type = __debug_less<_Comp>;
+#else // ^^^ _LIBCUDACXX_ENABLE_DEBUG_MODE ^^^ / vvv !_LIBCUDACXX_ENABLE_DEBUG_MODE vvv
+template <class _Comp>
+using __comp_ref_type = _Comp&;
+#endif // !_LIBCUDACXX_ENABLE_DEBUG_MODE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_COMP_REF_TYPE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/equal.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/equal.h
new file mode 100644
index 00000000..b708bcab
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/equal.h
@@ -0,0 +1,132 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_EQUAL_H
+#define _CUDA_STD___ALGORITHM_EQUAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__iterator/distance.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _InputIterator1, class _InputIterator2, class _BinaryPredicate>
+[[nodiscard]] _CCCL_API constexpr bool
+equal(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _BinaryPredicate __pred)
+{
+  bool __result = true;
+  for (; __first1 != __last1; ++__first1, (void) ++__first2)
+  {
+    if (!__pred(*__first1, *__first2))
+    {
+      __result = false;
+      break;
+    }
+  }
+  return __result;
+}
+
+template <class _InputIterator1, class _InputIterator2>
+[[nodiscard]] _CCCL_API constexpr bool equal(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2)
+{
+  return ::cuda::std::equal(__first1, __last1, __first2, __equal_to{});
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _BinaryPredicate, class _InputIterator1, class _InputIterator2>
+[[nodiscard]] _CCCL_API constexpr bool __equal(
+  _InputIterator1 __first1,
+  _InputIterator1 __last1,
+  _InputIterator2 __first2,
+  _InputIterator2 __last2,
+  _BinaryPredicate __pred,
+  input_iterator_tag,
+  input_iterator_tag)
+{
+  bool __result = true;
+  for (; __first1 != __last1 && __first2 != __last2; ++__first1, (void) ++__first2)
+  {
+    if (!__pred(*__first1, *__first2))
+    {
+      __result = false;
+      break;
+    }
+  }
+  return __result && __first1 == __last1 && __first2 == __last2;
+}
+
+template <class _BinaryPredicate, class _RandomAccessIterator1, class _RandomAccessIterator2>
+[[nodiscard]] _CCCL_API constexpr bool __equal(
+  _RandomAccessIterator1 __first1,
+  _RandomAccessIterator1 __last1,
+  _RandomAccessIterator2 __first2,
+  _RandomAccessIterator2 __last2,
+  _BinaryPredicate __pred,
+  random_access_iterator_tag,
+  random_access_iterator_tag)
+{
+  if (__last1 - __first1 != __last2 - __first2)
+  {
+    return false;
+  }
+  return ::cuda::std::equal<_RandomAccessIterator1, _RandomAccessIterator2, add_lvalue_reference_t<_BinaryPredicate>>(
+    __first1, __last1, __first2, __pred);
+}
+
+template <class _InputIterator1, class _InputIterator2, class _BinaryPredicate>
+[[nodiscard]] _CCCL_API constexpr bool
+equal(_InputIterator1 __first1,
+      _InputIterator1 __last1,
+      _InputIterator2 __first2,
+      _InputIterator2 __last2,
+      _BinaryPredicate __pred)
+{
+  return ::cuda::std::__equal<add_lvalue_reference_t<_BinaryPredicate>>(
+    __first1,
+    __last1,
+    __first2,
+    __last2,
+    __pred,
+    __iterator_traits_category_or_concept_t<_InputIterator1>(),
+    __iterator_traits_category_or_concept_t<_InputIterator2>());
+}
+
+template <class _InputIterator1, class _InputIterator2>
+[[nodiscard]] _CCCL_API constexpr bool
+equal(_InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _InputIterator2 __last2)
+{
+  return ::cuda::std::__equal(
+    __first1,
+    __last1,
+    __first2,
+    __last2,
+    __equal_to{},
+    __iterator_traits_category_or_concept_t<_InputIterator1>(),
+    __iterator_traits_category_or_concept_t<_InputIterator2>());
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_EQUAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/fill_n.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/fill_n.h
new file mode 100644
index 00000000..f10148bc
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/fill_n.h
@@ -0,0 +1,51 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_FILL_N_H
+#define _CUDA_STD___ALGORITHM_FILL_N_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__utility/convert_to_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _OutputIterator, class _Size, class _Tp>
+_CCCL_API constexpr _OutputIterator __fill_n(_OutputIterator __first, _Size __n, const _Tp& __value_)
+{
+  for (; __n > 0; ++__first, (void) --__n)
+  {
+    *__first = __value_;
+  }
+  return __first;
+}
+
+template <class _OutputIterator, class _Size, class _Tp>
+_CCCL_API constexpr _OutputIterator fill_n(_OutputIterator __first, _Size __n, const _Tp& __value_)
+{
+  return ::cuda::std::__fill_n(__first, __convert_to_integral(__n), __value_);
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_FILL_N_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iter_swap.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iter_swap.h
new file mode 100644
index 00000000..52b473dd
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iter_swap.h
@@ -0,0 +1,85 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_ITER_SWAP_H
+#define _CUDA_STD___ALGORITHM_ITER_SWAP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/swap.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+//! Intentionally not an algorithm to avoid breaking types that pull in `::std::iter_swap` via ADL
+_CCCL_BEGIN_NAMESPACE_CPO(__iter_swap)
+// "Poison pill" overload to intentionally create ambiguity with the unconstrained
+// `std::iter_swap` function.
+template <class _ForwardIterator1, class _ForwardIterator2>
+void iter_swap(_ForwardIterator1, _ForwardIterator2) = delete;
+
+template <class _ForwardIterator1, class _ForwardIterator2>
+_CCCL_CONCEPT __unqualified_iter_swap =
+  _CCCL_REQUIRES_EXPR((_ForwardIterator1, _ForwardIterator2), _ForwardIterator1&& __a, _ForwardIterator2&& __b)(
+    iter_swap(::cuda::std::forward<_ForwardIterator1>(__a), ::cuda::std::forward<_ForwardIterator2>(__b)));
+
+template <class _ForwardIterator1, class _ForwardIterator2>
+_CCCL_CONCEPT __readable_swappable =
+  _CCCL_REQUIRES_EXPR((_ForwardIterator1, _ForwardIterator2), _ForwardIterator1 __a, _ForwardIterator2 __b)(
+    requires(!__unqualified_iter_swap<_ForwardIterator1, _ForwardIterator2>), swap(*__a, *__b));
+
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _ForwardIterator1, class _ForwardIterator2)
+  _CCCL_REQUIRES(__unqualified_iter_swap<_ForwardIterator1, _ForwardIterator2>)
+  _CCCL_API constexpr void operator()(_ForwardIterator1&& __a, _ForwardIterator2&& __b) const
+    noexcept(noexcept(iter_swap(::cuda::std::declval<_ForwardIterator1>(), ::cuda::std::declval<_ForwardIterator2>())))
+  {
+    (void) iter_swap(::cuda::std::forward<_ForwardIterator1>(__a), ::cuda::std::forward<_ForwardIterator2>(__b));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _ForwardIterator1, class _ForwardIterator2)
+  _CCCL_REQUIRES(__readable_swappable<_ForwardIterator1, _ForwardIterator2>)
+  _CCCL_API constexpr void operator()(_ForwardIterator1&& __a, _ForwardIterator2&& __b) const
+    noexcept(noexcept(swap(*::cuda::std::declval<_ForwardIterator1>(), *::cuda::std::declval<_ForwardIterator2>())))
+  {
+    swap(*__a, *__b);
+  }
+};
+
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+// This is a global constant to avoid breaking types that pull in `::std::iter_swap` via ADL
+_CCCL_GLOBAL_CONSTANT auto iter_swap = __iter_swap::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __iter_swap_cpo = __iter_swap::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_ITER_SWAP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iterator_operations.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iterator_operations.h
new file mode 100644
index 00000000..829a95b6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/iterator_operations.h
@@ -0,0 +1,179 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_ITERATOR_OPERATIONS_H
+#define _CUDA_STD___ALGORITHM_ITERATOR_OPERATIONS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/iter_swap.h>
+#include <cuda/std/__algorithm/ranges_iterator_concept.h>
+#include <cuda/std/__iterator/advance.h>
+#include <cuda/std/__iterator/distance.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iter_move.h>
+#include <cuda/std/__iterator/iter_swap.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__iterator/next.h>
+#include <cuda/std/__iterator/prev.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _AlgPolicy>
+struct _IterOps;
+
+struct _RangeAlgPolicy
+{};
+
+template <>
+struct _IterOps<_RangeAlgPolicy>
+{
+  template <class _Iter>
+  using __value_type = iter_value_t<_Iter>;
+
+  template <class _Iter>
+  using __difference_type = iter_difference_t<_Iter>;
+
+  static constexpr auto advance      = ::cuda::std::ranges::__advance_cpo{};
+  static constexpr auto distance     = ::cuda::std::ranges::__distance_cpo{};
+  static constexpr auto __iter_move  = ::cuda::std::ranges::__iter_move_cpo{};
+  static constexpr auto iter_swap    = ::cuda::std::ranges::__iter_swap_cpo{};
+  static constexpr auto next         = ::cuda::std::ranges::__next_cpo{};
+  static constexpr auto prev         = ::cuda::std::ranges::__prev_cpo{};
+  static constexpr auto __advance_to = ::cuda::std::ranges::__advance_cpo{};
+};
+
+struct _ClassicAlgPolicy
+{};
+
+template <>
+struct _IterOps<_ClassicAlgPolicy>
+{
+  template <class _Iter>
+  using __value_type = typename iterator_traits<_Iter>::value_type;
+
+  template <class _Iter>
+  using __difference_type = typename iterator_traits<_Iter>::difference_type;
+
+  // advance
+  template <class _Iter, class _Distance>
+  _CCCL_API constexpr static void advance(_Iter& __iter, _Distance __count)
+  {
+    ::cuda::std::advance(__iter, __count);
+  }
+
+  // distance
+  template <class _Iter>
+  _CCCL_API constexpr static typename iterator_traits<_Iter>::difference_type distance(_Iter __first, _Iter __last)
+  {
+    return ::cuda::std::distance(__first, __last);
+  }
+
+  template <class _Iter>
+  using __deref_t = decltype(*::cuda::std::declval<_Iter&>());
+
+  template <class _Iter>
+  using __move_t = decltype(::cuda::std::move(*::cuda::std::declval<_Iter&>()));
+
+  template <class _Iter>
+  _CCCL_API constexpr static void __validate_iter_reference()
+  {
+    static_assert(
+      is_same_v<__deref_t<_Iter>, typename iterator_traits<remove_cvref_t<_Iter>>::reference>,
+      "It looks like your iterator's `iterator_traits<It>::reference` does not match the return type of "
+      "dereferencing the iterator, i.e., calling `*it`. This is undefined behavior according to [input.iterators] "
+      "and can lead to dangling reference issues at runtime, so we are flagging this.");
+  }
+
+  // iter_move
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter, enable_if_t<is_reference_v<__deref_t<_Iter>>, int> = 0>
+  _CCCL_API constexpr static
+    // If the result of dereferencing `_Iter` is a reference type, deduce the result of calling `::cuda::std::move` on
+    // it. Note that the C++03 mode doesn't support `decltype(auto)` as the return type.
+    __move_t<_Iter>
+    __iter_move(_Iter&& __i)
+  {
+    __validate_iter_reference<_Iter>();
+
+    return ::cuda::std::move(*::cuda::std::forward<_Iter>(__i));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter, enable_if_t<!is_reference_v<__deref_t<_Iter>>, int> = 0>
+  _CCCL_API constexpr static
+    // If the result of dereferencing `_Iter` is a value type, deduce the return value of this function to also be a
+    // value -- otherwise, after `operator*` returns a temporary, this function would return a dangling reference to
+    // that temporary. Note that the C++03 mode doesn't support `auto` as the return type.
+    __deref_t<_Iter>
+    __iter_move(_Iter&& __i)
+  {
+    __validate_iter_reference<_Iter>();
+
+    return *::cuda::std::forward<_Iter>(__i);
+  }
+
+  // iter_swap
+  template <class _Iter1, class _Iter2>
+  _CCCL_API constexpr static void iter_swap(_Iter1&& __a, _Iter2&& __b)
+  {
+    ::cuda::std::__iter_swap_cpo{}(::cuda::std::forward<_Iter1>(__a), ::cuda::std::forward<_Iter2>(__b));
+  }
+
+  // next
+  template <class _Iterator>
+  _CCCL_API static constexpr _Iterator next(_Iterator, _Iterator __last)
+  {
+    return __last;
+  }
+
+  template <class _Iter>
+  _CCCL_API static constexpr remove_cvref_t<_Iter> next(_Iter&& __it, __difference_type<remove_cvref_t<_Iter>> __n = 1)
+  {
+    return ::cuda::std::next(::cuda::std::forward<_Iter>(__it), __n);
+  }
+
+  // prev
+  template <class _Iter>
+  _CCCL_API static constexpr remove_cvref_t<_Iter> prev(_Iter&& __iter, __difference_type<remove_cvref_t<_Iter>> __n = 1)
+  {
+    return ::cuda::std::prev(::cuda::std::forward<_Iter>(__iter), __n);
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter>
+  _CCCL_API static constexpr void __advance_to(_Iter& __first, _Iter __last)
+  {
+    __first = __last;
+  }
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_ITERATOR_OPERATIONS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/lexicographical_compare.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/lexicographical_compare.h
new file mode 100644
index 00000000..39b146d0
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/lexicographical_compare.h
@@ -0,0 +1,70 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_LEXICOGRAPHICAL_COMPARE_H
+#define _CUDA_STD___ALGORITHM_LEXICOGRAPHICAL_COMPARE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__algorithm/comp_ref_type.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Compare, class _InputIterator1, class _InputIterator2>
+[[nodiscard]] _CCCL_API constexpr bool __lexicographical_compare(
+  _InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _InputIterator2 __last2, _Compare __comp)
+{
+  bool __result = false;
+  for (; __first2 != __last2; ++__first1, (void) ++__first2)
+  {
+    if (__first1 == __last1 || __comp(*__first1, *__first2))
+    {
+      __result = true;
+      break;
+    }
+    if (__comp(*__first2, *__first1))
+    {
+      break;
+    }
+  }
+  return __result;
+}
+
+template <class _InputIterator1, class _InputIterator2, class _Compare>
+[[nodiscard]] _CCCL_API constexpr bool lexicographical_compare(
+  _InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _InputIterator2 __last2, _Compare __comp)
+{
+  return __lexicographical_compare<__comp_ref_type<_Compare>>(__first1, __last1, __first2, __last2, __comp);
+}
+
+template <class _InputIterator1, class _InputIterator2>
+[[nodiscard]] _CCCL_API constexpr bool lexicographical_compare(
+  _InputIterator1 __first1, _InputIterator1 __last1, _InputIterator2 __first2, _InputIterator2 __last2)
+{
+  return ::cuda::std::lexicographical_compare(__first1, __last1, __first2, __last2, __less{});
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_LEXICOGRAPHICAL_COMPARE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max.h
new file mode 100644
index 00000000..bf1008da
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max.h
@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_MAX_H
+#define _CUDA_STD___ALGORITHM_MAX_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__algorithm/comp_ref_type.h>
+#include <cuda/std/__algorithm/max_element.h>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp, class _Compare>
+[[nodiscard]] _CCCL_API constexpr const _Tp&
+max(const _Tp& __a _CCCL_LIFETIMEBOUND, const _Tp& __b _CCCL_LIFETIMEBOUND, _Compare __comp)
+{
+  return __comp(__a, __b) ? __b : __a;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr const _Tp& max(const _Tp& __a _CCCL_LIFETIMEBOUND, const _Tp& __b _CCCL_LIFETIMEBOUND)
+{
+  return __a < __b ? __b : __a;
+}
+
+template <class _Tp, class _Compare>
+[[nodiscard]] _CCCL_API constexpr _Tp max(initializer_list<_Tp> __t, _Compare __comp)
+{
+  return *::cuda::std::__max_element<__comp_ref_type<_Compare>>(__t.begin(), __t.end(), __comp);
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp max(initializer_list<_Tp> __t)
+{
+  return *::cuda::std::max_element(__t.begin(), __t.end(), __less{});
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_MAX_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max_element.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max_element.h
new file mode 100644
index 00000000..bba33f1b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/max_element.h
@@ -0,0 +1,67 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_MAX_ELEMENT_H
+#define _CUDA_STD___ALGORITHM_MAX_ELEMENT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__algorithm/comp_ref_type.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Compare, class _ForwardIterator>
+_CCCL_API constexpr _ForwardIterator __max_element(_ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
+{
+  static_assert(__has_forward_traversal<_ForwardIterator>, "::cuda::std::max_element requires a ForwardIterator");
+  if (__first != __last)
+  {
+    _ForwardIterator __i = __first;
+    while (++__i != __last)
+    {
+      if (__comp(*__first, *__i))
+      {
+        __first = __i;
+      }
+    }
+  }
+  return __first;
+}
+
+template <class _ForwardIterator, class _Compare>
+[[nodiscard]] _CCCL_API constexpr _ForwardIterator
+max_element(_ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
+{
+  return ::cuda::std::__max_element<__comp_ref_type<_Compare>>(__first, __last, __comp);
+}
+
+template <class _ForwardIterator>
+[[nodiscard]] _CCCL_API constexpr _ForwardIterator max_element(_ForwardIterator __first, _ForwardIterator __last)
+{
+  return ::cuda::std::max_element(__first, __last, __less{});
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_MAX_ELEMENT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min.h
new file mode 100644
index 00000000..5335d7b8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min.h
@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_MIN_H
+#define _CUDA_STD___ALGORITHM_MIN_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__algorithm/comp_ref_type.h>
+#include <cuda/std/__algorithm/min_element.h>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp, class _Compare>
+[[nodiscard]] _CCCL_API constexpr const _Tp&
+min(const _Tp& __a _CCCL_LIFETIMEBOUND, const _Tp& __b _CCCL_LIFETIMEBOUND, _Compare __comp)
+{
+  return __comp(__b, __a) ? __b : __a;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr const _Tp& min(const _Tp& __a _CCCL_LIFETIMEBOUND, const _Tp& __b _CCCL_LIFETIMEBOUND)
+{
+  return __b < __a ? __b : __a;
+}
+
+template <class _Tp, class _Compare>
+[[nodiscard]] _CCCL_API constexpr _Tp min(initializer_list<_Tp> __t, _Compare __comp)
+{
+  return *::cuda::std::__min_element<__comp_ref_type<_Compare>>(__t.begin(), __t.end(), __comp);
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp min(initializer_list<_Tp> __t)
+{
+  return *::cuda::std::min_element(__t.begin(), __t.end(), __less{});
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_MIN_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min_element.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min_element.h
new file mode 100644
index 00000000..99185fed
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/min_element.h
@@ -0,0 +1,87 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_MIN_ELEMENT_H
+#define _CUDA_STD___ALGORITHM_MIN_ELEMENT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/comp.h>
+#include <cuda/std/__algorithm/comp_ref_type.h>
+#include <cuda/std/__functional/identity.h>
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/is_callable.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Comp, class _Iter, class _Sent, class _Proj>
+_CCCL_API constexpr _Iter __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj)
+{
+  if (__first == __last)
+  {
+    return __first;
+  }
+
+  _Iter __i = __first;
+  while (++__i != __last)
+  {
+    if (::cuda::std::invoke(__comp, ::cuda::std::invoke(__proj, *__i), ::cuda::std::invoke(__proj, *__first)))
+    {
+      __first = __i;
+    }
+  }
+
+  return __first;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Comp, class _Iter, class _Sent>
+_CCCL_API constexpr _Iter __min_element(_Iter __first, _Sent __last, _Comp __comp)
+{
+  auto __proj = identity();
+  return ::cuda::std::__min_element<_Comp>(::cuda::std::move(__first), ::cuda::std::move(__last), __comp, __proj);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _ForwardIterator, class _Compare>
+[[nodiscard]] _CCCL_API constexpr _ForwardIterator
+min_element(_ForwardIterator __first, _ForwardIterator __last, _Compare __comp)
+{
+  static_assert(__has_forward_traversal<_ForwardIterator>, "std::min_element requires a ForwardIterator");
+  static_assert(__is_callable<_Compare, decltype(*__first), decltype(*__first)>::value,
+                "The comparator has to be callable");
+
+  return ::cuda::std::__min_element<__comp_ref_type<_Compare>>(
+    ::cuda::std::move(__first), ::cuda::std::move(__last), __comp);
+}
+
+template <class _ForwardIterator>
+[[nodiscard]] _CCCL_API constexpr _ForwardIterator min_element(_ForwardIterator __first, _ForwardIterator __last)
+{
+  return ::cuda::std::min_element(__first, __last, __less{});
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_MIN_ELEMENT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/ranges_iterator_concept.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/ranges_iterator_concept.h
new file mode 100644
index 00000000..07d7172d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/ranges_iterator_concept.h
@@ -0,0 +1,65 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_RANGES_ITERATOR_CONCEPT_H
+#define _CUDA_STD___ALGORITHM_RANGES_ITERATOR_CONCEPT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+template <class _IterMaybeQualified>
+_CCCL_API constexpr auto __get_iterator_concept()
+{
+  using _Iter = remove_cvref_t<_IterMaybeQualified>;
+
+  if constexpr (contiguous_iterator<_Iter>)
+  {
+    return contiguous_iterator_tag();
+  }
+  else if constexpr (random_access_iterator<_Iter>)
+  {
+    return random_access_iterator_tag();
+  }
+  else if constexpr (bidirectional_iterator<_Iter>)
+  {
+    return bidirectional_iterator_tag();
+  }
+  else if constexpr (forward_iterator<_Iter>)
+  {
+    return forward_iterator_tag();
+  }
+  else if constexpr (input_iterator<_Iter>)
+  {
+    return input_iterator_tag();
+  }
+}
+
+template <class _Iter>
+using __iterator_concept = decltype(__get_iterator_concept<_Iter>());
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_RANGES_ITERATOR_CONCEPT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/swap_ranges.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/swap_ranges.h
new file mode 100644
index 00000000..ae6e6611
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/swap_ranges.h
@@ -0,0 +1,78 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_SWAP_RANGES_H
+#define _CUDA_STD___ALGORITHM_SWAP_RANGES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/iterator_operations.h>
+#include <cuda/std/__utility/move.h>
+#include <cuda/std/__utility/pair.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// 2+2 iterators: the shorter size will be used.
+_CCCL_EXEC_CHECK_DISABLE
+template <class _AlgPolicy, class _ForwardIterator1, class _Sentinel1, class _ForwardIterator2, class _Sentinel2>
+_CCCL_API constexpr pair<_ForwardIterator1, _ForwardIterator2>
+__swap_ranges(_ForwardIterator1 __first1, _Sentinel1 __last1, _ForwardIterator2 __first2, _Sentinel2 __last2)
+{
+  while (__first1 != __last1 && __first2 != __last2)
+  {
+    _IterOps<_AlgPolicy>::iter_swap(__first1, __first2);
+    ++__first1;
+    ++__first2;
+  }
+
+  return pair<_ForwardIterator1, _ForwardIterator2>(::cuda::std::move(__first1), ::cuda::std::move(__first2));
+}
+
+// 2+1 iterators: size2 >= size1.
+_CCCL_EXEC_CHECK_DISABLE
+template <class _AlgPolicy, class _ForwardIterator1, class _Sentinel1, class _ForwardIterator2>
+_CCCL_API constexpr pair<_ForwardIterator1, _ForwardIterator2>
+__swap_ranges(_ForwardIterator1 __first1, _Sentinel1 __last1, _ForwardIterator2 __first2)
+{
+  while (__first1 != __last1)
+  {
+    _IterOps<_AlgPolicy>::iter_swap(__first1, __first2);
+    ++__first1;
+    ++__first2;
+  }
+
+  return pair<_ForwardIterator1, _ForwardIterator2>(::cuda::std::move(__first1), ::cuda::std::move(__first2));
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _ForwardIterator1, class _ForwardIterator2>
+_CCCL_API constexpr _ForwardIterator2
+swap_ranges(_ForwardIterator1 __first1, _ForwardIterator1 __last1, _ForwardIterator2 __first2)
+{
+  return ::cuda::std::__swap_ranges<_ClassicAlgPolicy>(
+           ::cuda::std::move(__first1), ::cuda::std::move(__last1), ::cuda::std::move(__first2))
+    .second;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_SWAP_RANGES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/unwrap_iter.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/unwrap_iter.h
new file mode 100644
index 00000000..b0bc5313
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__algorithm/unwrap_iter.h
@@ -0,0 +1,95 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ALGORITHM_UNWRAP_ITER_H
+#define _CUDA_STD___ALGORITHM_UNWRAP_ITER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__memory/pointer_traits.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_copy_constructible.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// TODO: Change the name of __unwrap_iter_impl to something more appropriate
+// The job of __unwrap_iter is to remove iterator wrappers (like reverse_iterator or __wrap_iter),
+// to reduce the number of template instantiations and to enable pointer-based optimizations e.g. in ::cuda::std::copy.
+// In debug mode, we don't do this.
+//
+// Some algorithms (e.g. ::cuda::std::copy, but not ::cuda::std::sort) need to convert an
+// "unwrapped" result back into the original iterator type. Doing that is the job of __rewrap_iter.
+
+// Default case - we can't unwrap anything
+template <class _Iter, bool = __has_contiguous_traversal<_Iter>>
+struct __unwrap_iter_impl
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  static _CCCL_API constexpr _Iter __rewrap(_Iter, _Iter __iter)
+  {
+    return __iter;
+  }
+  _CCCL_EXEC_CHECK_DISABLE
+  static _CCCL_API constexpr _Iter __unwrap(_Iter __i) noexcept
+  {
+    return __i;
+  }
+};
+
+// It's a contiguous iterator, so we can use a raw pointer instead
+template <class _Iter>
+struct __unwrap_iter_impl<_Iter, true>
+{
+  using _ToAddressT = decltype(::cuda::std::__to_address(::cuda::std::declval<_Iter>()));
+
+  _CCCL_EXEC_CHECK_DISABLE
+  static _CCCL_API constexpr _Iter __rewrap(_Iter __orig_iter, _ToAddressT __unwrapped_iter)
+  {
+    return __orig_iter + (__unwrapped_iter - ::cuda::std::__to_address(__orig_iter));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  static _CCCL_API constexpr _ToAddressT __unwrap(_Iter __i) noexcept
+  {
+    return ::cuda::std::__to_address(__i);
+  }
+};
+
+template <class _Iter, class _Impl = __unwrap_iter_impl<_Iter>, enable_if_t<is_copy_constructible_v<_Iter>, int> = 0>
+_CCCL_API constexpr decltype(_Impl::__unwrap(::cuda::std::declval<_Iter>())) __unwrap_iter(_Iter __i) noexcept
+{
+  return _Impl::__unwrap(__i);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _OrigIter, class _Iter, class _Impl = __unwrap_iter_impl<_OrigIter>>
+_CCCL_API constexpr _OrigIter __rewrap_iter(_OrigIter __orig_iter, _Iter __iter) noexcept
+{
+  return _Impl::__rewrap(::cuda::std::move(__orig_iter), ::cuda::std::move(__iter));
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ALGORITHM_UNWRAP_ITER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__bit/bit_cast.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__bit/bit_cast.h
new file mode 100644
index 00000000..7cd58887
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__bit/bit_cast.h
@@ -0,0 +1,86 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024-26 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___BIT_BIT_CAST_H
+#define _CUDA_STD___BIT_BIT_CAST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__type_traits/is_trivially_copyable.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__cstring/memcpy.h>
+#include <cuda/std/__type_traits/is_default_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+// MSVC supports __builtin_bit_cast from 19.25 on
+#if _CCCL_CHECK_BUILTIN(builtin_bit_cast) || _CCCL_COMPILER(MSVC, >, 19, 25)
+#  define _CCCL_BUILTIN_BIT_CAST(...) __builtin_bit_cast(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_bit_cast)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_BIT_CAST)
+#  define _CCCL_CONSTEXPR_BIT_CAST       constexpr
+#  define _CCCL_HAS_CONSTEXPR_BIT_CAST() 1
+#else // ^^^ _CCCL_BUILTIN_BIT_CAST ^^^ / vvv !_CCCL_BUILTIN_BIT_CAST vvv
+#  define _CCCL_CONSTEXPR_BIT_CAST
+#  define _CCCL_HAS_CONSTEXPR_BIT_CAST() 0
+#endif // !_CCCL_BUILTIN_BIT_CAST
+
+#if _CCCL_COMPILER(GCC, >=, 8)
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_GCC("-Wclass-memaccess")
+#endif // _CCCL_COMPILER(GCC, >=, 8)
+
+template <class _To, class _From>
+[[nodiscard]] _CCCL_API inline _To __bit_cast_memcpy(const _From& __from) noexcept
+{
+  static_assert(::cuda::std::is_default_constructible_v<_To>,
+                "bit_cast memcpy fallback requires the destination type to be default constructible");
+  _To __temp;
+  ::cuda::std::memcpy(&__temp, &__from, sizeof(_To));
+  return __temp;
+}
+
+#if _CCCL_COMPILER(GCC, >=, 8)
+_CCCL_DIAG_POP
+#endif // _CCCL_COMPILER(GCC, >=, 8)
+
+_CCCL_TEMPLATE(class _To, class _From)
+_CCCL_REQUIRES((sizeof(_To) == sizeof(_From)) _CCCL_AND(::cuda::is_trivially_copyable_v<_To>)
+                 _CCCL_AND(::cuda::is_trivially_copyable_v<_From>))
+[[nodiscard]] _CCCL_API inline _CCCL_CONSTEXPR_BIT_CAST _To bit_cast(const _From& __from) noexcept
+{
+#if defined(_CCCL_BUILTIN_BIT_CAST)
+  if constexpr (::cuda::std::is_trivially_copyable_v<_To> && ::cuda::std::is_trivially_copyable_v<_From>)
+  {
+    return _CCCL_BUILTIN_BIT_CAST(_To, __from);
+  }
+  else
+#endif // _CCCL_BUILTIN_BIT_CAST
+  {
+    return ::cuda::std::__bit_cast_memcpy<_To>(__from);
+  }
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___BIT_BIT_CAST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/architecture.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/architecture.h
new file mode 100644
index 00000000..807b6b63
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/architecture.h
@@ -0,0 +1,128 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_ARCH_H
+#define __CCCL_ARCH_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/preprocessor.h>
+
+// The header provides the following macros to determine the host architecture:
+//
+// _CCCL_HOST_ARCH(ARM64)  ARM64
+// _CCCL_HOST_ARCH(X86_64) X86 64 bit
+// CCCL_HOST_ARCH(ARM64)   ARM64
+// CCCL_HOST_ARCH(X86_64)  X86 64 bit
+
+// Determine the host architecture
+
+// Arm 64-bit
+#if (defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC) /*emulation*/)
+#  define _CCCL_HOST_ARCH_ARM64_() 1
+#else
+#  define _CCCL_HOST_ARCH_ARM64_() 0
+#endif
+
+// X86 64-bit
+
+// _M_X64 is defined even if we are compiling in Arm64 emulation mode
+#if (defined(_M_X64) && !defined(_M_ARM64EC)) || defined(__amd64__) || defined(__x86_64__)
+#  define _CCCL_HOST_ARCH_X86_64_() 1
+#else
+#  define _CCCL_HOST_ARCH_X86_64_() 0
+#endif
+
+#define _CCCL_HOST_ARCH(...) _CCCL_HOST_ARCH_##__VA_ARGS__##_()
+
+//! @def CCCL_HOST_ARCH(ARCH) /* implementation defined */
+//!
+//! @brief Detect the current host architecture.
+//!
+//! @param ARCH The name of the host architecture to test.
+//!
+//! @note This macro is made available when including any libcu++ header. Users that wish to
+//! include the smallest possible header for this macro should include `<cuda/std/version>`.
+//!
+//! For supported host architectures, the macro expands to an implementation-defined true value
+//! if the current host architecture matches, or false otherwise. These values may be used in
+//! boolean expressions (preprocessor or otherwise), but no other guarantees are made.
+//!
+//! Available values for `ARCH` include:
+//!
+//! - ``ARM64``: ARM 64-bit, including MSVC ARM64EC emulation.
+//! - ``X86_64``: X86 64-bit. This is false when compiling in MSVC ARM64EC emulation mode.
+//!
+//! Passing any other value will result in an undefined expansion, which may or may not be
+//! diagnosed by the compiler.
+//!
+//! @par Example
+//! @code
+//! #define MY_OTHER_MACRO 1
+//!
+//! // Expansion value can be used in ordinary macro conditionals
+//! #if CCCL_HOST_ARCH(X86_64) && MY_OTHER_MACRO
+//!   // ...
+//! #endif
+//!
+//! // Can be negated as usual
+//! #if !CCCL_HOST_ARCH(ARM64)
+//!   // ...
+//! #endif
+//! @endcode
+//!
+//! @return true if the specified host architecture is being compiled for, false otherwise.
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CCCL_HOST_ARCH(ARCH) /* implementation defined */
+#else
+#  define CCCL_HOST_ARCH(__arch__) _CCCL_HOST_ARCH_##__arch__##_()
+#endif
+
+// Note: the public API is single-arg to constrain the API and allow for future expansion. The
+// implementation is duplicated to guard against the architecture targets being accidentally
+// defined by the user.
+
+// Determine the endianness
+
+#define _CCCL_ENDIAN_LITTLE() 0xDEAD
+#define _CCCL_ENDIAN_BIG()    0xFACE
+#define _CCCL_ENDIAN_PDP()    0xBEEF
+
+#if _CCCL_COMPILER(NVRTC) || (_CCCL_COMPILER(MSVC) && (_CCCL_HOST_ARCH(X86_64) || _CCCL_HOST_ARCH(ARM64))) \
+  || __LITTLE_ENDIAN__
+#  define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_LITTLE()
+#elif __BIG_ENDIAN__
+#  define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_BIG()
+#elif defined(__BYTE_ORDER__)
+#  if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_LITTLE()
+#  elif __BYTE_ORDER__ == __ORDER_PDP_ENDIAN__
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_PDP()
+#  elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_BIG()
+#  endif // __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
+#elif __has_include(<endian.h>)
+#  include <endian.h>
+#  if __BYTE_ORDER == __LITTLE_ENDIAN
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_LITTLE()
+#  elif __BYTE_ORDER == __PDP_ENDIAN
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_PDP()
+#  elif __BYTE_ORDER == __BIG_ENDIAN
+#    define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_BIG()
+#  endif // __BYTE_ORDER == __BIG_ENDIAN
+#endif // ^^^ has endian.h ^^^
+
+#if !defined(_CCCL_ENDIAN_NATIVE)
+_CCCL_WARNING("failed to determine the endianness of the host architecture, defaulting to little-endian")
+#  define _CCCL_ENDIAN_NATIVE() _CCCL_ENDIAN_LITTLE()
+#endif // !_CCCL_ENDIAN_NATIVE
+
+#define _CCCL_ENDIAN(_NAME) (_CCCL_ENDIAN_NATIVE() == _CCCL_ENDIAN_##_NAME())
+
+#endif // __CCCL_ARCH_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/assert.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/assert.h
new file mode 100644
index 00000000..8fd3dc9b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/assert.h
@@ -0,0 +1,169 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_ASSERT_H
+#define __CCCL_ASSERT_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/attributes.h>
+#include <cuda/std/__cccl/builtin.h>
+#include <cuda/std/__cccl/execution_space.h>
+#include <cuda/std/__cccl/preprocessor.h>
+
+#if _CCCL_HOSTED()
+#  include <assert.h>
+#endif // _CCCL_HOSTED()
+
+#include <nv/target>
+
+#if defined(_DEBUG) || defined(DEBUG)
+#  ifndef _CCCL_ENABLE_DEBUG_MODE
+#    define _CCCL_ENABLE_DEBUG_MODE
+#  endif // !_CCCL_ENABLE_DEBUG_MODE
+#endif // _DEBUG || DEBUG
+
+// Automatically enable assertions when debug mode is enabled
+#ifdef _CCCL_ENABLE_DEBUG_MODE
+#  ifndef CCCL_ENABLE_ASSERTIONS
+#    define CCCL_ENABLE_ASSERTIONS
+#  endif // !CCCL_ENABLE_ASSERTIONS
+#endif // _CCCL_ENABLE_DEBUG_MODE
+
+//! Ensure that we switch on host assertions when all assertions are enabled
+#ifndef CCCL_ENABLE_HOST_ASSERTIONS
+#  ifdef CCCL_ENABLE_ASSERTIONS
+#    define CCCL_ENABLE_HOST_ASSERTIONS
+#  endif // CCCL_ENABLE_ASSERTIONS
+#endif // !CCCL_ENABLE_HOST_ASSERTIONS
+
+//! Ensure that we switch on device assertions when all assertions are enabled
+#ifndef CCCL_ENABLE_DEVICE_ASSERTIONS
+#  if defined(CCCL_ENABLE_ASSERTIONS) || defined(__CUDACC_DEBUG__)
+#    define CCCL_ENABLE_DEVICE_ASSERTIONS
+#  endif // CCCL_ENABLE_ASSERTIONS
+#endif // !CCCL_ENABLE_DEVICE_ASSERTIONS
+
+//! Use the different standard library implementations to implement host side asserts
+//! _CCCL_ASSERT_IMPL_HOST should never be used directly
+#if _CCCL_OS(QNX)
+#  define _CCCL_ASSERT_IMPL_HOST(expression, message) ((void) 0)
+#elif _CCCL_COMPILER(NVRTC) // There is no host standard library in nvrtc
+#  define _CCCL_ASSERT_IMPL_HOST(expression, message) ((void) 0)
+#elif __has_include(<yvals.h>) && _CCCL_OS(WINDOWS) // Windows uses _STL_VERIFY from <yvals.h>
+#  include <yvals.h>
+#  define _CCCL_ASSERT_IMPL_HOST(expression, message) _STL_VERIFY(expression, message)
+#else // ^^^ MSVC STL ^^^ / vvv !MSVC STL vvv
+#  ifdef NDEBUG
+// Reintroduce the __assert_fail / __assert_rtn declaration
+extern "C" {
+#    if !_CCCL_CUDA_COMPILER(CLANG)
+_CCCL_HOST_DEVICE
+#    endif // !_CCCL_CUDA_COMPILER(CLANG)
+#    if _CCCL_OS(APPLE)
+void __assert_rtn(const char* __function, const char* __assertion, const char* __file, unsigned int __line) noexcept
+  __attribute__((__noreturn__));
+#    else // ^^^ _CCCL_OS(APPLE) ^^^ / vvv !_CCCL_OS(APPLE) ^^^
+void __assert_fail(const char* __assertion, const char* __file, unsigned int __line, const char* __function) noexcept
+  __attribute__((__noreturn__));
+#    endif // !_CCCL_OS(APPLE)
+}
+#  endif // NDEBUG
+
+#  if _CCCL_OS(APPLE)
+#    define _CCCL_ASSERT_IMPL_HOST(expression, message)      \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : __assert_rtn(__func__, __FILE__, __LINE__, message)
+#  elif _CCCL_OS(ANDROID)
+#    define _CCCL_ASSERT_IMPL_HOST(expression, message)      \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : __assert2(__FILE__, __LINE__, __func__, message)
+#  else // ^^^ _CCCL_OS(APPLE) ^^^ / vvv !_CCCL_OS(APPLE) ^^^
+#    define _CCCL_ASSERT_IMPL_HOST(expression, message)      \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : __assert_fail(message, __FILE__, __LINE__, __func__)
+#  endif // !_CCCL_OS(APPLE)
+#endif // !MSVC STL
+
+//! Use custom implementations with nvcc on device and the host ones with clang-cuda and nvhpc
+//! _CCCL_ASSERT_IMPL_DEVICE should never be used directly
+#if _CCCL_OS(QNX) || _CCCL_OS(APPLE)
+#  define _CCCL_ASSERT_IMPL_DEVICE(expression, message) ((void) 0)
+#elif _CCCL_COMPILER(NVRTC)
+#  define _CCCL_ASSERT_IMPL_DEVICE(expression, message)    \
+    _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+    ? (void) 0 : __assertfail(message, __FILE__, __LINE__, __func__, sizeof(char))
+#elif _CCCL_CUDA_COMPILER(NVCC) //! Use __assert_fail to implement device side asserts
+#  if _CCCL_COMPILER(MSVC)
+#    define _CCCL_ASSERT_IMPL_DEVICE(expression, message)    \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : _wassert(_CRT_WIDE(#message), __FILEW__, __LINE__)
+#  elif _CCCL_OS(ANDROID)
+#    define _CCCL_ASSERT_IMPL_DEVICE(expression, message)    \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : __assert2(__FILE__, __LINE__, __func__, message)
+#  else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#    define _CCCL_ASSERT_IMPL_DEVICE(expression, message)    \
+      _CCCL_BUILTIN_EXPECT(static_cast<bool>(expression), 1) \
+      ? (void) 0 : __assert_fail(message, __FILE__, __LINE__, __func__)
+#  endif // !_CCCL_COMPILER(MSVC)
+#elif _CCCL_CUDA_COMPILATION()
+#  define _CCCL_ASSERT_IMPL_DEVICE(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
+#else // ^^^ _CCCL_CUDA_COMPILATION() ^^^ / vvv !_CCCL_CUDA_COMPILATION() vvv
+#  define _CCCL_ASSERT_IMPL_DEVICE(expression, message) ((void) 0)
+#endif // !_CCCL_CUDA_COMPILATION()
+
+//! _CCCL_ASSERT_HOST is enabled conditionally depending on CCCL_ENABLE_HOST_ASSERTIONS
+#ifdef CCCL_ENABLE_HOST_ASSERTIONS
+#  define _CCCL_ASSERT_HOST(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
+#else // ^^^ CCCL_ENABLE_HOST_ASSERTIONS ^^^ / vvv !CCCL_ENABLE_HOST_ASSERTIONS vvv
+#  define _CCCL_ASSERT_HOST(expression, message) ((void) 0)
+#endif // !CCCL_ENABLE_HOST_ASSERTIONS
+
+//! _CCCL_ASSERT_DEVICE is enabled conditionally depending on CCCL_ENABLE_DEVICE_ASSERTIONS
+#ifdef CCCL_ENABLE_DEVICE_ASSERTIONS
+#  define _CCCL_ASSERT_DEVICE(expression, message) _CCCL_ASSERT_IMPL_DEVICE(expression, message)
+#else // ^^^ CCCL_ENABLE_DEVICE_ASSERTIONS ^^^ / vvv !CCCL_ENABLE_DEVICE_ASSERTIONS vvv
+#  define _CCCL_ASSERT_DEVICE(expression, message) ((void) 0)
+#endif // !CCCL_ENABLE_DEVICE_ASSERTIONS
+
+//! _CCCL_VERIFY is enabled unconditionally and reserved for critical checks that are required to always be on
+//! _CCCL_ASSERT is enabled conditionally depending on CCCL_ENABLE_HOST_ASSERTIONS and CCCL_ENABLE_DEVICE_ASSERTIONS
+#if _CCCL_CUDA_COMPILER(NVHPC) // NVHPC can't have different behavior for host and device.
+                               // The host version of the assert will also work in device code.
+#  define _CCCL_VERIFY(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
+#  if defined(CCCL_ENABLE_HOST_ASSERTIONS) || defined(CCCL_ENABLE_DEVICE_ASSERTIONS)
+#    define _CCCL_ASSERT(expression, message) _CCCL_ASSERT_HOST(expression, message)
+#  else
+#    define _CCCL_ASSERT(expression, message) ((void) 0)
+#  endif
+#elif _CCCL_CUDA_COMPILATION()
+#  if _CCCL_DEVICE_COMPILATION()
+#    define _CCCL_VERIFY(expression, message) _CCCL_ASSERT_IMPL_DEVICE(expression, message)
+#    define _CCCL_ASSERT(expression, message) _CCCL_ASSERT_DEVICE(expression, message)
+#  else // ^^^ _CCCL_DEVICE_COMPILATION() ^^^ / vvv !_CCCL_DEVICE_COMPILATION() vvv
+#    define _CCCL_VERIFY(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
+#    define _CCCL_ASSERT(expression, message) _CCCL_ASSERT_HOST(expression, message)
+#  endif // !_CCCL_DEVICE_COMPILATION()
+#else // ^^^ _CCCL_CUDA_COMPILATION() ^^^ / vvv !_CCCL_CUDA_COMPILATION() vvv
+#  define _CCCL_VERIFY(expression, message) _CCCL_ASSERT_IMPL_HOST(expression, message)
+#  define _CCCL_ASSERT(expression, message) _CCCL_ASSERT_HOST(expression, message)
+#endif // !_CCCL_CUDA_COMPILATION()
+
+#endif // __CCCL_ASSERT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/attributes.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/attributes.h
new file mode 100644
index 00000000..bc8bb34d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/attributes.h
@@ -0,0 +1,221 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_ATTRIBUTES_H
+#define __CCCL_ATTRIBUTES_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/diagnostic.h>
+#include <cuda/std/__cccl/dialect.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#ifdef __has_attribute
+#  define _CCCL_HAS_ATTRIBUTE(__x) __has_attribute(__x)
+#else // ^^^ __has_attribute ^^^ / vvv !__has_attribute vvv
+#  define _CCCL_HAS_ATTRIBUTE(__x) 0
+#endif // !__has_attribute
+
+#ifdef __has_cpp_attribute
+#  define _CCCL_HAS_CPP_ATTRIBUTE(__x) __has_cpp_attribute(__x)
+#else // ^^^ __has_cpp_attribute ^^^ / vvv !__has_cpp_attribute vvv
+#  define _CCCL_HAS_CPP_ATTRIBUTE(__x) 0
+#endif // !__has_cpp_attribute
+
+#ifdef __has_declspec_attribute
+#  define _CCCL_HAS_DECLSPEC_ATTRIBUTE(__x) __has_declspec_attribute(__x)
+#else // ^^^ __has_declspec_attribute ^^^ / vvv !__has_declspec_attribute vvv
+#  define _CCCL_HAS_DECLSPEC_ATTRIBUTE(__x) 0
+#endif // !__has_declspec_attribute
+
+// MSVC needs extra help with empty base classes
+#if _CCCL_COMPILER(MSVC) || _CCCL_HAS_DECLSPEC_ATTRIBUTE(empty_bases)
+#  define _CCCL_DECLSPEC_EMPTY_BASES __declspec(empty_bases)
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_DECLSPEC_EMPTY_BASES
+#endif // !_CCCL_COMPILER(MSVC)
+
+#if _CCCL_HAS_ATTRIBUTE(__nodebug__)
+#  define _CCCL_NODEBUG __attribute__((__nodebug__))
+#else // ^^^ _CCCL_HAS_ATTRIBUTE(__nodebug__) ^^^ / vvv !_CCCL_HAS_ATTRIBUTE(__nodebug__) vvv
+#  define _CCCL_NODEBUG
+#endif // !_CCCL_HAS_ATTRIBUTE(__nodebug__)
+
+// Debuggers do not step into functions marked with __attribute__((__artificial__)). This
+// is useful for small wrapper functions that just dispatch to other functions and that
+// are inlined into the caller.
+#if _CCCL_HAS_ATTRIBUTE(__artificial__) && !_CCCL_CUDA_COMPILER(NVCC)
+#  define _CCCL_ARTIFICIAL __attribute__((__artificial__))
+#else // ^^^ _CCCL_HAS_ATTRIBUTE(__artificial__) ^^^ / vvv !_CCCL_HAS_ATTRIBUTE(__artificial__) vvv
+#  define _CCCL_ARTIFICIAL
+#endif // !_CCCL_HAS_ATTRIBUTE(__artificial__)
+
+// The nodebug attribute flattens aliases down to the actual type rather typename meow<T>::type
+#if _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_NODEBUG_ALIAS _CCCL_NODEBUG
+#else // ^^^ _CCCL_CUDA_COMPILER(CLANG) ^^^ / vvv !_CCCL_CUDA_COMPILER(CLANG) vvv
+#  define _CCCL_NODEBUG_ALIAS
+#endif // !_CCCL_CUDA_COMPILER(CLANG)
+
+// _CCCL_ASSUME
+// NVCC does not properly respect [[assume()]], so use __builtin_assume, see nvbug5458663
+#if _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+#  define _CCCL_ASSUME(...) __builtin_assume(__VA_ARGS__)
+#elif _CCCL_HAS_CPP_ATTRIBUTE(assume)
+#  define _CCCL_ASSUME(...) [[assume(__VA_ARGS__)]]
+#else
+#  define _CCCL_ASSUME(...) _CCCL_BUILTIN_ASSUME(__VA_ARGS__)
+#endif
+
+#if _CCCL_TILE_COMPILATION() // nvbug6100910: __builtin_assume is not supported in tile mode
+#  undef _CCCL_ASSUME
+#  define _CCCL_ASSUME(...)
+#endif // _CCCL_TILE_COMPILATION()
+
+// _CCCL_CONST
+
+#if _CCCL_HAS_CPP_ATTRIBUTE(__gnu__::__const__)
+#  define _CCCL_CONST [[__gnu__::__const__]]
+#else // ^^^ has gnu::const ^^^ / vvv no gnu::const vvv
+#  define _CCCL_CONST _CCCL_PURE
+#endif // ^^^ no gnu::const ^^^
+
+// _CCCL_DIAGNOSE_IF
+
+#if _CCCL_HAS_ATTRIBUTE(__diagnose_if__)
+#  define _CCCL_DIAGNOSE_IF(_COND, _MSG, _TYPE) __attribute__((__diagnose_if__(_COND, _MSG, _TYPE)))
+#else // ^^^ _CCCL_HAS_ATTRIBUTE(diagnose_if) ^^^ / vvv !_CCCL_HAS_ATTRIBUTE(diagnose_if) vvv
+#  define _CCCL_DIAGNOSE_IF(_COND, _MSG, _TYPE)
+#endif // !_CCCL_HAS_ATTRIBUTE(diagnose_if)
+
+// _CCCL_INTRINSIC
+
+// MSVC provides a way to mark functions as intrinsic provided the function's body consists of a single
+// return statement of a cast expression (e.g., move(x) or forward<T>(u)).
+#if _CCCL_COMPILER(MSVC) && _CCCL_HAS_CPP_ATTRIBUTE(msvc::intrinsic)
+#  define _CCCL_INTRINSIC [[msvc::intrinsic]]
+#else
+#  define _CCCL_INTRINSIC
+#endif
+
+// _CCCL_PURE
+
+#if _CCCL_CUDA_COMPILER(NVCC, >=, 12, 5)
+#  define _CCCL_PURE __nv_pure__
+#elif _CCCL_HAS_CPP_ATTRIBUTE(__gnu__::__pure__)
+#  define _CCCL_PURE [[__gnu__::__pure__]]
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_PURE __declspec(noalias)
+#else
+#  define _CCCL_PURE
+#endif
+
+// _CCCL_NO_CFI
+
+#if !_CCCL_COMPILER(GCC)
+#  define _CCCL_NO_CFI _CCCL_NO_SANITIZE("cfi")
+#else
+#  define _CCCL_NO_CFI
+#endif
+
+// _CCCL_NO_SANITIZE
+
+#if _CCCL_HAS_ATTRIBUTE(__no_sanitize__)
+#  define _CCCL_NO_SANITIZE(_STR) __attribute__((__no_sanitize__(_STR)))
+#else // ^^^ _CCCL_HAS_ATTRIBUTE(no_sanitize) ^^^ / vvv !_CCCL_HAS_ATTRIBUTE(no_sanitize) vvv
+#  define _CCCL_NO_SANITIZE(_STR)
+#endif // !_CCCL_HAS_ATTRIBUTE(no_sanitize)
+
+// _CCCL_NO_SPECIALIZATIONS
+
+#if _CCCL_HAS_CPP_ATTRIBUTE(clang::__no_specializations__)
+#  define _CCCL_NO_SPECIALIZATIONS_BECAUSE(_MSG)   [[clang::__no_specializations__(_MSG)]]
+#  define _CCCL_HAS_ATTRIBUTE_NO_SPECIALIZATIONS() 1
+#elif _CCCL_HAS_CPP_ATTRIBUTE(msvc::no_specializations)
+#  define _CCCL_NO_SPECIALIZATIONS_BECAUSE(_MSG)   [[msvc::no_specializations(_MSG)]]
+#  define _CCCL_HAS_ATTRIBUTE_NO_SPECIALIZATIONS() 1
+#else // ^^^ has attribute no_specializations ^^^ / vvv hasn't attribute no_specializations vvv
+#  define _CCCL_NO_SPECIALIZATIONS_BECAUSE(_MSG)
+#  define _CCCL_HAS_ATTRIBUTE_NO_SPECIALIZATIONS() 0
+#endif // ^^^ hasn't attribute no_specializations ^^^
+
+#define _CCCL_NO_SPECIALIZATIONS \
+  _CCCL_NO_SPECIALIZATIONS_BECAUSE("Users are not allowed to specialize this cccl entity")
+
+// _CCCL_LIFETIMEBOUND
+
+#if _CCCL_HAS_CPP_ATTRIBUTE(clang::lifetimebound) || _CCCL_COMPILER(CLANG)
+#  define _CCCL_LIFETIMEBOUND [[clang::lifetimebound]]
+#elif _CCCL_HAS_CPP_ATTRIBUTE(msvc::lifetimebound) || _CCCL_COMPILER(MSVC, >=, 19, 37)
+#  define _CCCL_LIFETIMEBOUND [[msvc::lifetimebound]]
+#else
+#  define _CCCL_LIFETIMEBOUND
+#endif
+
+// _CCCL_NO_UNIQUE_ADDRESS
+
+#if _CCCL_COMPILER(MSVC) || _CCCL_HAS_CPP_ATTRIBUTE(no_unique_address) < 201803L
+// MSVC implementation has lead to multiple issues with silent runtime corruption when passing data into kernels
+#  define _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() 0
+#  define _CCCL_NO_UNIQUE_ADDRESS
+#elif _CCCL_HAS_CPP_ATTRIBUTE(no_unique_address)
+#  define _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() 1
+#  define _CCCL_NO_UNIQUE_ADDRESS                 [[no_unique_address]]
+#else
+#  define _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() 0
+#  define _CCCL_NO_UNIQUE_ADDRESS
+#endif
+
+// Passing objects with nested [[no_unique_address]] to kernels leads to data corruption.
+// This is caused by cudafe++ not honoring [[no_unique_address]] when compiling for C++17
+// with clang as the host compiler. See nvbug 5265027 for more details.
+#if _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() && _CCCL_COMPILER(CLANG) && _CCCL_STD_VER < 2020 \
+  && _CCCL_CUDA_COMPILER(NVCC)
+#  undef _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS
+#  undef _CCCL_NO_UNIQUE_ADDRESS
+#  define _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() 0
+#  define _CCCL_NO_UNIQUE_ADDRESS
+#endif // _CCCL_HAS_ATTRIBUTE_NO_UNIQUE_ADDRESS() && _CCCL_COMPILER(CLANG)
+
+// _CCCL_PREFERRED_NAME
+
+#if _CCCL_HAS_ATTRIBUTE(__preferred_name__)
+#  define _CCCL_PREFERRED_NAME(x) __attribute__((__preferred_name__(x)))
+#else
+#  define _CCCL_PREFERRED_NAME(x)
+#endif
+
+#if _CCCL_HAS_ATTRIBUTE(__require_constant_initialization__)
+#  define _CCCL_REQUIRE_CONSTANT_INITIALIZATION __attribute__((__require_constant_initialization__))
+#else
+#  define _CCCL_REQUIRE_CONSTANT_INITIALIZATION
+#endif
+
+// _CCCL_RESTRICT
+
+#if _CCCL_COMPILER(MSVC) // vvv _CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_RESTRICT __restrict
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_RESTRICT __restrict__
+#endif // ^^^ !_CCCL_COMPILER(MSVC) ^^^
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CCCL_ATTRIBUTES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/builtin.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/builtin.h
new file mode 100644
index 00000000..5580b83f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/builtin.h
@@ -0,0 +1,474 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_BUILTIN_H
+#define __CCCL_BUILTIN_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/preprocessor.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/cuda_capabilities.h>
+#include <cuda/std/__cccl/extended_data_types.h>
+#include <cuda/std/__cccl/host_std_lib.h>
+
+//! This file consolidates all compiler builtin detection for CCCL.
+//!
+//! To work around older compilers not supporting `__has_builtin` we use `_CCCL_CHECK_BUILTIN` that detects more
+//! cases
+//!
+//! * We work around old clang versions (before clang-10) not supporting __has_builtin via _CCCL_CHECK_BUILTIN
+//! * We work around old intel versions (before 2021.3)   not supporting __has_builtin via _CCCL_CHECK_BUILTIN
+//! * We work around old nvhpc versions (before 2022.11)  not supporting __has_builtin via _CCCL_CHECK_BUILTIN
+//! * MSVC needs manual handling, has no real way of checking builtins so all is manual
+//! * GCC  needs manual handling, before gcc-10 as that finally supports __has_builtin
+//!
+//! In case compiler support for a builtin is advertised but leads to regressions we explicitly undef the macro
+//!
+//! Finally, because `_CCCL_CHECK_BUILTIN` may lead to false positives, we move detection of new builtins over towards
+//! just using _CCCL_HAS_BUILTIN
+
+#ifdef __has_builtin
+#  define _CCCL_HAS_BUILTIN(__x) __has_builtin(__x)
+#else // ^^^ __has_builtin ^^^ / vvv !__has_builtin vvv
+#  define _CCCL_HAS_BUILTIN(__x) 0
+#endif // !__has_builtin
+
+#ifdef __has_feature
+#  define _CCCL_HAS_FEATURE(__x) __has_feature(__x)
+#else // ^^^ __has_feature ^^^ / vvv !__has_feature vvv
+#  define _CCCL_HAS_FEATURE(__x) 0
+#endif // !__has_feature
+
+// '__is_identifier' returns '0' if '__x' is a reserved identifier provided by the compiler and '1' otherwise.
+#ifdef __is_identifier
+#  define _CCCL_IS_IDENTIFIER(__x) __is_identifier(__x)
+#else // ^^^ __is_identifier ^^^ / vvv !__is_identifier vvv
+#  define _CCCL_IS_IDENTIFIER(__x) 1
+#endif // !__is_identifier
+
+#define _CCCL_HAS_KEYWORD(__x) !(_CCCL_IS_IDENTIFIER(__x))
+
+// https://bugs.llvm.org/show_bug.cgi?id=44517
+#define _CCCL_CHECK_BUILTIN(__x) (_CCCL_HAS_BUILTIN(__##__x) || _CCCL_HAS_KEYWORD(__##__x) || _CCCL_HAS_FEATURE(__x))
+
+// NVCC has issues with function pointers
+#if _CCCL_HAS_BUILTIN(__add_lvalue_reference) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_ADD_LVALUE_REFERENCE(...) __add_lvalue_reference(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__add_lvalue_reference)
+
+// NVCC has issues with function pointers
+#if _CCCL_HAS_BUILTIN(__add_pointer) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_ADD_POINTER(...) __add_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__add_pointer)
+
+// NVCC has issues with function pointers
+#if _CCCL_HAS_BUILTIN(__add_rvalue_reference) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_ADD_RVALUE_REFERENCE(...) __add_rvalue_reference(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__add_rvalue_reference)
+
+// TODO: Enable using the builtin __array_rank when https://llvm.org/PR57133 is resolved
+#if 0 // _CCCL_CHECK_BUILTIN(array_rank)
+#  define _CCCL_BUILTIN_ARRAY_RANK(...) __array_rank(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(array_rank)
+
+// nvhpc has a bug where it supports __builtin_addressof but does not mark it via _CCCL_CHECK_BUILTIN
+#if _CCCL_CHECK_BUILTIN(builtin_addressof) || _CCCL_COMPILER(GCC, >=, 7) || _CCCL_COMPILER(MSVC) \
+  || _CCCL_COMPILER(NVHPC) || _CCCL_COMPILER(NVRTC, >=, 12, 3)
+#  define _CCCL_BUILTIN_ADDRESSOF(...) __builtin_addressof(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_addressof)
+
+#if _CCCL_CHECK_BUILTIN(builtin_assume) || _CCCL_COMPILER(CLANG) || _CCCL_COMPILER(NVHPC)
+#  define _CCCL_BUILTIN_ASSUME(...) __builtin_assume(__VA_ARGS__)
+#elif _CCCL_COMPILER(GCC, >=, 13)
+#  define _CCCL_BUILTIN_ASSUME(...) __attribute__((__assume__(__VA_ARGS__)))
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_BUILTIN_ASSUME(...) __assume(__VA_ARGS__)
+#else
+#  define _CCCL_BUILTIN_ASSUME(...)
+#endif // _CCCL_CHECK_BUILTIN(builtin_assume)
+
+#if _CCCL_TILE_COMPILATION() // nvbug6100910: __builtin_assume is not supported in tile mode
+#  undef _CCCL_BUILTIN_ASSUME
+#  define _CCCL_BUILTIN_ASSUME(...)
+#endif // _CCCL_TILE_COMPILATION()
+
+#if _CCCL_HAS_BUILTIN(__builtin_assume_aligned) || _CCCL_COMPILER(MSVC, >=, 19, 23) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_ASSUME_ALIGNED(...) __builtin_assume_aligned(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__builtin_assume_aligned)
+
+#if _CCCL_CHECK_BUILTIN(builtin_constant_p) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_CONSTANT_P(...) __builtin_constant_p(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_constant_p)
+
+#if _CCCL_CHECK_BUILTIN(builtin_expect) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_EXPECT(_EXPR, _VAL) __builtin_expect(_EXPR, _VAL)
+#else // ^^^ has __builtin_expect ^^^ / vvv no __builtin_expect vvv
+#  define _CCCL_BUILTIN_EXPECT(_EXPR, _VAL) (_EXPR)
+#endif // ^^^ no __builtin_expect ^^^
+
+#if _CCCL_TILE_COMPILATION() // nvbug6100927:  __builtin_expect is unsupported in tile mode
+#  undef _CCCL_BUILTIN_EXPECT
+#  define _CCCL_BUILTIN_EXPECT(_EXPR, _VAL) (_EXPR)
+#endif // _CCCL_TILE_COMPILATION()
+
+#if _CCCL_CHECK_BUILTIN(builtin_huge_valf) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_HUGE_VALF() __builtin_huge_valf()
+#endif // _CCCL_CHECK_BUILTIN(builtin_huge_valf)
+
+#if _CCCL_CHECK_BUILTIN(builtin_huge_val) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_HUGE_VAL() __builtin_huge_val()
+#endif // _CCCL_CHECK_BUILTIN(builtin_huge_val)
+
+#if _CCCL_CHECK_BUILTIN(builtin_huge_vall) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_HUGE_VALL() __builtin_huge_vall()
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_BUILTIN_HUGE_VALL() static_cast<long double>(__builtin_huge_val())
+#endif // _CCCL_CHECK_BUILTIN(builtin_huge_vall)
+
+#if _CCCL_HAS_FLOAT128()
+#  if _CCCL_CHECK_BUILTIN(builtin_huge_valf128) || _CCCL_COMPILER(GCC, >=, 7)
+#    define _CCCL_BUILTIN_HUGE_VALF128() __builtin_huge_valf128()
+#  endif // _CCCL_CHECK_BUILTIN(builtin_huge_valf128) || _CCCL_COMPILER(GCC, >=, 7)
+
+// nvcc does not implement __builtin_huge_valf128
+#  if _CCCL_CUDA_COMPILER(NVCC)
+#    undef _CCCL_BUILTIN_HUGE_VALF128
+#  endif // _CCCL_CUDA_COMPILER(NVCC)
+#endif // _CCCL_HAS_FLOAT128()
+
+#if _CCCL_CHECK_BUILTIN(builtin_is_constant_evaluated) || _CCCL_COMPILER(GCC, >=, 9) || _CCCL_COMPILER(MSVC, >, 19, 24)
+#  define _CCCL_BUILTIN_IS_CONSTANT_EVALUATED(...) __builtin_is_constant_evaluated(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_is_constant_evaluated)
+
+#if _CCCL_TILE_COMPILATION() // nvbug6067464: __builtin_is_constant_evaluated is unsupported in tile mode
+#  undef _CCCL_BUILTIN_IS_CONSTANT_EVALUATED
+#endif // _CCCL_TILE_COMPILATION()
+
+#if _CCCL_CHECK_BUILTIN(builtin_is_corresponding_member)
+#  define _CCCL_BUILTIN_IS_CORRESPONDING_MEMBER(_C1, _C2, _MPtr1, _MPtr2) \
+    __builtin_is_corresponding_member(_MPtr1, _MPtr2)
+#elif _CCCL_COMPILER(MSVC, >=, 19, 29)
+// using __is_corresponding_member with msvc outside of constexpr context causes linker errors, see
+// https://developercommunity.visualstudio.com/t/Using-compiler-builtins-causes-linking-n/10888080
+// #  define _CCCL_BUILTIN_IS_CORRESPONDING_MEMBER(_C1, _C2, _MPtr1, _MPtr2) __is_corresponding_member(_C1, _C2, _MPtr1,
+// _MPtr2)
+#endif // ^^^ _CCCL_COMPILER(MSVC, >=, 19, 29) ^^^
+
+#if _CCCL_CHECK_BUILTIN(builtin_is_pointer_interconvertible_with_class)
+#  define _CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_WITH_CLASS(_S, _MPtr) \
+    __builtin_is_pointer_interconvertible_with_class(_MPtr)
+#elif _CCCL_COMPILER(MSVC, >=, 19, 29)
+// using __is_pointer_interconvertible_with_class with msvc outside of constexpr context causes linker errors, see
+// https://developercommunity.visualstudio.com/t/Using-compiler-builtins-causes-linking-n/10888080
+// #  define _CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_WITH_CLASS(_S, _MPtr)
+// __is_pointer_interconvertible_with_class(_S, _MPtr)
+#endif // ^^^ _CCCL_COMPILER(MSVC, >=, 19, 29) ^^^
+
+#if _CCCL_CHECK_BUILTIN(builtin_is_virtual_base_of)
+#  define _CCCL_BUILTIN_IS_VIRTUAL_BASE_OF(...) __builtin_is_virtual_base_of(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_is_virtual_base_of)
+
+// nvcc < 13.3 doesn't implement __builtin_is_virtual_base_of
+#if _CCCL_CUDA_COMPILER(NVCC, <, 13, 3)
+#  undef _CCCL_BUILTIN_IS_VIRTUAL_BASE_OF
+#endif // _CCCL_CUDA_COMPILER(NVCC, <, 13, 3)
+
+#if _CCCL_CHECK_BUILTIN(builtin_nanf) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NANF(...) __builtin_nanf(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_nanf)
+
+#if _CCCL_CHECK_BUILTIN(builtin_nan) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NAN(...) __builtin_nan(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_nan)
+
+#if _CCCL_CHECK_BUILTIN(builtin_nanl) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NANL(...) __builtin_nanl(__VA_ARGS__)
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_BUILTIN_NANL(...) static_cast<long double>(__builtin_nan(__VA_ARGS__))
+#endif // _CCCL_CHECK_BUILTIN(builtin_nanl)
+
+#if _CCCL_HAS_FLOAT128()
+#  if _CCCL_CHECK_BUILTIN(builtin_nanf128) || _CCCL_COMPILER(GCC, >=, 7)
+#    define _CCCL_BUILTIN_NANF128(...) __builtin_nanf128(__VA_ARGS__)
+#  endif // _CCCL_CHECK_BUILTIN(builtin_nanf128) || _CCCL_COMPILER(GCC, >=, 7)
+
+// nvcc does not implement __builtin_nanf128
+#  if _CCCL_CUDA_COMPILER(NVCC)
+#    undef _CCCL_BUILTIN_NANF128
+#  endif // _CCCL_CUDA_COMPILER(NVCC)
+#endif // _CCCL_HAS_FLOAT128()
+
+#if _CCCL_CHECK_BUILTIN(builtin_nansf) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NANSF(...) __builtin_nansf(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_nansf)
+
+#if _CCCL_CHECK_BUILTIN(builtin_nans) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NANS(...) __builtin_nans(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_nans)
+
+#if _CCCL_CHECK_BUILTIN(builtin_nansl) || _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_BUILTIN_NANSL(...) __builtin_nansl(__VA_ARGS__)
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_BUILTIN_NANSL(...) static_cast<long double>(__builtin_nans(__VA_ARGS__))
+#endif // _CCCL_CHECK_BUILTIN(builtin_nansl)
+
+#if _CCCL_HAS_FLOAT128()
+#  if _CCCL_CHECK_BUILTIN(builtin_nansf128) || _CCCL_COMPILER(GCC, >=, 7)
+#    define _CCCL_BUILTIN_NANSF128(...) __builtin_nansf128(__VA_ARGS__)
+#  endif // _CCCL_CHECK_BUILTIN(builtin_nansf128) || _CCCL_COMPILER(GCC, >=, 7)
+
+// nvcc does not implement __builtin_nansf128
+#  if _CCCL_CUDA_COMPILER(NVCC)
+#    undef _CCCL_BUILTIN_NANSF128
+#  endif // _CCCL_CUDA_COMPILER(NVCC)
+#endif // _CCCL_HAS_FLOAT128()
+
+#if _CCCL_CHECK_BUILTIN(builtin_memcmp) || _CCCL_COMPILER(GCC) || _CCCL_COMPILER(MSVC, >=, 19, 28)
+#  define _CCCL_BUILTIN_MEMCMP(...) __builtin_memcmp(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_memcmp) || _CCCL_COMPILER(GCC) || _CCCL_COMPILER(MSVC, >=, 19, 28)
+
+#if _CCCL_CUDA_COMPILER(NVCC) || _CCCL_CUDA_COMPILER(CLANG)
+#  undef _CCCL_BUILTIN_MEMCMP
+#endif // _CCCL_CUDA_COMPILER(NVCC) || _CCCL_CUDA_COMPILER(CLANG)
+
+#if _CCCL_CHECK_BUILTIN(builtin_memmove) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_MEMMOVE(...) __builtin_memmove(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_memmove) || _CCCL_COMPILER(GCC)
+
+#if _CCCL_CUDA_COMPILER(NVCC)
+#  undef _CCCL_BUILTIN_MEMMOVE
+#endif // _CCCL_CUDA_COMPILER(NVCC)
+
+#if _CCCL_CHECK_BUILTIN(builtin_operator_new) && _CCCL_CHECK_BUILTIN(builtin_operator_delete) \
+  && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_OPERATOR_DELETE(...) __builtin_operator_delete(__VA_ARGS__)
+#  define _CCCL_BUILTIN_OPERATOR_NEW(...)    __builtin_operator_new(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(builtin_operator_new) && _CCCL_CHECK_BUILTIN(builtin_operator_delete)
+
+#if _CCCL_CHECK_BUILTIN(builtin_prefetch) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_PREFETCH(...) NV_IF_TARGET(NV_IS_HOST, __builtin_prefetch(__VA_ARGS__);)
+#else
+#  define _CCCL_BUILTIN_PREFETCH(...)
+#endif // _CCCL_CHECK_BUILTIN(builtin_prefetch)
+
+#if _CCCL_HAS_BUILTIN(__decay) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_DECAY(...) __decay(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__decay) && clang-cuda
+
+#if _CCCL_CHECK_BUILTIN(has_nothrow_assign) || _CCCL_COMPILER(GCC, >=, 4, 3) || _CCCL_COMPILER(MSVC) \
+  || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_HAS_NOTHROW_ASSIGN(...) __has_nothrow_assign(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(has_nothrow_assign) && gcc >= 4.3
+
+#if _CCCL_CHECK_BUILTIN(has_nothrow_constructor) || _CCCL_COMPILER(GCC, >=, 4, 3) || _CCCL_COMPILER(MSVC) \
+  || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_HAS_NOTHROW_CONSTRUCTOR(...) __has_nothrow_constructor(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(has_nothrow_constructor) && gcc >= 4.3
+
+#if _CCCL_CHECK_BUILTIN(has_nothrow_copy) || _CCCL_COMPILER(GCC, >=, 4, 3) || _CCCL_COMPILER(MSVC) \
+  || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_HAS_NOTHROW_COPY(...) __has_nothrow_copy(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(has_nothrow_copy) && gcc >= 4.3
+
+#if _CCCL_HAS_BUILTIN(__integer_pack)
+#  define _CCCL_BUILTIN_INTEGER_PACK(...) __integer_pack(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__integer_pack)
+
+#if _CCCL_CHECK_BUILTIN(is_array)
+#  define _CCCL_BUILTIN_IS_ARRAY(...) __is_array(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_array)
+
+// clang prior to clang-19 gives wrong results for __is_array of _Tp[0]
+#if _CCCL_COMPILER(CLANG, <, 19)
+#  undef _CCCL_BUILTIN_IS_ARRAY
+#endif // clang < 19
+
+#if _CCCL_CHECK_BUILTIN(is_assignable) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(GCC, >=, 9)
+#  define _CCCL_BUILTIN_IS_ASSIGNABLE(...) __is_assignable(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_assignable) && gcc >= 9.0
+
+#if _CCCL_CHECK_BUILTIN(is_constructible) || _CCCL_COMPILER(GCC, >=, 8) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_IS_CONSTRUCTIBLE(...) __is_constructible(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_constructible) && gcc >= 8.0
+
+#if _CCCL_CHECK_BUILTIN(is_convertible_to) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_IS_CONVERTIBLE_TO(...) __is_convertible_to(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_convertible_to)
+
+#if _CCCL_CHECK_BUILTIN(is_destructible) || _CCCL_COMPILER(MSVC)
+#  define _CCCL_BUILTIN_IS_DESTRUCTIBLE(...) __is_destructible(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_destructible)
+
+#if _CCCL_CHECK_BUILTIN(is_layout_compatible) || _CCCL_COMPILER(MSVC, >=, 19, 29)
+#  define _CCCL_BUILTIN_IS_LAYOUT_COMPATIBLE(...) __is_layout_compatible(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_layout_compatible) || _CCCL_COMPILER(MSVC, >=, 19, 29)
+
+#if _CCCL_CHECK_BUILTIN(is_lvalue_reference)
+#  define _CCCL_BUILTIN_IS_LVALUE_REFERENCE(...) __is_lvalue_reference(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_lvalue_reference)
+
+#if _CCCL_HAS_BUILTIN(__is_member_function_pointer)
+#  define _CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER(...) __is_member_function_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_member_function_pointer)
+
+#if _CCCL_HAS_BUILTIN(__is_member_object_pointer)
+#  define _CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER(...) __is_member_object_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_member_object_pointer)
+
+#if _CCCL_HAS_BUILTIN(__is_member_pointer)
+#  define _CCCL_BUILTIN_IS_MEMBER_POINTER(...) __is_member_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_member_pointer)
+
+#if _CCCL_CHECK_BUILTIN(is_nothrow_assignable) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(...) __is_nothrow_assignable(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_nothrow_assignable)
+
+#if _CCCL_CHECK_BUILTIN(is_nothrow_constructible) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(...) __is_nothrow_constructible(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_nothrow_constructible)
+
+#if _CCCL_CHECK_BUILTIN(is_nothrow_destructible) || _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE(...) __is_nothrow_destructible(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_nothrow_destructible)
+
+#if _CCCL_CHECK_BUILTIN(is_object)
+#  define _CCCL_BUILTIN_IS_OBJECT(...) __is_object(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_object)
+
+// Disabled due to libstdc++ conflict
+#if 0 // _CCCL_HAS_BUILTIN(__is_pointer)
+#  define _CCCL_BUILTIN_IS_POINTER(...) __is_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_pointer)
+
+#if _CCCL_CHECK_BUILTIN(is_pointer_interconvertible_base_of) || _CCCL_COMPILER(MSVC, >=, 19, 29)
+#  define _CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_BASE_OF(...) __is_pointer_interconvertible_base_of(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_pointer_interconvertible_base_of) || _CCCL_COMPILER(MSVC, >=, 19, 29)
+
+#if _CCCL_HAS_BUILTIN(__is_reference)
+#  define _CCCL_BUILTIN_IS_REFERENCE(...) __is_reference(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_reference)
+
+// Disabled due to libstdc++ conflict
+#if 0 // _CCCL_HAS_BUILTIN(__is_referenceable)
+#  define _CCCL_BUILTIN_IS_REFERENCEABLE(...) __is_referenceable(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_referenceable)
+
+#if _CCCL_HAS_BUILTIN(__is_rvalue_reference)
+#  define _CCCL_BUILTIN_IS_RVALUE_REFERENCE(...) __is_rvalue_reference(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_rvalue_reference)
+
+// Disabled due to libstdc++ conflict
+#if 0 // _CCCL_HAS_BUILTIN(__is_scalar)
+#  define _CCCL_BUILTIN_IS_SCALAR(...) __is_scalar(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_scalar)
+
+#if _CCCL_CHECK_BUILTIN(make_integer_seq) || _CCCL_COMPILER(MSVC, >=, 19, 23)
+#  define _CCCL_BUILTIN_MAKE_INTEGER_SEQ(...) __make_integer_seq<__VA_ARGS__>
+#endif // _CCCL_CHECK_BUILTIN(make_integer_seq)
+
+#if _CCCL_HAS_BUILTIN(__reference_constructs_from_temporary)
+#  define _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY(...) __reference_constructs_from_temporary(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__reference_constructs_from_temporary)
+
+#if _CCCL_HAS_BUILTIN(__reference_converts_from_temporary)
+#  define _CCCL_BUILTIN_REFERENCE_CONVERTS_FROM_TEMPORARY(...) __reference_converts_from_temporary(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__reference_converts_from_temporary)
+
+#if _CCCL_HAS_BUILTIN(__remove_const) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_CONST(...) __remove_const(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_const)
+
+#if _CCCL_HAS_BUILTIN(__remove_cv) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_CV(...) __remove_cv(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_cv)
+
+#if _CCCL_HAS_BUILTIN(__remove_cvref) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_CVREF(...) __remove_cvref(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_cvref)
+
+#if _CCCL_COMPILER(NVRTC, <, 12, 4) // NVRTC below 12.4 fails to properly compile that builtin
+#  undef _CCCL_BUILTIN_REMOVE_CVREF
+#endif // _CCCL_COMPILER(NVRTC, <, 12, 4)
+
+#if _CCCL_HAS_BUILTIN(__remove_extent) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_EXTENT(...) __remove_extent(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_extent)
+
+#if _CCCL_HAS_BUILTIN(__remove_pointer) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_POINTER(...) __remove_pointer(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_pointer)
+
+#if _CCCL_HAS_BUILTIN(__remove_reference)
+#  define _CCCL_BUILTIN_REMOVE_REFERENCE_T(...) __remove_reference(__VA_ARGS__)
+#elif _CCCL_HAS_BUILTIN(__remove_reference_t) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_REFERENCE_T(...) __remove_reference_t(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_reference_t)
+
+#if _CCCL_COMPILER(NVRTC, <, 12, 4) // NVRTC below 12.4 fails to properly compile cuda::std::move with that
+#  undef _CCCL_BUILTIN_REMOVE_REFERENCE_T
+#endif // _CCCL_COMPILER(NVRTC, <, 12, 4)
+
+#if _CCCL_HAS_BUILTIN(__remove_volatile) && _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_BUILTIN_REMOVE_VOLATILE(...) __remove_volatile(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__remove_volatile)
+
+#if _CCCL_HAS_BUILTIN(__type_pack_element)
+#  define _CCCL_BUILTIN_TYPE_PACK_ELEMENT(...) __type_pack_element<__VA_ARGS__>
+#endif // _CCCL_HAS_BUILTIN(__type_pack_element)
+
+#if _CCCL_HAS_BUILTIN(__is_complete_type)
+#  define _CCCL_BUILTIN_IS_COMPLETE_TYPE(...) __is_complete_type(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_complete_type)
+
+#if _CCCL_HAS_BUILTIN(__builtin_clear_padding) \
+  && (_CCCL_HOST_COMPILATION() || !(_CCCL_COMPILER(GCC) || _CCCL_COMPILER(NVHPC)))
+#  define _CCCL_BUILTIN_CLEAR_PADDING(...) __builtin_clear_padding(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__builtin_clear_padding) && (_CCCL_HOST_COMPILATION() || !(_CCCL_COMPILER(GCC) ||
+       // _CCCL_COMPILER(NVHPC)))
+
+// NVCC prior to 12.2 have trouble with pack expansion into __type_pack_element in an alias template
+#if _CCCL_CUDACC_BELOW(12, 2)
+#  undef _CCCL_BUILTIN_TYPE_PACK_ELEMENT
+#endif // _CCCL_CUDACC_BELOW(12, 2)
+
+#if _CCCL_COMPILER(MSVC) // To use __builtin_FUNCSIG(), both MSVC and nvcc need to support it
+#  if _CCCL_COMPILER(MSVC, >=, 19, 35) && _CCCL_CUDACC_AT_LEAST(12, 3)
+#    define _CCCL_BUILTIN_PRETTY_FUNCTION() __builtin_FUNCSIG()
+#  else // ^^^ _CCCL_COMPILER(MSVC, >=, 19, 35) ^^^ / vvv _CCCL_COMPILER(MSVC, <, 19, 35) vvv
+#    define _CCCL_BUILTIN_PRETTY_FUNCTION() __FUNCSIG__
+#    define _CCCL_BROKEN_MSVC_FUNCSIG
+#  endif // _CCCL_COMPILER(MSVC, <, 19, 35)
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_BUILTIN_PRETTY_FUNCTION() __PRETTY_FUNCTION__
+#endif // !_CCCL_COMPILER(MSVC)
+
+// GCC's builtin_strlen isn't reliable at constexpr time
+// NVRTC does not expose builtin_strlen
+#if !_CCCL_COMPILER(GCC) && !_CCCL_COMPILER(NVRTC)
+#  define _CCCL_BUILTIN_STRLEN(...) __builtin_strlen(__VA_ARGS__)
+#endif
+
+// The new __nv_atomic builtins are available when __CUDACC_DEVICE_ATOMIC_BUILTINS__ is defined
+#if defined(__CUDACC_DEVICE_ATOMIC_BUILTINS__) && _CCCL_PTX_ARCH() >= 600 && !_CCCL_COMPILER(MSVC)
+#  define _CCCL_HAS_NV_ATOMIC_BUILTINS() 1
+#else // ^^^ has intrinsics ^^^ / vvv no intrinsics
+#  define _CCCL_HAS_NV_ATOMIC_BUILTINS() 0
+#endif // no intrinsics
+
+#endif // __CCCL_BUILTIN_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/compiler.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/compiler.h
new file mode 100644
index 00000000..da41e2f1
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/compiler.h
@@ -0,0 +1,238 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_COMPILER_H
+#define __CCCL_COMPILER_H
+
+#include <cuda/std/__cccl/preprocessor.h>
+
+// Utility to compare version numbers. To use:
+// 1) Define a macro that makes a pair of (major, minor) numbers:
+//    #define MYPRODUCT_MAKE_VERSION(_MAJOR, _MINOR) (_MAJOR * 100 + _MINOR)
+// 2) Define a macro that you will use to compare versions, e.g.:
+//    #define MYPRODUCT(...) _CCCL_VERSION_COMPARE(MYPRODUCT, MYPRODUCT_##__VA_ARGS__)
+//    Signatures:
+//       MYPRODUCT(_PROD)                      - is the product _PROD version non-zero?
+//       MYPRODUCT(_PROD, _OP, _MAJOR)         - compare the product _PROD major version to _MAJOR using operator _OP
+//       MYPRODUCT(_PROD, _OP, _MAJOR, _MINOR) - compare the product _PROD version to _MAJOR._MINOR using operator _OP
+// 3) Define the product version macros as a function-like macro that returns the version number or
+//    _CCCL_VERSION_INVALID() if the version cannot be determined, e. g.:
+//    #define MYPRODUCT_<_PROD>() (1, 2)
+//      or
+//    #define MYPRODUCT_<_PROD>() _CCCL_VERSION_INVALID()
+#define _CCCL_VERSION_MAJOR_(_MAJOR, _MINOR)   _MAJOR
+#define _CCCL_VERSION_MAJOR(_PAIR)             _CCCL_VERSION_MAJOR_ _PAIR
+#define _CCCL_VERSION_INVALID()                (-1, -1)
+#define _CCCL_MAKE_VERSION(_PREFIX, _PAIR)     (_CCCL_PP_EVAL(_CCCL_PP_CAT(_PREFIX, MAKE_VERSION), _CCCL_PP_EXPAND _PAIR))
+#define _CCCL_VERSION_IS_INVALID(_PAIR)        (_CCCL_VERSION_MAJOR(_PAIR) == _CCCL_VERSION_MAJOR(_CCCL_VERSION_INVALID()))
+#define _CCCL_VERSION_COMPARE_1(_PREFIX, _VER) (!_CCCL_VERSION_IS_INVALID(_VER()))
+#define _CCCL_VERSION_COMPARE_3(_PREFIX, _VER, _OP, _MAJOR) \
+  (!_CCCL_VERSION_IS_INVALID(_VER()) && (_CCCL_VERSION_MAJOR(_VER()) _OP _MAJOR))
+#define _CCCL_VERSION_COMPARE_4(_PREFIX, _VER, _OP, _MAJOR, _MINOR) \
+  (!_CCCL_VERSION_IS_INVALID(_VER())                                \
+   && (_CCCL_MAKE_VERSION(_PREFIX, _VER()) _OP _CCCL_MAKE_VERSION(_PREFIX, (_MAJOR, _MINOR))))
+#define _CCCL_VERSION_SELECT_COUNT(_ARG1, _ARG2, _ARG3, _ARG4, _ARG5, ...) _ARG5
+#define _CCCL_VERSION_SELECT2(_ARGS)                                       _CCCL_VERSION_SELECT_COUNT _ARGS
+// MSVC traditonal preprocessor requires an extra level of indirection
+#define _CCCL_VERSION_SELECT(...)         \
+  _CCCL_VERSION_SELECT2(                  \
+    (__VA_ARGS__,                         \
+     _CCCL_VERSION_COMPARE_4,             \
+     _CCCL_VERSION_COMPARE_3,             \
+     _CCCL_VERSION_COMPARE_BAD_ARG_COUNT, \
+     _CCCL_VERSION_COMPARE_1,             \
+     _CCCL_VERSION_COMPARE_BAD_ARG_COUNT))
+#define _CCCL_VERSION_COMPARE(_PREFIX, ...) _CCCL_VERSION_SELECT(__VA_ARGS__)(_PREFIX, __VA_ARGS__)
+
+#define _CCCL_COMPILER_MAKE_VERSION(_MAJOR, _MINOR) ((_MAJOR) * 100 + (_MINOR))
+#define _CCCL_COMPILER(...)                         _CCCL_VERSION_COMPARE(_CCCL_COMPILER_, _CCCL_COMPILER_##__VA_ARGS__)
+
+#define _CCCL_COMPILER_NVHPC()    _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_CLANG()    _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_GCC()      _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_MSVC()     _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_MSVC2019() _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_MSVC2022() _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_MSVC2026() _CCCL_VERSION_INVALID()
+#define _CCCL_COMPILER_NVRTC()    _CCCL_VERSION_INVALID()
+
+// Determine the host compiler and its version
+#if defined(__INTEL_COMPILER)
+#  ifndef CCCL_IGNORE_DEPRECATED_COMPILER
+#    warning \
+      "The Intel C++ Compiler Classic (icc/icpc) is not supported by CCCL. Define CCCL_IGNORE_DEPRECATED_COMPILER to suppress this message."
+#  endif // !CCCL_IGNORE_DEPRECATED_COMPILER
+#elif defined(__NVCOMPILER)
+#  undef _CCCL_COMPILER_NVHPC
+#  define _CCCL_COMPILER_NVHPC() (__NVCOMPILER_MAJOR__, __NVCOMPILER_MINOR__)
+#elif defined(__clang__)
+#  undef _CCCL_COMPILER_CLANG
+#  define _CCCL_COMPILER_CLANG() (__clang_major__, __clang_minor__)
+#elif defined(__GNUC__)
+#  undef _CCCL_COMPILER_GCC
+#  define _CCCL_COMPILER_GCC() (__GNUC__, __GNUC_MINOR__)
+#elif defined(_MSC_VER)
+// see https://learn.microsoft.com/en-us/cpp/overview/compiler-versions?view=msvc-180#version-macros
+#  undef _CCCL_COMPILER_MSVC
+#  define _CCCL_COMPILER_MSVC() (_MSC_VER / 100, _MSC_VER % 100)
+#  if _CCCL_COMPILER(MSVC, <, 19, 20)
+#    ifndef CCCL_IGNORE_DEPRECATED_COMPILER
+#      error \
+        "Visual Studio 2017 (MSC_VER < 1920) and older are not supported by CCCL. Define CCCL_IGNORE_DEPRECATED_COMPILER to suppress this error."
+#    endif
+#  endif // _CCCL_COMPILER(MSVC, <, 19, 20)
+#  if _CCCL_COMPILER(MSVC, >=, 19, 20) && _CCCL_COMPILER(MSVC, <, 19, 30)
+#    undef _CCCL_COMPILER_MSVC2019
+#    define _CCCL_COMPILER_MSVC2019() _CCCL_COMPILER_MSVC()
+#  endif // _CCCL_COMPILER(MSVC, >=, 19, 20) && _CCCL_COMPILER(MSVC, <, 19, 30)
+#  if _CCCL_COMPILER(MSVC, >=, 19, 30) && _CCCL_COMPILER(MSVC, <, 19, 50)
+#    undef _CCCL_COMPILER_MSVC2022
+#    define _CCCL_COMPILER_MSVC2022() _CCCL_COMPILER_MSVC()
+#  endif // _CCCL_COMPILER(MSVC, >=, 19, 30) && _CCCL_COMPILER(MSVC, <, 19, 50)
+#  if _CCCL_COMPILER(MSVC, >=, 19, 50)
+#    undef _CCCL_COMPILER_MSVC2026
+#    define _CCCL_COMPILER_MSVC2026() _CCCL_COMPILER_MSVC()
+#  endif // _CCCL_COMPILER(MSVC, >=, 19, 45)
+#elif defined(__CUDACC_RTC__)
+#  undef _CCCL_COMPILER_NVRTC
+#  define _CCCL_COMPILER_NVRTC() (__CUDACC_VER_MAJOR__, __CUDACC_VER_MINOR__)
+#endif
+
+#if defined(__CUDACC__) || defined(_NVHPC_CUDA)
+#  define _CCCL_CUDA_COMPILATION() 1
+#else // ^^^ compiling .cu file ^^^ / vvv not compiling .cu file vvv
+#  define _CCCL_CUDA_COMPILATION() 0
+#endif // ^^^ not compiling .cu file ^^^
+
+// The CUDA compiler version shares the implementation with the C++ compiler
+#define _CCCL_CUDA_COMPILER_MAKE_VERSION(_MAJOR, _MINOR) _CCCL_COMPILER_MAKE_VERSION(_MAJOR, _MINOR)
+#define _CCCL_CUDA_COMPILER(...)                         _CCCL_VERSION_COMPARE(_CCCL_CUDA_COMPILER_, _CCCL_CUDA_COMPILER_##__VA_ARGS__)
+
+#define _CCCL_CUDA_COMPILER_NVCC()  _CCCL_VERSION_INVALID()
+#define _CCCL_CUDA_COMPILER_NVHPC() _CCCL_VERSION_INVALID()
+#define _CCCL_CUDA_COMPILER_CLANG() _CCCL_VERSION_INVALID()
+#define _CCCL_CUDA_COMPILER_NVRTC() _CCCL_VERSION_INVALID()
+
+// Determine the cuda compiler
+#if _CCCL_CUDA_COMPILATION()
+#  if defined(__NVCC__)
+#    undef _CCCL_CUDA_COMPILER_NVCC
+#    define _CCCL_CUDA_COMPILER_NVCC() (__CUDACC_VER_MAJOR__, __CUDACC_VER_MINOR__)
+#  elif defined(_NVHPC_CUDA)
+#    undef _CCCL_CUDA_COMPILER_NVHPC
+#    define _CCCL_CUDA_COMPILER_NVHPC() _CCCL_COMPILER_NVHPC()
+#  elif defined(__CUDA__) && _CCCL_COMPILER(CLANG)
+#    undef _CCCL_CUDA_COMPILER_CLANG
+#    define _CCCL_CUDA_COMPILER_CLANG() _CCCL_COMPILER_CLANG()
+#  elif _CCCL_COMPILER(NVRTC)
+#    undef _CCCL_CUDA_COMPILER_NVRTC
+#    define _CCCL_CUDA_COMPILER_NVRTC() _CCCL_COMPILER_NVRTC()
+#  endif // ^^^ _CCCL_COMPILER(NVRTC) ^^^
+#endif // _CCCL_CUDA_COMPILATION()
+
+// Determine if we are compiling host code, this includes both CUDA and C++ compilation
+// nvc++ does not define __CUDA_ARCH__, but it compiles both host and device code at the same time
+#if !defined(__CUDA_ARCH__)
+#  define _CCCL_HOST_COMPILATION() 1
+#else // ^^^ compiling host code ^^^ / vvv not compiling host code vvv
+#  define _CCCL_HOST_COMPILATION() 0
+#endif // ^^^ not compiling host code ^^^
+
+#if (_CCCL_CUDA_COMPILATION() && defined(__CUDA_ARCH__)) || _CCCL_CUDA_COMPILER(NVHPC)
+#  define _CCCL_DEVICE_COMPILATION() 1
+#else // ^^^ compiling device code ^^^ / vvv not compiling device code vvv
+#  define _CCCL_DEVICE_COMPILATION() 0
+#endif // ^^^ not compiling device code ^^^
+
+#if defined(__CUDACC_TILE__) && _CCCL_CUDA_COMPILER(NVCC, >, 13, 3)
+#  define _CCCL_TILE_COMPILATION() 1
+#else // ^^^ compiling .cu file in tile mode ^^^ / vvv not compiling in tile mode vvv
+#  define _CCCL_TILE_COMPILATION() 0
+#endif // ^^^ not compiling .cu file ^^^
+
+#define _CCCL_CUDACC_MAKE_VERSION(_MAJOR, _MINOR) ((_MAJOR) * 1000 + (_MINOR) * 10)
+
+// clang-cuda does not define __CUDACC_VER_MAJOR__ and friends. They are instead retrieved from the CUDA_VERSION macro
+// defined in "cuda.h". clang-cuda automatically pre-includes "__clang_cuda_runtime_wrapper.h" which includes "cuda.h"
+#if _CCCL_CUDA_COMPILER(NVCC) || _CCCL_CUDA_COMPILER(NVHPC) || _CCCL_CUDA_COMPILER(NVRTC)
+#  define _CCCL_CUDACC() (__CUDACC_VER_MAJOR__, __CUDACC_VER_MINOR__)
+#elif _CCCL_CUDA_COMPILER(CLANG)
+#  define _CCCL_CUDACC() (CUDA_VERSION / 1000, (CUDA_VERSION % 1000) / 10)
+#endif // ^^^ has cuda compiler ^^^
+
+#if !defined(_CCCL_CUDACC) || !_CCCL_CUDA_COMPILATION()
+#  undef _CCCL_CUDACC
+#  define _CCCL_CUDACC() _CCCL_VERSION_INVALID()
+#endif // !_CCCL_CUDACC || !_CCCL_CUDA_COMPILATION()
+
+#define _CCCL_CUDACC_EQUAL(...)    _CCCL_VERSION_COMPARE(_CCCL_CUDACC_, _CCCL_CUDACC, ==, __VA_ARGS__)
+#define _CCCL_CUDACC_BELOW(...)    _CCCL_VERSION_COMPARE(_CCCL_CUDACC_, _CCCL_CUDACC, <, __VA_ARGS__)
+#define _CCCL_CUDACC_AT_LEAST(...) _CCCL_VERSION_COMPARE(_CCCL_CUDACC_, _CCCL_CUDACC, >=, __VA_ARGS__)
+
+#if _CCCL_CUDA_COMPILATION() && _CCCL_CUDACC_BELOW(12) && !defined(CCCL_IGNORE_DEPRECATED_CUDA_BELOW_12)
+#  error "CUDA versions below 12 are not supported." \
+"Define CCCL_IGNORE_DEPRECATED_CUDA_BELOW_12 to suppress this message."
+#endif
+
+// Define the pragma for the host compiler
+#if _CCCL_COMPILER(MSVC)
+#  define _CCCL_PRAGMA(_ARG) __pragma(_ARG)
+#else
+#  define _CCCL_PRAGMA(_ARG) _Pragma(_CCCL_TO_STRING(_ARG))
+#endif // _CCCL_COMPILER(MSVC)
+
+// Define the proper object format for NVHPC and NVRTC
+#if (_CCCL_COMPILER(NVHPC) && defined(__linux__)) || _CCCL_COMPILER(NVRTC)
+#  ifndef __ELF__
+#    define __ELF__
+#  endif // !__ELF__
+#endif // _CCCL_COMPILER(NVHPC) || _CCCL_COMPILER(NVRTC)
+
+#if _CCCL_DEVICE_COMPILATION()
+#  define _CCCL_PRAGMA_UNROLL(_N)    _CCCL_PRAGMA(unroll _N)
+#  define _CCCL_PRAGMA_UNROLL_FULL() _CCCL_PRAGMA(unroll)
+#elif _CCCL_COMPILER(NVHPC) || _CCCL_COMPILER(NVRTC) || _CCCL_COMPILER(CLANG)
+#  define _CCCL_PRAGMA_UNROLL(_N)    _CCCL_PRAGMA(unroll _N)
+#  define _CCCL_PRAGMA_UNROLL_FULL() _CCCL_PRAGMA(unroll)
+#elif _CCCL_COMPILER(GCC, >=, 8)
+// gcc supports only #pragma GCC unroll, but that causes problems when compiling with nvcc. So, we use #pragma unroll
+// when compiling device code, and #pragma GCC unroll when compiling host code, but we need to suppress the warning
+// about the unknown pragma for nvcc.
+// #pragma GCC unroll does not support full unrolling, so we use the maximum value that it supports.
+#  define _CCCL_PRAGMA_UNROLL(_N) \
+    _CCCL_BEGIN_NV_DIAG_SUPPRESS(1675) _CCCL_PRAGMA(GCC unroll _N) _CCCL_END_NV_DIAG_SUPPRESS()
+#  define _CCCL_PRAGMA_UNROLL_FULL() _CCCL_PRAGMA_UNROLL(65534)
+#else // ^^^ has pragma unroll support ^^^ / vvv no pragma unroll support vvv
+#  define _CCCL_PRAGMA_UNROLL(_N)
+#  define _CCCL_PRAGMA_UNROLL_FULL()
+#endif // ^^^ no pragma unroll support ^^^
+
+#define _CCCL_PRAGMA_NOUNROLL() _CCCL_PRAGMA_UNROLL(1)
+
+#if _CCCL_COMPILER(MSVC)
+#  define _CCCL_WARNING(_MSG) _CCCL_PRAGMA(message(__FILE__ ":" _CCCL_TO_STRING(__LINE__) ": warning: " _MSG))
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_WARNING(_MSG) _CCCL_PRAGMA(GCC warning _MSG)
+#endif // !_CCCL_COMPILER(MSVC)
+
+// Freestanding environment detection
+// NVRTC is treated as freestanding since it has no access to the host standard library
+#if defined(_CCCL_ENABLE_FREESTANDING) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_FREESTANDING() 1
+#  define _CCCL_HOSTED()       0
+#  define _CCCL_HOSTJIT()      (!_CCCL_COMPILER(NVRTC))
+#  define _CCCL_NO_TYPEID
+#else // ^^^ _CCCL_ENABLE_FREESTANDING ||  _CCCL_COMPILER(NVRTC) ^^^ / vvv Hosted environment vvv
+#  define _CCCL_FREESTANDING() 0
+#  define _CCCL_HOSTED()       1
+#  define _CCCL_HOSTJIT()      0
+#endif // Hosted environment
+
+#endif // __CCCL_COMPILER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_capabilities.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_capabilities.h
new file mode 100644
index 00000000..2182bf5f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_capabilities.h
@@ -0,0 +1,118 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_CUDA_CAPABILITIES
+#define __CCCL_CUDA_CAPABILITIES
+
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/cuda_toolkit.h>
+
+#include <nv/target>
+
+/// In device code, _CCCL_PTX_ARCH() expands to the PTX version for which we are compiling.
+/// In host code, _CCCL_PTX_ARCH()'s value is implementation defined.
+#if !defined(__CUDA_ARCH__)
+#  define _CCCL_PTX_ARCH() 0
+#else
+#  define _CCCL_PTX_ARCH() __CUDA_ARCH__
+#endif
+
+#ifdef _CCCL_DOXYGEN_INVOKED // Only parse this during doxygen passes:
+//! When this macro is defined, Programmatic Dependent Launch (PDL) is disabled across CCCL
+#  define CCCL_DISABLE_PDL
+#endif // _CCCL_DOXYGEN_INVOKED
+
+#ifdef CCCL_DISABLE_PDL
+#  define _CCCL_HAS_PDL() 0
+#else // CCCL_DISABLE_PDL
+#  define _CCCL_HAS_PDL() 1
+#endif // CCCL_DISABLE_PDL
+
+#if _CCCL_HAS_PDL()
+// Waits for the previous kernel to complete (when it reaches its final membar). Should be put before the first global
+// memory access in a kernel.
+#  define _CCCL_PDL_GRID_DEPENDENCY_SYNC() NV_IF_TARGET(NV_PROVIDES_SM_90, ::cudaGridDependencySynchronize();)
+// Allows the subsequent kernel in the same stream to launch. Can be put anywhere in a kernel.
+// Heuristic(ahendriksen): put it after the last load.
+#  define _CCCL_PDL_TRIGGER_NEXT_LAUNCH() NV_IF_TARGET(NV_PROVIDES_SM_90, ::cudaTriggerProgrammaticLaunchCompletion();)
+#else // _CCCL_HAS_PDL()
+#  define _CCCL_PDL_GRID_DEPENDENCY_SYNC()
+#  define _CCCL_PDL_TRIGGER_NEXT_LAUNCH()
+#endif // _CCCL_HAS_PDL()
+
+// Check whether the relocatable device code (RDC) is being generated.
+#if defined(__CUDACC_RDC__) || defined(__CLANG_RDC__) || defined(_NVHPC_RDC)
+#  define _CCCL_HAS_RDC() 1
+#else // ^^^ has RDC ^^^ / vvv no RDC vvv
+#  define _CCCL_HAS_RDC() 0
+#endif // ^^^ no RDC ^^^
+
+// Check whether extensible whole program is being compiled.
+#if defined(__CUDACC_EWP__)
+#  define _CCCL_HAS_EWP() 1
+#else // ^^^ has EWP ^^^ / vvv no EWP vvv
+#  define _CCCL_HAS_EWP() 0
+#endif // ^^^ no EWP ^^^
+
+// Control whether device runtime APIs can be used, because they require libcudadevrt to be linked. Defaults to true
+// when RDC or EWP are enabled. Can be disabled by defining CCCL_DISABLE_DEVICE_RUNTIME.
+#if (_CCCL_HAS_RDC() || _CCCL_HAS_EWP()) && !defined(CCCL_DISABLE_DEVICE_RUNTIME)
+#  define _CCCL_HAS_DEVICE_RUNTIME() 1
+#else // ^^^ has device runtime ^^^ / vvv no device runtime vvv
+#  define _CCCL_HAS_DEVICE_RUNTIME() 0
+#endif // ^^^ no device runtime ^^^
+
+// Some functions can be called from host or device code and launch kernels inside. Thus, they use CUDA Dynamic
+// Parallelism (CDP) and require compiling with Relocatable Device Code (RDC) or extensible whole program (EWP) and link
+// with device runtime library. CDP is unsupported with clang-cuda below 22.
+// TODO(bgruber): remove CUB_DISABLE_CDP in CCCL 4.0
+#if _CCCL_HAS_DEVICE_RUNTIME() && !defined(CCCL_DISABLE_CDP) && !defined(CUB_DISABLE_CDP) \
+  && !_CCCL_CUDA_COMPILER(CLANG, <, 22)
+// We have CDP, so host and device APIs can call kernels
+#  define _CCCL_HAS_CDP() 1
+#else // ^^^ has CDP ^^^ / vvv no CDP vvv
+// We don't have CDP, only host APIs can call kernels
+#  define _CCCL_HAS_CDP() 0
+#endif // ^^^ no CDP ^^^
+
+// When RDC is enabled, __launch_bounds__ cannot be used reliably. See #902.
+#if !_CCCL_HAS_RDC() && !defined(CCCL_DISABLE_LAUNCH_BOUNDS)
+#  define _CCCL_LAUNCH_BOUNDS(...) __launch_bounds__(__VA_ARGS__)
+#else // ^^^ has launch bounds attribute ^^^ / vvv no launch bounds attribute vvv
+#  define _CCCL_LAUNCH_BOUNDS(...)
+#endif // ^^^ no launch bounds attribute ^^^
+
+// __block_size__ attribute is available for nvcc and nvrtc 12.9+ for hopper+ architectures. For older nvcc and nvrtc,
+// we can fallback to __cluster_dims__ attribute only specifying the ncta per cluster.
+// This attribute should be used only for cluster launches.
+#if (_CCCL_CUDA_COMPILER(NVCC, >=, 12, 9) || _CCCL_CUDA_COMPILER(NVRTC, >=, 12, 9)) && _CCCL_PTX_ARCH() >= 900
+#  define _CCCL_BLOCK_SIZE(_NTID, _NCTA_PER_CLUSTER) __block_size__(_NTID, _NCTA_PER_CLUSTER)
+#elif (_CCCL_CUDA_COMPILER(NVCC) || _CCCL_CUDA_COMPILER(NVRTC)) && _CCCL_PTX_ARCH() >= 900
+#  define _CCCL_BLOCK_SIZE(_NTID, _NCTA_PER_CLUSTER) __cluster_dims__ _NCTA_PER_CLUSTER
+#else // ^^ has __block_size__ attribute ^^^ / vvv no __block_size__ attribute vvv
+#  define _CCCL_BLOCK_SIZE(_NTID, _NCTA_PER_CLUSTER)
+#endif // ^^^ no __block_size__ attribute ^^^
+
+#if _CCCL_HAS_CDP()
+#  ifdef CUDA_FORCE_CDP1_IF_SUPPORTED
+#    error "CUDA Dynamic Parallelism 1 is no longer supported. Please undefine CUDA_FORCE_CDP1_IF_SUPPORTED."
+#  endif // CUDA_FORCE_CDP1_IF_SUPPORTED
+#endif // _CCCL_HAS_CDP()
+
+#endif // __CCCL_CUDA_CAPABILITIES
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_toolkit.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_toolkit.h
new file mode 100644
index 00000000..e29e1057
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/cuda_toolkit.h
@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_CUDA_TOOLKIT_H
+#define __CCCL_CUDA_TOOLKIT_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_CUDA_COMPILATION() || __has_include(<cuda_runtime_api.h>)
+#  define _CCCL_HAS_CTK() 1
+#else // ^^^ has cuda toolkit ^^^ / vvv no cuda toolkit vvv
+#  define _CCCL_HAS_CTK() 0
+#endif // ^^^ no cuda toolkit ^^^
+
+// CUDA compilers preinclude cuda_runtime.h, so we need to include it here to get the CUDART_VERSION macro
+#if _CCCL_HAS_CTK() && !_CCCL_CUDA_COMPILATION()
+#  include <cuda_runtime_api.h>
+#endif // _CCCL_HAS_CTK() && !_CCCL_CUDA_COMPILATION()
+
+// Check compatibility of the CUDA compiler and CUDA toolkit headers
+// Some users might want to use a newer version of the CTK than the compiler ships. Enable that on their own peril
+#ifndef CCCL_DISABLE_CTK_COMPATIBILITY_CHECK
+#  if _CCCL_CUDA_COMPILATION()
+#    if !_CCCL_CUDACC_EQUAL((CUDART_VERSION / 1000), (CUDART_VERSION % 1000) / 10)
+#      error "CUDA compiler and CUDA toolkit headers are incompatible, please check your include paths"
+#    endif // !_CCCL_CUDACC_EQUAL((CUDART_VERSION / 1000), (CUDART_VERSION % 1000) / 10)
+#  endif // _CCCL_CUDA_COMPILATION()
+#endif // CCCL_DISABLE_CTK_COMPATIBILITY_CHECK
+
+#if _CCCL_HAS_CTK()
+#  define _CCCL_CTK() (CUDART_VERSION / 1000, (CUDART_VERSION % 1000) / 10)
+#else // ^^^ has cuda toolkit ^^^ / vvv no cuda toolkit vvv
+#  define _CCCL_CTK() _CCCL_VERSION_INVALID()
+#endif // ^^^ no cuda toolkit ^^^
+
+#define _CCCL_CTK_MAKE_VERSION(_MAJOR, _MINOR) ((_MAJOR) * 1000 + (_MINOR) * 10)
+#define _CCCL_CTK_BELOW(...)                   _CCCL_VERSION_COMPARE(_CCCL_CTK_, _CCCL_CTK, <, __VA_ARGS__)
+#define _CCCL_CTK_AT_LEAST(...)                _CCCL_VERSION_COMPARE(_CCCL_CTK_, _CCCL_CTK, >=, __VA_ARGS__)
+
+#endif // __CCCL_CUDA_TOOLKIT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/deprecated.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/deprecated.h
new file mode 100644
index 00000000..a8e47095
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/deprecated.h
@@ -0,0 +1,87 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_DEPRECATED_H
+#define __CCCL_DEPRECATED_H
+
+#include <cuda/std/__cccl/attributes.h>
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/dialect.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// Check for deprecation opt outs
+#if defined(LIBCUDACXX_IGNORE_DEPRECATED_CPP_DIALECT)
+#  if !defined(CCCL_IGNORE_DEPRECATED_CPP_DIALECT)
+#    define CCCL_IGNORE_DEPRECATED_CPP_DIALECT
+#  endif
+#endif // suppress all dialect deprecation warnings
+#if defined(LIBCUDACXX_IGNORE_DEPRECATED_CPP_14) || defined(CCCL_IGNORE_DEPRECATED_CPP_DIALECT)
+#  if !defined(CCCL_IGNORE_DEPRECATED_CPP_14)
+#    define CCCL_IGNORE_DEPRECATED_CPP_14
+#  endif
+#endif // suppress all c++14 dialect deprecation warnings
+#if defined(LIBCUDACXX_IGNORE_DEPRECATED_CPP_11) || defined(CCCL_IGNORE_DEPRECATED_CPP_DIALECT) \
+  || defined(CCCL_IGNORE_DEPRECATED_CPP_14)
+#  if !defined(CCCL_IGNORE_DEPRECATED_CPP_11)
+#    define CCCL_IGNORE_DEPRECATED_CPP_11
+#  endif
+#endif // suppress all c++11 dialect deprecation warnings
+#if defined(LIBCUDACXX_IGNORE_DEPRECATED_COMPILER) || defined(THRUST_IGNORE_DEPRECATED_COMPILER) \
+  || defined(CUB_IGNORE_DEPRECATED_COMPILER) || defined(CCCL_IGNORE_DEPRECATED_CPP_DIALECT)      \
+  || defined(CCCL_IGNORE_DEPRECATED_CPP_14) || defined(CCCL_IGNORE_DEPRECATED_CPP_11)
+#  if !defined(CCCL_IGNORE_DEPRECATED_COMPILER)
+#    define CCCL_IGNORE_DEPRECATED_COMPILER
+#  endif
+#endif // suppress all compiler deprecation warnings
+#if defined(LIBCUDACXX_IGNORE_DEPRECATED_API) || defined(THRUST_IGNORE_DEPRECATED_API) \
+  || defined(CUB_IGNORE_DEPRECATED_API)
+#  if !defined(CCCL_IGNORE_DEPRECATED_API)
+#    define CCCL_IGNORE_DEPRECATED_API
+#  endif
+#endif // suppress all API deprecation warnings
+
+#if defined(CCCL_IGNORE_DEPRECATED_API) || defined(_LIBCUDACXX_DISABLE_DEPRECATION_WARNINGS)
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED_BECAUSE(MSG)
+#elif _CCCL_HAS_ATTRIBUTE(deprecated)
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED              __attribute__((deprecated))
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED_BECAUSE(MSG) __attribute__((deprecated(MSG)))
+#else // ^^^ attribute deprecated ^^^ / vvv standard deprecated attribute vvv
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED              [[deprecated]]
+//! deprecated [Since 2.8]
+#  define CCCL_DEPRECATED_BECAUSE(MSG) [[deprecated(MSG)]]
+#endif // ^^^ standard deprecated attribute ^^^
+
+#if _CCCL_STD_VER >= 2020
+#  define _CCCL_DEPRECATED_IN_CXX20 CCCL_DEPRECATED
+#else // ^^^ _CCCL_STD_VER >= 2020 ^^^ / vvv _CCCL_STD_VER < 2020 vvv
+#  define _CCCL_DEPRECATED_IN_CXX20
+#endif // ^^^ _CCCL_STD_VER < 2020 ^^^
+
+#if _CCCL_STD_VER >= 2023
+#  define _CCCL_DEPRECATED_IN_CXX23 CCCL_DEPRECATED
+#else // ^^^ _CCCL_STD_VER >= 2023 ^^^ / vvv _CCCL_STD_VER < 2023 vvv
+#  define _CCCL_DEPRECATED_IN_CXX23
+#endif // ^^^ _CCCL_STD_VER < 2023 ^^^
+
+#endif // __CCCL_DEPRECATED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/diagnostic.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/diagnostic.h
new file mode 100644
index 00000000..a5daafb2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/diagnostic.h
@@ -0,0 +1,145 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_DIAGNOSTIC_H
+#define __CCCL_DIAGNOSTIC_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// Enable us to selectively silence host compiler warnings
+#if _CCCL_COMPILER(CLANG)
+#  define _CCCL_DIAG_PUSH                     _CCCL_PRAGMA(clang diagnostic push)
+#  define _CCCL_DIAG_POP                      _CCCL_PRAGMA(clang diagnostic pop)
+#  define _CCCL_DIAG_SUPPRESS_CLANG(_WARNING) _CCCL_PRAGMA(clang diagnostic ignored _WARNING)
+#  define _CCCL_DIAG_SUPPRESS_GCC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_NVHPC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_MSVC(_WARNING)
+#elif _CCCL_COMPILER(GCC)
+#  define _CCCL_DIAG_PUSH _CCCL_PRAGMA(GCC diagnostic push)
+#  define _CCCL_DIAG_POP  _CCCL_PRAGMA(GCC diagnostic pop)
+#  define _CCCL_DIAG_SUPPRESS_CLANG(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_GCC(_WARNING) _CCCL_PRAGMA(GCC diagnostic ignored _WARNING)
+#  define _CCCL_DIAG_SUPPRESS_NVHPC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_MSVC(_WARNING)
+#elif _CCCL_COMPILER(NVHPC)
+#  define _CCCL_DIAG_PUSH _CCCL_PRAGMA(diagnostic push)
+#  define _CCCL_DIAG_POP  _CCCL_PRAGMA(diagnostic pop)
+#  define _CCCL_DIAG_SUPPRESS_CLANG(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_GCC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_NVHPC(_WARNING) _CCCL_PRAGMA(diag_suppress _WARNING)
+#  define _CCCL_DIAG_SUPPRESS_MSVC(_WARNING)
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_DIAG_PUSH _CCCL_PRAGMA(warning(push))
+#  define _CCCL_DIAG_POP  _CCCL_PRAGMA(warning(pop))
+#  define _CCCL_DIAG_SUPPRESS_CLANG(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_GCC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_NVHPC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_MSVC(_WARNING) _CCCL_PRAGMA(warning(disable : _WARNING))
+#else
+#  define _CCCL_DIAG_PUSH
+#  define _CCCL_DIAG_POP
+#  define _CCCL_DIAG_SUPPRESS_CLANG(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_GCC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_NVHPC(_WARNING)
+#  define _CCCL_DIAG_SUPPRESS_MSVC(_WARNING)
+#endif
+
+// Enable us to selectively silence cuda compiler warnings
+#if _CCCL_CUDA_COMPILER(NVCC) || _CCCL_COMPILER(NVRTC)
+#  if defined(__NVCC_DIAG_PRAGMA_SUPPORT__)
+#    define _CCCL_NV_DIAG_PUSH()               _CCCL_PRAGMA(nv_diagnostic push)
+#    define _CCCL_NV_DIAG_POP()                _CCCL_PRAGMA(nv_diagnostic pop)
+#    define _CCCL_DIAG_SUPPRESS_NVCC(_WARNING) _CCCL_PRAGMA(nv_diag_suppress _WARNING)
+#    define _CCCL_BEGIN_NV_DIAG_SUPPRESS(...) \
+      _CCCL_NV_DIAG_PUSH() _CCCL_PP_FOR_EACH(_CCCL_DIAG_SUPPRESS_NVCC, __VA_ARGS__)
+#    define _CCCL_END_NV_DIAG_SUPPRESS() _CCCL_NV_DIAG_POP()
+#  else // ^^^ __NVCC_DIAG_PRAGMA_SUPPORT__ ^^^ / vvv !__NVCC_DIAG_PRAGMA_SUPPORT__ vvv
+#    define _CCCL_NV_DIAG_PUSH()               _CCCL_PRAGMA(diagnostic push)
+#    define _CCCL_NV_DIAG_POP()                _CCCL_PRAGMA(diagnostic pop)
+#    define _CCCL_DIAG_SUPPRESS_NVCC(_WARNING) _CCCL_PRAGMA(diag_suppress _WARNING)
+#    define _CCCL_BEGIN_NV_DIAG_SUPPRESS(...) \
+      _CCCL_NV_DIAG_PUSH() _CCCL_PP_FOR_EACH(_CCCL_DIAG_SUPPRESS_NVCC, __VA_ARGS__)
+#    define _CCCL_END_NV_DIAG_SUPPRESS() _CCCL_NV_DIAG_POP()
+#  endif // !__NVCC_DIAG_PRAGMA_SUPPORT__
+#else // ^^^ _CCCL_CUDA_COMPILER(NVCC) ^^^ / vvv !_CCCL_CUDA_COMPILER(NVCC) vvv
+#  define _CCCL_NV_DIAG_PUSH()
+#  define _CCCL_NV_DIAG_POP()
+#  define _CCCL_DIAG_SUPPRESS_NVCC(_WARNING)
+#  define _CCCL_BEGIN_NV_DIAG_SUPPRESS(...)
+#  define _CCCL_END_NV_DIAG_SUPPRESS()
+#endif // !_CCCL_CUDA_COMPILER(NVCC)
+
+// Convenient shortcuts to silence common warnings
+#if _CCCL_COMPILER(CLANG)
+#  define _CCCL_SUPPRESS_DEPRECATED_PUSH                   \
+    _CCCL_DIAG_PUSH                                        \
+    _CCCL_DIAG_SUPPRESS_CLANG("-Wdeprecated")              \
+    _CCCL_DIAG_SUPPRESS_CLANG("-Wdeprecated-declarations") \
+    _CCCL_BEGIN_NV_DIAG_SUPPRESS(1444, 20199)
+#  define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  define _CCCL_SUPPRESS_DEPRECATED_POP _CCCL_NV_DIAG_POP() _CCCL_DIAG_POP
+#elif _CCCL_COMPILER(GCC)
+#  define _CCCL_SUPPRESS_DEPRECATED_PUSH                 \
+    _CCCL_DIAG_PUSH                                      \
+    _CCCL_DIAG_SUPPRESS_GCC("-Wdeprecated")              \
+    _CCCL_DIAG_SUPPRESS_GCC("-Wdeprecated-declarations") \
+    _CCCL_BEGIN_NV_DIAG_SUPPRESS(1444, 20199)
+#  define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  define _CCCL_SUPPRESS_DEPRECATED_POP _CCCL_NV_DIAG_POP() _CCCL_DIAG_POP
+#elif _CCCL_COMPILER(NVHPC)
+#  define _CCCL_SUPPRESS_DEPRECATED_PUSH                             \
+    _CCCL_DIAG_PUSH                                                  \
+    _CCCL_DIAG_SUPPRESS_NVHPC(deprecated_entity)                     \
+    _CCCL_DIAG_SUPPRESS_NVHPC(deprecated_entity_with_custom_message) \
+    _CCCL_BEGIN_NV_DIAG_SUPPRESS(1444, 20199)
+#  define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  define _CCCL_SUPPRESS_DEPRECATED_POP _CCCL_NV_DIAG_POP() _CCCL_DIAG_POP
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_SUPPRESS_DEPRECATED_PUSH \
+    _CCCL_DIAG_PUSH                      \
+    _CCCL_DIAG_SUPPRESS_MSVC(4996)       \
+    _CCCL_BEGIN_NV_DIAG_SUPPRESS(1444)
+#  define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  define _CCCL_SUPPRESS_DEPRECATED_POP _CCCL_NV_DIAG_POP() _CCCL_DIAG_POP
+#elif _CCCL_COMPILER(NVRTC)
+#  if _CCCL_COMPILER(NVRTC, >=, 13, 3) && defined(__NVCC_DIAG_PRAGMA_SUPPORT__)
+#    define _CCCL_SUPPRESS_DEPRECATED_PUSH       _CCCL_NV_DIAG_PUSH()
+// NVRTC 13.3 does not honor nv_diag_suppress when it is emitted in the same macro expansion as
+// nv_diagnostic push. Keep the suppression in a separate source-level macro invocation.
+// See https://github.com/NVIDIA/cccl/issues/9170 and nvbug 6239043.
+#    define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG _Pragma("nv_diag_suppress 1444,20199")
+#  else // ^^^ NVRTC >= 13.3 with __NVCC_DIAG_PRAGMA_SUPPORT__ ^^^
+#    define _CCCL_SUPPRESS_DEPRECATED_PUSH _CCCL_BEGIN_NV_DIAG_SUPPRESS(1444, 20199)
+#    define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  endif // ^^^ NVRTC >= 13.3 with __NVCC_DIAG_PRAGMA_SUPPORT__ ^^^
+#  define _CCCL_SUPPRESS_DEPRECATED_POP _CCCL_NV_DIAG_POP()
+#else // unknown compiler
+#  define _CCCL_SUPPRESS_DEPRECATED_PUSH
+#  define _CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+#  define _CCCL_SUPPRESS_DEPRECATED_POP
+#endif // unknown compiler
+
+#if _CCCL_COMPILER(MSVC)
+#  define _CCCL_HAS_PRAGMA_MSVC_WARNING
+#  if !defined(_LIBCUDACXX_DISABLE_PRAGMA_MSVC_WARNING)
+#    define _CCCL_USE_PRAGMA_MSVC_WARNING
+#  endif // !_LIBCUDACXX_DISABLE_PRAGMA_MSVC_WARNING
+#endif // !_CCCL_COMPILER(MSVC)
+
+#endif // __CCCL_DIAGNOSTIC_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/dialect.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/dialect.h
new file mode 100644
index 00000000..989490bb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/dialect.h
@@ -0,0 +1,230 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_DIALECT_H
+#define __CCCL_DIALECT_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/builtin.h>
+#include <cuda/std/__cccl/host_std_lib.h>
+
+///////////////////////////////////////////////////////////////////////////////
+// Determine the C++ standard dialect
+///////////////////////////////////////////////////////////////////////////////
+#if _CCCL_COMPILER(MSVC)
+#  if _MSVC_LANG <= 201103L
+#    define _CCCL_STD_VER 2011
+#  elif _MSVC_LANG <= 201402L
+#    define _CCCL_STD_VER 2014
+#  elif _MSVC_LANG <= 201703L
+#    define _CCCL_STD_VER 2017
+#  elif _MSVC_LANG <= 202002L
+#    define _CCCL_STD_VER 2020
+#  else
+#    define _CCCL_STD_VER 2023 // current year, or date of c++2b ratification
+#  endif
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  if __cplusplus <= 199711L
+#    define _CCCL_STD_VER 2003
+#  elif __cplusplus <= 201103L
+#    define _CCCL_STD_VER 2011
+#  elif __cplusplus <= 201402L
+#    define _CCCL_STD_VER 2014
+#  elif __cplusplus <= 201703L
+#    define _CCCL_STD_VER 2017
+#  elif __cplusplus <= 202002L
+#    define _CCCL_STD_VER 2020
+#  elif __cplusplus <= 202302L
+#    define _CCCL_STD_VER 2023
+#  else
+#    define _CCCL_STD_VER 2024 // current year, or date of c++2c ratification
+#  endif
+#endif // !_CCCL_COMPILER(MSVC)
+
+///////////////////////////////////////////////////////////////////////////////
+// Conditionally enable constexpr per standard dialect
+///////////////////////////////////////////////////////////////////////////////
+
+#if _CCCL_STD_VER >= 2020
+#  define _CCCL_CONSTEXPR_CXX20 constexpr
+#else // ^^^ C++20 ^^^ / vvv C++17 vvv
+#  define _CCCL_CONSTEXPR_CXX20
+#endif // _CCCL_STD_VER <= 2017
+
+#if _CCCL_STD_VER >= 2023
+#  define _CCCL_CONSTEXPR_CXX23 constexpr
+#else // ^^^ C++23 ^^^ / vvv C++20 vvv
+#  define _CCCL_CONSTEXPR_CXX23
+#endif // _CCCL_STD_VER <= 2020
+
+///////////////////////////////////////////////////////////////////////////////
+// Detect whether we can use some language features based on standard dialect
+///////////////////////////////////////////////////////////////////////////////
+
+// concepts are only available from C++20 onwards
+#if _CCCL_STD_VER <= 2017 || __cpp_concepts < 201907L
+#  define _CCCL_HAS_CONCEPTS() 0
+#else // ^^^ no concepts ^^^ / vvv has concepts vvv
+#  define _CCCL_HAS_CONCEPTS() 1
+#endif // ^^^ has concepts ^^^
+
+// Three way comparison is only available from C++20 onwards
+#if _CCCL_STD_VER <= 2017 || __cpp_impl_three_way_comparison < 201907L
+#  define _CCCL_NO_THREE_WAY_COMPARISON
+#endif // _CCCL_STD_VER <= 2017 || __cpp_impl_three_way_comparison < 201907L
+
+// Some compilers turn on pack indexing in pre-C++26 code. We want to use it if it is
+// available.
+#if __cpp_pack_indexing >= 202311L && !_CCCL_CUDA_COMPILER(NVCC) && !_CCCL_COMPILER(CLANG, <, 20)
+#  define _CCCL_HAS_PACK_INDEXING() 1
+#else // ^^^ has pack indexing ^^^ / vvv no pack indexing vvv
+#  define _CCCL_HAS_PACK_INDEXING() 0
+#endif // no pack indexing
+
+#if _CCCL_STD_VER <= 2017 || __cpp_consteval < 201811L
+#  define _CCCL_NO_CONSTEVAL
+#  define _CCCL_CONSTEVAL constexpr
+#else
+#  define _CCCL_CONSTEVAL consteval
+#endif
+
+///////////////////////////////////////////////////////////////////////////////
+// Conditionally use certain language features depending on availability
+///////////////////////////////////////////////////////////////////////////////
+
+// We need to treat host and device separately
+#if _CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC)
+#  define _CCCL_GLOBAL_CONSTANT _CCCL_DEVICE constexpr
+#else // ^^^ _CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC) ^^^ /
+      // vvv !_CCCL_DEVICE_COMPILATION() || _CCCL_CUDA_COMPILER(NVHPC) vvv
+#  define _CCCL_GLOBAL_CONSTANT inline constexpr
+#endif // !_CCCL_DEVICE_COMPILATION() || _CCCL_CUDA_COMPILER(NVHPC)
+
+#if _CCCL_STD_VER >= 2020 && __cpp_constinit >= 201907L
+#  define _CCCL_CONSTINIT constinit
+#else // ^^^ has constinit ^^^ / vvv no constinit vvv
+#  define _CCCL_CONSTINIT _CCCL_REQUIRE_CONSTANT_INITIALIZATION
+#endif // ^^^ no constinit ^^^
+
+// nvcc and nvrtc don't implement multiarg operator[] even in C++23 mode
+#if __cpp_multidimensional_subscript >= 202110L && !_CCCL_CUDA_COMPILER(NVCC) && !_CCCL_CUDA_COMPILER(NVRTC)
+#  define _CCCL_HAS_MULTIARG_OPERATOR_BRACKETS() 1
+#else // ^^^ has multiarg operator[] ^^^ / vvv no multiarg operator[] vvv
+#  define _CCCL_HAS_MULTIARG_OPERATOR_BRACKETS() 0
+#endif // ^^^ no mutiarg operator[] ^^^
+
+// clang 16+, gcc 13+ and nvc++ 25.9+ backport the static subscript operator back to c++17.
+#if __cpp_multidimensional_subscript >= 202211L                                        \
+  || ((_CCCL_COMPILER(CLANG, >=, 16) || _CCCL_COMPILER(GCC, >=, 13)                    \
+       || (_CCCL_COMPILER(NVHPC, >=, 25, 9) && _CCCL_HOST_STD_LIB(LIBSTDCXX, >=, 12))) \
+      && (!_CCCL_CUDA_COMPILATION() || _CCCL_CUDA_COMPILER(CLANG)))
+#  define _CCCL_HAS_STATIC_SUBSCRIPT_OPERATOR() 1
+#else // ^^^ has static operator[] ^^^ / vvv no static operator[] vvv
+#  define _CCCL_HAS_STATIC_SUBSCRIPT_OPERATOR() 0
+#endif // ^^^ no static operator[] ^^^
+
+// nvcc 13+, clang 16+ and gcc 13+ backport the static call operator back to c++17.
+#if __cpp_static_call_operator >= 202207L                                              \
+  || ((_CCCL_COMPILER(CLANG, >=, 16) || _CCCL_COMPILER(GCC, >=, 13)                    \
+       || (_CCCL_COMPILER(NVHPC, >=, 26, 1) && _CCCL_HOST_STD_LIB(LIBSTDCXX, >=, 13))) \
+      && (!_CCCL_CUDA_COMPILATION() || _CCCL_CUDA_COMPILER(NVCC, >=, 13, 0) || _CCCL_CUDA_COMPILER(CLANG)))
+#  define _CCCL_HAS_STATIC_CALL_OPERATOR() 1
+#else // ^^^ has static operator() ^^^ / vvv no static operator() vvv
+#  define _CCCL_HAS_STATIC_CALL_OPERATOR() 0
+#endif // ^^^ no static operator() ^^^
+
+// if consteval requires C++23, but most compilers support it even in C++20 mode while emitting some warnings. Those are
+// silenced in prologue/epilogue. nvcc is happy about using it in C++20 since 13.0, but only when compiling host code.
+// nvc++ requires libstdc++ at least 12 to support if consteval.
+#if _CCCL_STD_VER == 2020                                  \
+  && (_CCCL_COMPILER(GCC, >=, 12) || _CCCL_COMPILER(CLANG) \
+      || (_CCCL_COMPILER(NVHPC) && _CCCL_HOST_STD_LIB(LIBSTDCXX, >=, 12)))
+#  define _CCCL_HAS_IF_CONSTEVAL_IN_CXX20() 1
+#else
+#  define _CCCL_HAS_IF_CONSTEVAL_IN_CXX20() 0
+#endif
+
+// nvcc before 13 doesn't support if consteval at all. Since 13, it accepts if consteval in host code (clang doesn't
+// work) and since 13.1 it works in device code, too.
+#if _CCCL_CUDA_COMPILER(NVCC, <, 13) || (_CCCL_CUDA_COMPILER(NVCC, <, 13, 1) && _CCCL_DEVICE_COMPILATION()) \
+  || (_CCCL_CUDA_COMPILER(NVCC) && _CCCL_COMPILER(CLANG))
+#  undef _CCCL_HAS_IF_CONSTEVAL_IN_CXX20
+#  define _CCCL_HAS_IF_CONSTEVAL_IN_CXX20() 0
+#endif // ^^^ disable if consteval in c++20 for nvcc ^^^
+
+#if __cpp_if_consteval >= 202106L || _CCCL_HAS_IF_CONSTEVAL_IN_CXX20()
+#  define _CCCL_IF_CONSTEVAL             if consteval
+#  define _CCCL_IF_CONSTEVAL_DEFAULT     _CCCL_IF_CONSTEVAL
+#  define _CCCL_IF_NOT_CONSTEVAL         if !consteval
+#  define _CCCL_IF_NOT_CONSTEVAL_DEFAULT _CCCL_IF_NOT_CONSTEVAL
+#elif defined(_CCCL_BUILTIN_IS_CONSTANT_EVALUATED)
+#  if _CCCL_HOST_COMPILATION() && _CCCL_COMPILER(GCC)
+#    define _CCCL_BEGIN_IF_CONSTEVAL_SUPPRESS() _CCCL_DIAG_PUSH _CCCL_DIAG_SUPPRESS_GCC("-Wtautological-compare")
+#    define _CCCL_END_IF_CONSTEVAL_SUPPRESS()   _CCCL_DIAG_POP
+#  else // ^^^ _CCCL_HOST_COMPILATION() && _CCCL_COMPILER(GCC) ^^^ /
+        // vvv !_CCCL_HOST_COMPILATION() || ! _CCCL_COMPILER(GCC) vvv
+#    define _CCCL_BEGIN_IF_CONSTEVAL_SUPPRESS()
+#    define _CCCL_END_IF_CONSTEVAL_SUPPRESS()
+#  endif // ^^^ !_CCCL_HOST_COMPILATION() || ! _CCCL_COMPILER(GCC) ^^^
+
+#  define _CCCL_IF_CONSTEVAL \
+    _CCCL_BEGIN_IF_CONSTEVAL_SUPPRESS() if (_CCCL_BUILTIN_IS_CONSTANT_EVALUATED()) _CCCL_END_IF_CONSTEVAL_SUPPRESS()
+#  define _CCCL_IF_CONSTEVAL_DEFAULT _CCCL_IF_CONSTEVAL
+#  define _CCCL_IF_NOT_CONSTEVAL \
+    _CCCL_BEGIN_IF_CONSTEVAL_SUPPRESS() if (!_CCCL_BUILTIN_IS_CONSTANT_EVALUATED()) _CCCL_END_IF_CONSTEVAL_SUPPRESS()
+#  define _CCCL_IF_NOT_CONSTEVAL_DEFAULT _CCCL_IF_NOT_CONSTEVAL
+#else // ^^^ has is constant evaluated ^^^ / vvv no is constant evaluated vvv
+#  define _CCCL_IF_CONSTEVAL             if constexpr (false)
+#  define _CCCL_IF_CONSTEVAL_DEFAULT     if constexpr (true)
+#  define _CCCL_IF_NOT_CONSTEVAL         if constexpr (true)
+#  define _CCCL_IF_NOT_CONSTEVAL_DEFAULT if constexpr (false)
+#endif // ^^^ no is constant evaluated ^^^
+
+#if _CCCL_STD_VER >= 2020 && __cpp_char8_t >= 201811L
+#  define _CCCL_HAS_CHAR8_T() 1
+#else // ^^^ has char8_t ^^^ / vvv no char8_t vvv
+#  define _CCCL_HAS_CHAR8_T() 0
+#endif // ^^^ no char8_t ^^^
+
+// We currently do not support any of the STL wchar facilities
+#define _CCCL_HAS_WCHAR_T() 0
+
+// Fixme: replace the condition with (!_CCCL_DEVICE_COMPILATION())
+// FIXME: Enable this for clang-cuda in a followup
+#if !_CCCL_CUDA_COMPILATION() && !defined(CCCL_DISABLE_LONG_DOUBLE_SUPPORT)
+#  define _CCCL_HAS_LONG_DOUBLE() 1
+#else // ^^^ has long double ^^^ / vvv no long double vvv
+#  define _CCCL_HAS_LONG_DOUBLE() 0
+#endif // ^^^ no long double ^^^
+
+// clang-21+ and gcc-16+ allow structured bindings to introduce a pack since C++17.
+#if __cpp_structured_bindings >= 202411L || _CCCL_COMPILER(CLANG, >=, 21) || _CCCL_COMPILER(GCC, >=, 16)
+#  define _CCCL_HAS_STRUCTURED_BINDINGS_PACK() 1
+#else // ^^^ has structured bindings with pack ^^^ / vvv no structured bindings with pack vvv
+#  define _CCCL_HAS_STRUCTURED_BINDINGS_PACK() 0
+#endif // ^^^ no structured bindings with pack ^^^
+
+// nvcc doesn't implement structured bindings pack yet.
+#if _CCCL_CUDA_COMPILER(NVCC)
+#  undef _CCCL_HAS_STRUCTURED_BINDINGS_PACK
+#  define _CCCL_HAS_STRUCTURED_BINDINGS_PACK() 0
+#endif // _CCCL_CUDA_COMPILER(NVCC)
+
+#endif // __CCCL_DIALECT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/epilogue.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/epilogue.h
new file mode 100644
index 00000000..70a1c999
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/epilogue.h
@@ -0,0 +1,390 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// !!! DO NOT EDIT THIS FILE !!! This file is generated by utils/generate_prologue_epilogue.py.
+
+// NO include guards here (this file is included multiple times)
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/diagnostic.h>
+
+#if !defined(_CCCL_PROLOGUE_INCLUDED)
+#  error "cccl internal error: <cuda/std/__cccl/prologue.h> must be included before <cuda/std/__cccl/epilogue.h>"
+#endif
+#undef _CCCL_PROLOGUE_INCLUDED
+
+_CCCL_NV_DIAG_POP()
+_CCCL_DIAG_POP
+
+// __declspec modifiers
+
+#if defined(align)
+#  error \
+    "cccl internal error: macro `align` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_align)
+#  pragma pop_macro("align")
+#  undef _CCCL_POP_MACRO_align
+#endif
+
+#if defined(allocate)
+#  error \
+    "cccl internal error: macro `allocate` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_allocate)
+#  pragma pop_macro("allocate")
+#  undef _CCCL_POP_MACRO_allocate
+#endif
+
+#if defined(allocator)
+#  error \
+    "cccl internal error: macro `allocator` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_allocator)
+#  pragma pop_macro("allocator")
+#  undef _CCCL_POP_MACRO_allocator
+#endif
+
+#if defined(appdomain)
+#  error \
+    "cccl internal error: macro `appdomain` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_appdomain)
+#  pragma pop_macro("appdomain")
+#  undef _CCCL_POP_MACRO_appdomain
+#endif
+
+#if defined(code_seg)
+#  error \
+    "cccl internal error: macro `code_seg` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_code_seg)
+#  pragma pop_macro("code_seg")
+#  undef _CCCL_POP_MACRO_code_seg
+#endif
+
+#if defined(deprecated)
+#  error \
+    "cccl internal error: macro `deprecated` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_deprecated)
+#  pragma pop_macro("deprecated")
+#  undef _CCCL_POP_MACRO_deprecated
+#endif
+
+#if defined(dllimport)
+#  error \
+    "cccl internal error: macro `dllimport` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_dllimport)
+#  pragma pop_macro("dllimport")
+#  undef _CCCL_POP_MACRO_dllimport
+#endif
+
+#if defined(dllexport)
+#  error \
+    "cccl internal error: macro `dllexport` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_dllexport)
+#  pragma pop_macro("dllexport")
+#  undef _CCCL_POP_MACRO_dllexport
+#endif
+
+#if defined(empty_bases)
+#  error \
+    "cccl internal error: macro `empty_bases` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_empty_bases)
+#  pragma pop_macro("empty_bases")
+#  undef _CCCL_POP_MACRO_empty_bases
+#endif
+
+#if defined(hybrid_patchable)
+#  error \
+    "cccl internal error: macro `hybrid_patchable` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_hybrid_patchable)
+#  pragma pop_macro("hybrid_patchable")
+#  undef _CCCL_POP_MACRO_hybrid_patchable
+#endif
+
+#if defined(jitintrinsic)
+#  error \
+    "cccl internal error: macro `jitintrinsic` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_jitintrinsic)
+#  pragma pop_macro("jitintrinsic")
+#  undef _CCCL_POP_MACRO_jitintrinsic
+#endif
+
+#if defined(lifetimebound)
+#  error \
+    "cccl internal error: macro `lifetimebound` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_lifetimebound)
+#  pragma pop_macro("lifetimebound")
+#  undef _CCCL_POP_MACRO_lifetimebound
+#endif
+
+#if defined(naked)
+#  error \
+    "cccl internal error: macro `naked` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_naked)
+#  pragma pop_macro("naked")
+#  undef _CCCL_POP_MACRO_naked
+#endif
+
+#if defined(noalias)
+#  error \
+    "cccl internal error: macro `noalias` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_noalias)
+#  pragma pop_macro("noalias")
+#  undef _CCCL_POP_MACRO_noalias
+#endif
+
+#if defined(noinline)
+#  error \
+    "cccl internal error: macro `noinline` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_noinline)
+#  pragma pop_macro("noinline")
+#  undef _CCCL_POP_MACRO_noinline
+#endif
+
+#if defined(noreturn)
+#  error \
+    "cccl internal error: macro `noreturn` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_noreturn)
+#  pragma pop_macro("noreturn")
+#  undef _CCCL_POP_MACRO_noreturn
+#endif
+
+#if defined(nothrow)
+#  error \
+    "cccl internal error: macro `nothrow` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_nothrow)
+#  pragma pop_macro("nothrow")
+#  undef _CCCL_POP_MACRO_nothrow
+#endif
+
+#if defined(novtable)
+#  error \
+    "cccl internal error: macro `novtable` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_novtable)
+#  pragma pop_macro("novtable")
+#  undef _CCCL_POP_MACRO_novtable
+#endif
+
+#if defined(no_sanitize_address)
+#  error \
+    "cccl internal error: macro `no_sanitize_address` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_no_sanitize_address)
+#  pragma pop_macro("no_sanitize_address")
+#  undef _CCCL_POP_MACRO_no_sanitize_address
+#endif
+
+#if defined(process)
+#  error \
+    "cccl internal error: macro `process` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_process)
+#  pragma pop_macro("process")
+#  undef _CCCL_POP_MACRO_process
+#endif
+
+#if defined(property)
+#  error \
+    "cccl internal error: macro `property` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_property)
+#  pragma pop_macro("property")
+#  undef _CCCL_POP_MACRO_property
+#endif
+
+#if defined(restrict)
+#  error \
+    "cccl internal error: macro `restrict` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_restrict)
+#  pragma pop_macro("restrict")
+#  undef _CCCL_POP_MACRO_restrict
+#endif
+
+#if defined(safebuffers)
+#  error \
+    "cccl internal error: macro `safebuffers` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_safebuffers)
+#  pragma pop_macro("safebuffers")
+#  undef _CCCL_POP_MACRO_safebuffers
+#endif
+
+#if defined(selectany)
+#  error \
+    "cccl internal error: macro `selectany` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_selectany)
+#  pragma pop_macro("selectany")
+#  undef _CCCL_POP_MACRO_selectany
+#endif
+
+#if defined(spectre)
+#  error \
+    "cccl internal error: macro `spectre` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_spectre)
+#  pragma pop_macro("spectre")
+#  undef _CCCL_POP_MACRO_spectre
+#endif
+
+#if defined(thread)
+#  error \
+    "cccl internal error: macro `thread` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_thread)
+#  pragma pop_macro("thread")
+#  undef _CCCL_POP_MACRO_thread
+#endif
+
+#if defined(uuid)
+#  error \
+    "cccl internal error: macro `uuid` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_uuid)
+#  pragma pop_macro("uuid")
+#  undef _CCCL_POP_MACRO_uuid
+#endif
+
+// [[msvc::attribute]] attributes
+
+#if defined(msvc)
+#  error \
+    "cccl internal error: macro `msvc` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_msvc)
+#  pragma pop_macro("msvc")
+#  undef _CCCL_POP_MACRO_msvc
+#endif
+
+#if defined(flatten)
+#  error \
+    "cccl internal error: macro `flatten` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_flatten)
+#  pragma pop_macro("flatten")
+#  undef _CCCL_POP_MACRO_flatten
+#endif
+
+#if defined(forceinline)
+#  error \
+    "cccl internal error: macro `forceinline` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_forceinline)
+#  pragma pop_macro("forceinline")
+#  undef _CCCL_POP_MACRO_forceinline
+#endif
+
+#if defined(forceinline_calls)
+#  error \
+    "cccl internal error: macro `forceinline_calls` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_forceinline_calls)
+#  pragma pop_macro("forceinline_calls")
+#  undef _CCCL_POP_MACRO_forceinline_calls
+#endif
+
+#if defined(intrinsic)
+#  error \
+    "cccl internal error: macro `intrinsic` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_intrinsic)
+#  pragma pop_macro("intrinsic")
+#  undef _CCCL_POP_MACRO_intrinsic
+#endif
+
+#if defined(noinline)
+#  error \
+    "cccl internal error: macro `noinline` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_noinline)
+#  pragma pop_macro("noinline")
+#  undef _CCCL_POP_MACRO_noinline
+#endif
+
+#if defined(noinline_calls)
+#  error \
+    "cccl internal error: macro `noinline_calls` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_noinline_calls)
+#  pragma pop_macro("noinline_calls")
+#  undef _CCCL_POP_MACRO_noinline_calls
+#endif
+
+#if defined(no_tls_guard)
+#  error \
+    "cccl internal error: macro `no_tls_guard` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_no_tls_guard)
+#  pragma pop_macro("no_tls_guard")
+#  undef _CCCL_POP_MACRO_no_tls_guard
+#endif
+
+// Windows nasty macros
+
+#if defined(min)
+#  error \
+    "cccl internal error: macro `min` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_min)
+#  pragma pop_macro("min")
+#  undef _CCCL_POP_MACRO_min
+#endif
+
+#if defined(max)
+#  error \
+    "cccl internal error: macro `max` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_max)
+#  pragma pop_macro("max")
+#  undef _CCCL_POP_MACRO_max
+#endif
+
+#if defined(interface)
+#  error \
+    "cccl internal error: macro `interface` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_interface)
+#  pragma pop_macro("interface")
+#  undef _CCCL_POP_MACRO_interface
+#endif
+
+// sal.h on Windows
+
+#if defined(__valid)
+#  error \
+    "cccl internal error: macro `__valid` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO___valid)
+#  pragma pop_macro("__valid")
+#  undef _CCCL_POP_MACRO___valid
+#endif
+
+#if defined(__callback)
+#  error \
+    "cccl internal error: macro `__callback` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO___callback)
+#  pragma pop_macro("__callback")
+#  undef _CCCL_POP_MACRO___callback
+#endif
+
+// other macros
+
+#if defined(clang)
+#  error \
+    "cccl internal error: macro `clang` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_clang)
+#  pragma pop_macro("clang")
+#  undef _CCCL_POP_MACRO_clang
+#endif
+
+// sys/sysmacros.h on linux
+
+#if defined(major)
+#  error \
+    "cccl internal error: macro `major` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_major)
+#  pragma pop_macro("major")
+#  undef _CCCL_POP_MACRO_major
+#endif
+
+#if defined(minor)
+#  error \
+    "cccl internal error: macro `minor` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_minor)
+#  pragma pop_macro("minor")
+#  undef _CCCL_POP_MACRO_minor
+#endif
+
+#if defined(makedev)
+#  error \
+    "cccl internal error: macro `makedev` was redefined between <cuda/std/__cccl/prologue.h> and <cuda/std/__cccl/epilogue.h>"
+#elif defined(_CCCL_POP_MACRO_makedev)
+#  pragma pop_macro("makedev")
+#  undef _CCCL_POP_MACRO_makedev
+#endif
+
+// NO include guards here (this file is included multiple times)
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/exceptions.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/exceptions.h
new file mode 100644
index 00000000..57edeaf5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/exceptions.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_EXCEPTIONS_H
+#define __CCCL_EXCEPTIONS_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/execution_space.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if defined(CCCL_DISABLE_EXCEPTIONS) // Escape hatch for users to manually disable exceptions
+#  define _CCCL_HAS_EXCEPTIONS() 0
+#elif _CCCL_COMPILER(NVRTC) // NVRTC has no exceptions
+#  define _CCCL_HAS_EXCEPTIONS() 0
+#elif _CCCL_COMPILER(MSVC) // MSVC needs special checks for `_HAS_EXCEPTIONS` and `_CPPUNWIND`
+#  define _CCCL_HAS_EXCEPTIONS() ((_HAS_EXCEPTIONS != 0) && (_CPPUNWIND != 0)) // disabled with /EH
+#else // other compilers use `__EXCEPTIONS`
+#  define _CCCL_HAS_EXCEPTIONS() (__EXCEPTIONS) // disabled with -fno-exceptions
+#endif // has exceptions
+
+#if _CCCL_HAS_EXCEPTIONS() && __cpp_constexpr_exceptions >= 202411L
+#  define _CCCL_HAS_CONSTEXPR_EXCEPTIONS() 1
+#else // ^^^ has constexpr exceptions ^^^ / vvv no constexpr exceptions vvv
+#  define _CCCL_HAS_CONSTEXPR_EXCEPTIONS() 0
+#endif // ^^^ no constexpr exceptions ^^^
+
+#endif // __CCCL_EXCEPTIONS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/execution_space.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/execution_space.h
new file mode 100644
index 00000000..ed3ddd9b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/execution_space.h
@@ -0,0 +1,87 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_EXECUTION_SPACE_H
+#define __CCCL_EXECUTION_SPACE_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/cuda_capabilities.h>
+
+#if _CCCL_CUDA_COMPILATION()
+#  define _CCCL_HOST        __host__
+#  define _CCCL_DEVICE      __device__
+#  define _CCCL_HOST_DEVICE __host__ __device__
+#else // ^^^ _CCCL_CUDA_COMPILATION ^^^ / vvv !_CCCL_CUDA_COMPILATION vvv
+#  define _CCCL_HOST
+#  define _CCCL_DEVICE
+#  define _CCCL_HOST_DEVICE
+#endif // !_CCCL_CUDA_COMPILATION
+
+#if _CCCL_TILE_COMPILATION()
+#  define _CCCL_TILE __tile__
+#else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION() vvv
+#  define _CCCL_TILE
+#endif // ^^^ !_CCCL_TILE_COMPILATION() ^^^
+
+// clang-cuda before version 22 requires __host__ __device__ annotations on deduction guides
+#if _CCCL_CUDA_COMPILER(CLANG, <, 22)
+#  define _CCCL_DEDUCTION_GUIDE_ATTRIBUTES _CCCL_HOST_DEVICE
+#else // ^^^ _CCCL_CUDA_COMPILER(CLANG, <, 22) ^^^ / vvv !_CCCL_CUDA_COMPILER(CLANG, <, 22) vvv
+#  define _CCCL_DEDUCTION_GUIDE_ATTRIBUTES
+#endif // ^^ !_CCCL_CUDA_COMPILER(CLANG, <, 22) ^^^
+
+// Global variables of non builtin types are only device accessible if they are marked as `__device__`
+#if _CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC)
+#  define _CCCL_GLOBAL_VARIABLE _CCCL_DEVICE
+#else // ^^^ _CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC) ^^^ /
+      // vvv !_CCCL_DEVICE_COMPILATION() || _CCCL_CUDA_COMPILER(NVHPC) vvv
+#  define _CCCL_GLOBAL_VARIABLE
+#endif // ^^^ !_CCCL_DEVICE_COMPILATION() || _CCCL_CUDA_COMPILER(NVHPC) ^^^
+
+#if (_CCCL_CUDA_COMPILER(NVCC, >=, 12, 8) || _CCCL_CUDA_COMPILER(NVRTC) || _CCCL_CUDA_COMPILER(CLANG, >=, 20)) \
+  && _CCCL_PTX_ARCH() >= 700
+#  define _CCCL_HAS_GRID_CONSTANT() 1
+#  define _CCCL_GRID_CONSTANT       __grid_constant__
+#else // ^^^ has __grid_constant__ ^^^ / vvv no __grid_constant__ vvv
+#  define _CCCL_HAS_GRID_CONSTANT() 0
+#  define _CCCL_GRID_CONSTANT
+#endif // ^^^ no __grid_constant__ ^^^
+
+#if !defined(_CCCL_EXEC_CHECK_DISABLE)
+#  if _CCCL_CUDA_COMPILER(NVCC)
+#    define _CCCL_EXEC_CHECK_DISABLE _CCCL_PRAGMA(nv_exec_check_disable)
+#  else
+#    define _CCCL_EXEC_CHECK_DISABLE
+#  endif // _CCCL_CUDA_COMPILER(NVCC)
+#endif // !_CCCL_EXEC_CHECK_DISABLE
+
+#if _CCCL_CUDA_COMPILER(NVHPC)
+#  define _CCCL_TARGET_CONSTEXPR
+#else // ^^^ _CCCL_CUDA_COMPILER(NVHPC) ^^^ / vvv !_CCCL_CUDA_COMPILER(NVHPC) vvv
+#  define _CCCL_TARGET_CONSTEXPR constexpr
+#endif // ^^^ !_CCCL_CUDA_COMPILER(NVHPC) ^^^
+
+//! @brief List of all known PTX architectures supported by this CCCL version.
+#define _CCCL_KNOWN_CUDA_ARCH_LIST 50, 52, 53, 60, 61, 62, 70, 75, 80, 86, 87, 88, 89, 90, 100, 103, 110, 120, 121
+
+//! @brief List of all known architecture specific architectures supported by this CCCL version.
+#define _CCCL_KNOWN_CUDA_ARCH_SPECIFIC_LIST 90, 100, 103, 110, 120, 121
+
+#endif // __CCCL_EXECUTION_SPACE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/extended_data_types.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/extended_data_types.h
new file mode 100644
index 00000000..658036c2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/extended_data_types.h
@@ -0,0 +1,148 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_EXTENDED_DATA_TYPES_H
+#define __CCCL_EXTENDED_DATA_TYPES_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/architecture.h>
+#include <cuda/std/__cccl/cuda_capabilities.h>
+#include <cuda/std/__cccl/cuda_toolkit.h>
+#include <cuda/std/__cccl/diagnostic.h>
+#include <cuda/std/__cccl/os.h>
+#include <cuda/std/__cccl/preprocessor.h>
+
+#define _CCCL_HAS_INT128()   0
+#define _CCCL_HAS_NVFP4()    0
+#define _CCCL_HAS_NVFP6()    0
+#define _CCCL_HAS_NVFP8()    0
+#define _CCCL_HAS_NVFP16()   0
+#define _CCCL_HAS_NVBF16()   0
+#define _CCCL_HAS_FLOAT128() 0
+
+#if _CCCL_TILE_COMPILATION() // TODO(miscco): Fix access to extended floating point types
+#  define CCCL_DISABLE_NVFP4_SUPPORT
+#  define CCCL_DISABLE_NVFP6_SUPPORT
+#  define CCCL_DISABLE_NVFP8_SUPPORT
+#  define CCCL_DISABLE_INT128_SUPPORT
+#  define CCCL_DISABLE_FLOAT128_SUPPORT
+#endif // _CCCL_TILE_COMPILATION()
+
+#if !defined(CCCL_DISABLE_INT128_SUPPORT) && _CCCL_OS(LINUX) \
+  && ((_CCCL_COMPILER(NVRTC) && defined(__CUDACC_RTC_INT128__)) || defined(__SIZEOF_INT128__))
+#  undef _CCCL_HAS_INT128
+#  define _CCCL_HAS_INT128() 1
+#endif
+
+#if __has_include(<cuda_fp16.h>) && (_CCCL_HAS_CTK() || defined(LIBCUDACXX_ENABLE_HOST_NVFP16)) \
+                      && !defined(CCCL_DISABLE_FP16_SUPPORT)
+#  undef _CCCL_HAS_NVFP16
+#  define _CCCL_HAS_NVFP16() 1
+struct __half;
+struct __half2;
+#endif
+
+#if __has_include(<cuda_bf16.h>) && _CCCL_HAS_NVFP16() && !defined(CCCL_DISABLE_BF16_SUPPORT)
+#  undef _CCCL_HAS_NVBF16
+#  define _CCCL_HAS_NVBF16() 1
+struct __nv_bfloat16;
+struct __nv_bfloat162;
+#endif
+
+#if __has_include(<cuda_fp8.h>) && _CCCL_HAS_NVFP16() && _CCCL_HAS_NVBF16() && !defined(CCCL_DISABLE_NVFP8_SUPPORT)
+#  undef _CCCL_HAS_NVFP8
+#  define _CCCL_HAS_NVFP8() 1
+struct __nv_fp8_e5m2;
+struct __nv_fp8x2_e5m2;
+struct __nv_fp8x4_e5m2;
+
+struct __nv_fp8_e4m3;
+struct __nv_fp8x2_e4m3;
+struct __nv_fp8x4_e4m3;
+
+#  if _CCCL_CTK_AT_LEAST(12, 8)
+struct __nv_fp8_e8m0;
+struct __nv_fp8x2_e8m0;
+struct __nv_fp8x4_e8m0;
+#  endif // _CCCL_CTK_AT_LEAST(12, 8)
+#endif
+
+#if __has_include(<cuda_fp6.h>) && _CCCL_HAS_NVFP8() && !_CCCL_CUDA_COMPILER(NVHPC) \
+                      && !defined(CCCL_DISABLE_NVFP6_SUPPORT)
+#  undef _CCCL_HAS_NVFP6
+#  define _CCCL_HAS_NVFP6() 1
+struct __nv_fp6_e3m2;
+struct __nv_fp6x2_e3m2;
+struct __nv_fp6x4_e3m2;
+
+struct __nv_fp6_e2m3;
+struct __nv_fp6x2_e2m3;
+struct __nv_fp6x4_e2m3;
+#endif
+
+#if __has_include(<cuda_fp4.h>) && _CCCL_HAS_NVFP6() && !defined(CCCL_DISABLE_NVFP4_SUPPORT)
+#  undef _CCCL_HAS_NVFP4
+#  define _CCCL_HAS_NVFP4() 1
+struct __nv_fp4_e2m1;
+struct __nv_fp4x2_e2m1;
+struct __nv_fp4x4_e2m1;
+#endif
+
+#define _CCCL_HAS_NVFP4_E2M1() _CCCL_HAS_NVFP4()
+#define _CCCL_HAS_NVFP6_E2M3() _CCCL_HAS_NVFP6()
+#define _CCCL_HAS_NVFP6_E3M2() _CCCL_HAS_NVFP6()
+#define _CCCL_HAS_NVFP8_E4M3() _CCCL_HAS_NVFP8()
+#define _CCCL_HAS_NVFP8_E5M2() _CCCL_HAS_NVFP8()
+#define _CCCL_HAS_NVFP8_E8M0() (_CCCL_HAS_NVFP8() && _CCCL_CTK_AT_LEAST(12, 8))
+
+/***********************************************************************************************************************
+ * __float128
+ **********************************************************************************************************************/
+
+#if !defined(CCCL_DISABLE_FLOAT128_SUPPORT) && _CCCL_HAS_INT128() && _CCCL_OS(LINUX) && !_CCCL_HOST_ARCH(ARM64) \
+  && !_CCCL_TILE_COMPILATION()
+// Detect host compiler support
+#  if (defined(__CUDACC_RTC_FLOAT128__) || defined(__SIZEOF_FLOAT128__) || defined(__FLOAT128__))
+#    if _CCCL_DEVICE_COMPILATION()
+// Only NVCC and NVRTC 12.8+ on architectures at least SM100 supports __float128 on device
+#      if (_CCCL_CUDA_COMPILER(NVCC, >=, 12, 8) || _CCCL_CUDA_COMPILER(NVRTC, >=, 12, 8)) && _CCCL_PTX_ARCH() >= 1000
+#        undef _CCCL_HAS_FLOAT128
+#        define _CCCL_HAS_FLOAT128() 1
+#      endif // _CCCL_CUDA_COMPILER(NVCC) && _CCCL_PTX_ARCH() >= 1000
+#    else // ^^^ _CCCL_DEVICE_COMPILATION() ^^^ / vvv !_CCCL_DEVICE_COMPILATION() vvv
+#      undef _CCCL_HAS_FLOAT128
+#      define _CCCL_HAS_FLOAT128() 1
+#    endif // ^^^ !_CCCL_DEVICE_COMPILATION() ^^^
+#  endif // Host compiler support
+#endif // !defined(CCCL_DISABLE_FLOAT128_SUPPORT) && _CCCL_HAS_INT128() && _CCCL_OS(LINUX) && !_CCCL_HOST_ARCH(ARM64)
+
+// gcc does not allow to use q/Q floating point literals when __STRICT_ANSI__ is defined. They may be allowed by
+// -fext-numeric-literals, but there is no way to detect it in the preprocessor. The user is required to define
+// CCCL_GCC_HAS_EXTENDED_NUMERIC_LITERALS in this case. Otherwise, we disable the __float128 support.
+//
+// Note: since GCC 13, we could use f128/F128 literals, but for values > DBL_MAX, the compilation with nvcc fails due to
+//       "floating constant is out of range".
+#if _CCCL_HAS_FLOAT128() && _CCCL_COMPILER(GCC) && defined(__STRICT_ANSI__) \
+  && !defined(CCCL_GCC_HAS_EXTENDED_NUMERIC_LITERALS)
+#  undef _CCCL_HAS_FLOAT128
+#  define _CCCL_HAS_FLOAT128() 0
+#endif // _CCCL_HAS_FLOAT128()
+
+#endif // __CCCL_EXTENDED_DATA_TYPES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/host_std_lib.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/host_std_lib.h
new file mode 100644
index 00000000..634100ea
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/host_std_lib.h
@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_HOST_STD_LIB_H
+#define __CCCL_HOST_STD_LIB_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/preprocessor.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#define _CCCL_HOST_STD_LIB_LIBSTDCXX() _CCCL_VERSION_INVALID()
+#define _CCCL_HOST_STD_LIB_LIBCXX()    _CCCL_VERSION_INVALID()
+#define _CCCL_HOST_STD_LIB_STL()       _CCCL_VERSION_INVALID()
+
+// include a minimal header
+#if __has_include(<version>)
+#  include <version>
+#elif __has_include(<ciso646>)
+#  include <ciso646>
+#endif // ^^^ __has_include(<ciso646>) ^^^
+
+#define _CCCL_HOST_STD_LIB_MAKE_VERSION(_MAJOR, _MINOR) ((_MAJOR) * 100 + (_MINOR))
+#define _CCCL_HOST_STD_LIB(...)                         _CCCL_VERSION_COMPARE(_CCCL_HOST_STD_LIB_, _CCCL_HOST_STD_LIB_##__VA_ARGS__)
+
+#if _CCCL_HOSTED()
+#  if defined(_MSVC_STL_VERSION)
+#    undef _CCCL_HOST_STD_LIB_STL
+#    define _CCCL_HOST_STD_LIB_STL() (_MSVC_STL_VERSION, 0)
+#  elif defined(__GLIBCXX__)
+#    undef _CCCL_HOST_STD_LIB_LIBSTDCXX
+#    define _CCCL_HOST_STD_LIB_LIBSTDCXX() (_GLIBCXX_RELEASE, 0)
+#  elif defined(_LIBCPP_VERSION)
+#    undef _CCCL_HOST_STD_LIB_LIBCXX
+// since llvm-16, the version scheme has been changed from MMppp to MMmmpp
+#    if _LIBCPP_VERSION / 10000 < 2
+#      define _CCCL_HOST_STD_LIB_LIBCXX() (_LIBCPP_VERSION / 1000, 0)
+#    else
+#      define _CCCL_HOST_STD_LIB_LIBCXX() (_LIBCPP_VERSION / 10000, (_LIBCPP_VERSION / 100) % 100)
+#    endif
+#  endif // ^^^ _LIBCPP_VERSION ^^^
+#endif // _CCCL_HOSTED()
+
+#define _CCCL_HAS_HOST_STD_LIB() \
+  (_CCCL_HOST_STD_LIB(LIBSTDCXX) || _CCCL_HOST_STD_LIB(LIBCXX) || _CCCL_HOST_STD_LIB(STL))
+
+#endif // __CCCL_HOST_STD_LIB_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/is_non_narrowing_convertible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/is_non_narrowing_convertible.h
new file mode 100644
index 00000000..101d95e6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/is_non_narrowing_convertible.h
@@ -0,0 +1,71 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_IS_NON_NARROWING_CONVERTIBLE_H
+#define __CCCL_IS_NON_NARROWING_CONVERTIBLE_H
+
+#include <cuda/std/__cccl/compiler.h>
+
+//! There is compiler bug that results in incorrect results for the below `__is_non_narrowing_convertible` check.
+//! This breaks some common functionality, so this *must* be included outside of a system header. See nvbug4867473.
+#if defined(_CCCL_FORCE_SYSTEM_HEADER_GCC) || defined(_CCCL_FORCE_SYSTEM_HEADER_CLANG) \
+  || defined(_CCCL_FORCE_SYSTEM_HEADER_MSVC)
+#  error \
+    "This header must be included only within the <cuda/std/__cccl/system_header>. This most likely means a mix and match of different versions of CCCL."
+#endif // system header detected
+
+namespace __cccl_internal
+{
+#if _CCCL_CUDA_COMPILATION()
+template <class _Tp>
+__host__ __device__ _Tp&& __cccl_declval(int);
+template <class _Tp>
+__host__ __device__ _Tp __cccl_declval(long);
+template <class _Tp>
+__host__ __device__ decltype(__cccl_internal::__cccl_declval<_Tp>(0)) __cccl_declval() noexcept;
+
+// This requires a type to be implicitly convertible (also non-arithmetic)
+template <class _Tp>
+__host__ __device__ void __cccl_accepts_implicit_conversion(_Tp) noexcept;
+#else // ^^^ CUDA compilation ^^^ / vvv no CUDA compilation
+template <class _Tp>
+_Tp&& __cccl_declval(int);
+template <class _Tp>
+_Tp __cccl_declval(long);
+template <class _Tp>
+decltype(__cccl_internal::__cccl_declval<_Tp>(0)) __cccl_declval() noexcept;
+
+// This requires a type to be implicitly convertible (also non-arithmetic)
+template <class _Tp>
+void __cccl_accepts_implicit_conversion(_Tp) noexcept;
+#endif // no CUDA compilation
+
+template <class...>
+using __cccl_void_t = void;
+
+template <class _Dest, class _Source, class = void>
+struct __is_non_narrowing_convertible
+{
+  static constexpr bool value = false;
+};
+
+// This also prohibits narrowing conversion in case of arithmetic types
+template <class _Dest, class _Source>
+struct __is_non_narrowing_convertible<_Dest,
+                                      _Source,
+                                      __cccl_void_t<decltype(__cccl_internal::__cccl_accepts_implicit_conversion<_Dest>(
+                                                      __cccl_internal::__cccl_declval<_Source>())),
+                                                    decltype(_Dest{__cccl_internal::__cccl_declval<_Source>()})>>
+{
+  static constexpr bool value = true;
+};
+} // namespace __cccl_internal
+
+#endif // __CCCL_IS_NON_NARROWING_CONVERTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/os.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/os.h
new file mode 100644
index 00000000..2b7c8843
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/os.h
@@ -0,0 +1,120 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_OS_H
+#define __CCCL_OS_H
+
+// The header provides the following macros to determine the host architecture:
+//
+// _CCCL_OS(WINDOWS)
+// _CCCL_OS(LINUX)
+// _CCCL_OS(ANDROID)
+// _CCCL_OS(QNX)
+
+// Determine the host compiler and its version
+#if defined(_WIN32) || defined(_WIN64) /* _WIN64 for NVRTC */
+#  define _CCCL_OS_WINDOWS_() 1
+#else
+#  define _CCCL_OS_WINDOWS_() 0
+#endif
+
+#if defined(__linux__) || defined(__LP64__) /* __LP64__ for NVRTC */
+#  define _CCCL_OS_LINUX_() 1
+#else
+#  define _CCCL_OS_LINUX_() 0
+#endif
+
+#if defined(__ANDROID__)
+#  define _CCCL_OS_ANDROID_() 1
+#else
+#  define _CCCL_OS_ANDROID_() 0
+#endif
+
+#if defined(__QNX__) || defined(__QNXNTO__)
+#  define _CCCL_OS_QNX_() 1
+#else
+#  define _CCCL_OS_QNX_() 0
+#endif
+
+#if defined(__APPLE__) || defined(__APPLE_CC__)
+#  define _CCCL_OS_APPLE_() 1
+#else
+#  define _CCCL_OS_APPLE_() 0
+#endif
+
+#define _CCCL_OS(...) _CCCL_OS_##__VA_ARGS__##_()
+
+//! @def CCCL_OS(os) /* implementation defined */
+//!
+//! @brief Detect the current operating system.
+//!
+//! @param os The name of the operating system to test.
+//!
+//! @note This macro is made available when including any libcu++ header. Users that wish to
+//! include the smallest possible header for this macro should include `<cuda/std/version>`.
+//!
+//! For supported operating systems, the macro expands to an implementation-defined true value
+//! if the current operating system matches, or false otherwise. These values may be used in
+//! boolean expressions (preprocessor or otherwise), but no other guarantees are made.
+//!
+//! Available values for `os` include:
+//!
+//! - ``WINDOWS``: Windows, either in 32-bit or 64-bit mode.
+//! - ``LINUX``: Any kind of Linux installation. Note that other unix-based operating systems will
+//!              also match against this.
+//! - ``ANDROID``: Android operating system.
+//! - ``QNX``: QNX real-time operating system.
+//! - ``APPLE``: macOS (Intel or Apple Silicon).
+//!
+//! Passing any other value will result in an undefined expansion, which may or may not be
+//! diagnosed by the compiler.
+//!
+//! @note Some operating systems may satisfy multiple conditions. For example macOS and Android
+//! satisfy both `APPLE`/`ANDROID` and `LINUX`.
+//!
+//! @par Example
+//! @code
+//! #define MY_OTHER_MACRO 1
+//!
+//! // Expansion value can be used in ordinary macro conditionals
+//! #if CCCL_OS(WINDOWS) && MY_OTHER_MACRO
+//!   // ...
+//! #endif
+//!
+//! // Can be negated as usual
+//! #if !CCCL_OS(QNX)
+//!   // ...
+//! #endif
+//!
+//! #if CCCL_OS(APPLE)
+//!   // Will be visible only on macOS
+//! #endif
+//!
+//! #if CCCL_OS(ANDROID)
+//!   // Will be visible only on Android
+//! #endif
+//!
+//! #if CCCL_OS(LINUX) && !CCCL_OS(APPLE) && !CCCL_OS(ANDROID)
+//!   // Only visible on Linux
+//! #endif
+//! @endcode
+//!
+//! @return true if the specified OS is begin compiled for, false otherwise.
+#ifdef _CCCL_DOXYGEN_INVOKED
+#  define CCCL_OS(os) /* implementation defined */
+#else
+#  define CCCL_OS(__os__) _CCCL_OS_##__os__##_()
+#endif
+
+// Note: the public API is single-arg to constrain the API and allow for future expansion. The
+// implementation is duplicated to guard against the OS targets being accidentally defined by
+// the user.
+
+#endif // __CCCL_OS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/preprocessor.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/preprocessor.h
new file mode 100644
index 00000000..a8f6cbbf
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/preprocessor.h
@@ -0,0 +1,1366 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_PREPROCESSOR_H
+#define __CCCL_PREPROCESSOR_H
+
+// Error when MSVC is used with the traditional preprocessor.
+// We can't use `#pragma message` here because MSVC will encounter
+// errors and exit before it processes pragma message directives.
+#if defined(_MSC_VER) && !defined(__clang__)
+#  if (!defined(_MSVC_TRADITIONAL) || _MSVC_TRADITIONAL == 1) \
+    && !defined(CCCL_IGNORE_MSVC_TRADITIONAL_PREPROCESSOR_WARNING)
+#    error \
+MSVC/cl.exe with traditional preprocessor is used. This may lead to unexpected compilation errors. Please \
+switch to the standard conforming preprocessor by passing `/Zc:preprocessor` to cl.exe. You can define \
+CCCL_IGNORE_MSVC_TRADITIONAL_PREPROCESSOR_WARNING to suppress this warning.
+#  endif // !defined(_MSVC_TRADITIONAL) || _MSVC_TRADITIONAL == 1
+#endif // defined(_MSC_VER) && !defined(__clang__)
+
+#ifdef __COUNTER__
+#  define _CCCL_COUNTER() __COUNTER__
+#else
+#  define _CCCL_COUNTER() __LINE__
+#endif
+
+// Convert parameter to string
+#define _CCCL_TO_STRING2(_STR) #_STR
+#define _CCCL_TO_STRING(_STR)  _CCCL_TO_STRING2(_STR)
+
+#define _CCCL_PP_FIRST(_FIRST, ...)      _FIRST
+#define _CCCL_PP_SECOND(_, _SECOND, ...) _SECOND
+#define _CCCL_PP_THIRD(_1, _2, _THIRD)   _THIRD
+
+#define _CCCL_PP_EXPAND(...) __VA_ARGS__
+#define _CCCL_PP_EAT(...)
+
+#define _CCCL_PP_CAT_(_Xp, ...) _Xp##__VA_ARGS__
+#define _CCCL_PP_CAT(_Xp, ...)  _CCCL_PP_CAT_(_Xp, __VA_ARGS__)
+
+#define _CCCL_PP_CAT2_(_Xp, ...) _Xp##__VA_ARGS__
+#define _CCCL_PP_CAT2(_Xp, ...)  _CCCL_PP_CAT2_(_Xp, __VA_ARGS__)
+
+#define _CCCL_PP_CAT3_(_Xp, ...) _Xp##__VA_ARGS__
+#define _CCCL_PP_CAT3(_Xp, ...)  _CCCL_PP_CAT3_(_Xp, __VA_ARGS__)
+
+#define _CCCL_PP_CAT4_(_Xp, ...) _Xp##__VA_ARGS__
+#define _CCCL_PP_CAT4(_Xp, ...)  _CCCL_PP_CAT4_(_Xp, __VA_ARGS__)
+
+#define _CCCL_PP_EVAL_(_Xp, _ARGS) _Xp _ARGS
+#define _CCCL_PP_EVAL(_Xp, ...)    _CCCL_PP_EVAL_(_Xp, (__VA_ARGS__))
+
+#define _CCCL_PP_EVAL2_(_Xp, _ARGS) _Xp _ARGS
+#define _CCCL_PP_EVAL2(_Xp, ...)    _CCCL_PP_EVAL2_(_Xp, (__VA_ARGS__))
+
+#define _CCCL_PP_CHECK(...)              _CCCL_PP_EXPAND(_CCCL_PP_CHECK_N(__VA_ARGS__, 0, ))
+#define _CCCL_PP_CHECK_N(_Xp, _Num, ...) _Num
+#define _CCCL_PP_PROBE(_Xp)              _Xp, 1,
+#define _CCCL_PP_PROBE_N(_Xp, _Num)      _Xp, _Num,
+
+#define _CCCL_PP_IS_PAREN(_Xp)       _CCCL_PP_CHECK(_CCCL_PP_IS_PAREN_PROBE _Xp)
+#define _CCCL_PP_IS_PAREN_PROBE(...) _CCCL_PP_PROBE(~)
+
+#define _CCCL_PP_IIF(_BIT)         _CCCL_PP_CAT_(_CCCL_PP_IIF_, _BIT)
+#define _CCCL_PP_IIF_0(_TRUE, ...) __VA_ARGS__
+#define _CCCL_PP_IIF_1(_TRUE, ...) _TRUE
+
+#define _CCCL_PP_LPAREN (
+#define _CCCL_PP_RPAREN )
+
+#define _CCCL_PP_NOT(_BIT) _CCCL_PP_CAT_(_CCCL_PP_NOT_, _BIT)
+#define _CCCL_PP_NOT_0     1
+#define _CCCL_PP_NOT_1     0
+
+#define _CCCL_PP_EMPTY()
+#define _CCCL_PP_COMMA()        ,
+#define _CCCL_PP_LBRACE()       {
+#define _CCCL_PP_RBRACE()       }
+#define _CCCL_PP_COMMA_IIF(_Xp) _CCCL_PP_IIF(_Xp)(_CCCL_PP_COMMA, _CCCL_PP_EMPTY)()
+
+#define _CCCL_PP_CASE(_ARG) _CCCL_PP_PROBE_N(~, _ARG)
+#define _CCCL_PP_SWITCH(_PREFIX, ...) \
+  _CCCL_PP_CAT(_PREFIX##_CASE_, _CCCP_PP_CASE_LABEL_(_PREFIX, __VA_ARGS__))(__VA_ARGS__)
+#define _CCCL_PP_SWITCH2(_PREFIX, ...) \
+  _CCCL_PP_CAT(_PREFIX##_CASE_, _CCCP_PP_CASE_LABEL_(_PREFIX, __VA_ARGS__))(__VA_ARGS__)
+#define _CCCP_PP_CASE_LABEL_(_PREFIX, ...) \
+  _CCCL_PP_EVAL(_CCCL_PP_CHECK, _CCCL_PP_CAT(_PREFIX##_SWITCH_, _CCCL_PP_FIRST(__VA_ARGS__)), _CCCL_SWITCH_DEFAULT, )
+
+#define _CCCL_PP_FOR_EACH(_Mp, ...)                          _CCCL_PP_FOR_EACH_N(_CCCL_PP_COUNT(__VA_ARGS__), _Mp, __VA_ARGS__)
+#define _CCCL_PP_FOR_EACH_N(_Np, _Mp, ...)                   _CCCL_PP_CAT2(_CCCL_PP_FOR_EACH_, _Np)(_Mp, __VA_ARGS__)
+#define _CCCL_PP_FOR_EACH_1(_Mp, _1)                         _Mp(_1)
+#define _CCCL_PP_FOR_EACH_2(_Mp, _1, _2)                     _Mp(_1) _Mp(_2)
+#define _CCCL_PP_FOR_EACH_3(_Mp, _1, _2, _3)                 _Mp(_1) _Mp(_2) _Mp(_3)
+#define _CCCL_PP_FOR_EACH_4(_Mp, _1, _2, _3, _4)             _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4)
+#define _CCCL_PP_FOR_EACH_5(_Mp, _1, _2, _3, _4, _5)         _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5)
+#define _CCCL_PP_FOR_EACH_6(_Mp, _1, _2, _3, _4, _5, _6)     _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6)
+#define _CCCL_PP_FOR_EACH_7(_Mp, _1, _2, _3, _4, _5, _6, _7) _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7)
+#define _CCCL_PP_FOR_EACH_8(_Mp, _1, _2, _3, _4, _5, _6, _7, _8) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8)
+#define _CCCL_PP_FOR_EACH_9(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9)
+#define _CCCL_PP_FOR_EACH_10(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10)
+#define _CCCL_PP_FOR_EACH_11(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11)
+#define _CCCL_PP_FOR_EACH_12(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12)
+#define _CCCL_PP_FOR_EACH_13(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13)
+#define _CCCL_PP_FOR_EACH_14(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14)
+#define _CCCL_PP_FOR_EACH_15(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15)                    \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14) \
+    _Mp(_15)
+#define _CCCL_PP_FOR_EACH_16(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16)               \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14) \
+    _Mp(_15) _Mp(_16)
+#define _CCCL_PP_FOR_EACH_17(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17)          \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14) \
+    _Mp(_15) _Mp(_16) _Mp(_17)
+#define _CCCL_PP_FOR_EACH_18(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18)     \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14) \
+    _Mp(_15) _Mp(_16) _Mp(_17) _Mp(_18)
+#define _CCCL_PP_FOR_EACH_19(_Mp, _1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19) \
+  _Mp(_1) _Mp(_2) _Mp(_3) _Mp(_4) _Mp(_5) _Mp(_6) _Mp(_7) _Mp(_8) _Mp(_9) _Mp(_10) _Mp(_11) _Mp(_12) _Mp(_13) _Mp(_14)  \
+    _Mp(_15) _Mp(_16) _Mp(_17) _Mp(_18) _Mp(_19)
+
+#define _CCCL_PP_PROBE_EMPTY_PROBE__CCCL_PP_PROBE_EMPTY _CCCL_PP_PROBE(~)
+
+#define _CCCL_PP_PROBE_EMPTY()
+#define _CCCL_PP_IS_NOT_EMPTY(...)                                                                             \
+  _CCCL_PP_EVAL(_CCCL_PP_CHECK, _CCCL_PP_CAT(_CCCL_PP_PROBE_EMPTY_PROBE_, _CCCL_PP_PROBE_EMPTY __VA_ARGS__())) \
+  /**/
+
+#define _CCCL_PP_TAIL(_, ...) __VA_ARGS__
+
+///////////////////////////////////////////////////////////////////////////////
+
+// Count the number of arguments. There must be at least one argument and fewer
+// than 126 arguments.
+// clang-format off
+#define _CCCL_PP_COUNT_IMPL(                                                                      \
+  _125, _124, _123, _122, _121, _120, _119, _118, _117, _116, _115, _114, _113, _112, _111, _110, \
+  _109, _108, _107, _106, _105, _104, _103, _102, _101, _100, _99, _98, _97, _96, _95, _94,       \
+  _93, _92, _91, _90, _89, _88, _87, _86, _85, _84, _83, _82, _81, _80, _79, _78,                 \
+  _77, _76, _75, _74, _73, _72, _71, _70, _69, _68, _67, _66, _65, _64, _63, _62,                 \
+  _61, _60, _59, _58, _57, _56, _55, _54, _53, _52, _51, _50, _49, _48, _47, _46,                 \
+  _45, _44, _43, _42, _41, _40, _39, _38, _37, _36, _35, _34, _33, _32, _31, _30,                 \
+  _29, _28, _27, _26, _25, _24, _23, _22, _21, _20, _19, _18, _17, _16, _15, _14,                 \
+  _13, _12, _11, _10, _9, _8, _7, _6, _5, _4, _3, _2, _1, _0, ...) _0
+
+#define _CCCL_PP_COUNT(...)                                                         \
+  _CCCL_PP_EXPAND(_CCCL_PP_COUNT_IMPL( __VA_ARGS__,                                 \
+    125, 124, 123, 122, 121, 120, 119, 118, 117, 116, 115, 114, 113, 112, 111, 110, \
+    109, 108, 107, 106, 105, 104, 103, 102, 101, 100, 99, 98, 97, 96, 95, 94,       \
+    93, 92, 91, 90, 89, 88, 87, 86, 85, 84, 83, 82, 81, 80, 79, 78,                 \
+    77, 76, 75, 74, 73, 72, 71, 70, 69, 68, 67, 66, 65, 64, 63, 62,                 \
+    61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46,                 \
+    45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30,                 \
+    29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14,                 \
+    13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0))
+// clang-format on
+
+///////////////////////////////////////////////////////////////////////////////
+
+#define _CCCL_PP_INC(_X) _CCCL_PP_INC_IMPL0(_X)
+
+#define _CCCL_PP_INC_IMPL0(_X) _CCCL_PP_CAT(_CCCL_PP_INC_IMPL_TAG, _X)
+
+#define _CCCL_PP_INC_IMPL_TAG0   1
+#define _CCCL_PP_INC_IMPL_TAG1   2
+#define _CCCL_PP_INC_IMPL_TAG2   3
+#define _CCCL_PP_INC_IMPL_TAG3   4
+#define _CCCL_PP_INC_IMPL_TAG4   5
+#define _CCCL_PP_INC_IMPL_TAG5   6
+#define _CCCL_PP_INC_IMPL_TAG6   7
+#define _CCCL_PP_INC_IMPL_TAG7   8
+#define _CCCL_PP_INC_IMPL_TAG8   9
+#define _CCCL_PP_INC_IMPL_TAG9   10
+#define _CCCL_PP_INC_IMPL_TAG10  11
+#define _CCCL_PP_INC_IMPL_TAG11  12
+#define _CCCL_PP_INC_IMPL_TAG12  13
+#define _CCCL_PP_INC_IMPL_TAG13  14
+#define _CCCL_PP_INC_IMPL_TAG14  15
+#define _CCCL_PP_INC_IMPL_TAG15  16
+#define _CCCL_PP_INC_IMPL_TAG16  17
+#define _CCCL_PP_INC_IMPL_TAG17  18
+#define _CCCL_PP_INC_IMPL_TAG18  19
+#define _CCCL_PP_INC_IMPL_TAG19  20
+#define _CCCL_PP_INC_IMPL_TAG20  21
+#define _CCCL_PP_INC_IMPL_TAG21  22
+#define _CCCL_PP_INC_IMPL_TAG22  23
+#define _CCCL_PP_INC_IMPL_TAG23  24
+#define _CCCL_PP_INC_IMPL_TAG24  25
+#define _CCCL_PP_INC_IMPL_TAG25  26
+#define _CCCL_PP_INC_IMPL_TAG26  27
+#define _CCCL_PP_INC_IMPL_TAG27  28
+#define _CCCL_PP_INC_IMPL_TAG28  29
+#define _CCCL_PP_INC_IMPL_TAG29  30
+#define _CCCL_PP_INC_IMPL_TAG30  31
+#define _CCCL_PP_INC_IMPL_TAG31  32
+#define _CCCL_PP_INC_IMPL_TAG32  33
+#define _CCCL_PP_INC_IMPL_TAG33  34
+#define _CCCL_PP_INC_IMPL_TAG34  35
+#define _CCCL_PP_INC_IMPL_TAG35  36
+#define _CCCL_PP_INC_IMPL_TAG36  37
+#define _CCCL_PP_INC_IMPL_TAG37  38
+#define _CCCL_PP_INC_IMPL_TAG38  39
+#define _CCCL_PP_INC_IMPL_TAG39  40
+#define _CCCL_PP_INC_IMPL_TAG40  41
+#define _CCCL_PP_INC_IMPL_TAG41  42
+#define _CCCL_PP_INC_IMPL_TAG42  43
+#define _CCCL_PP_INC_IMPL_TAG43  44
+#define _CCCL_PP_INC_IMPL_TAG44  45
+#define _CCCL_PP_INC_IMPL_TAG45  46
+#define _CCCL_PP_INC_IMPL_TAG46  47
+#define _CCCL_PP_INC_IMPL_TAG47  48
+#define _CCCL_PP_INC_IMPL_TAG48  49
+#define _CCCL_PP_INC_IMPL_TAG49  50
+#define _CCCL_PP_INC_IMPL_TAG50  51
+#define _CCCL_PP_INC_IMPL_TAG51  52
+#define _CCCL_PP_INC_IMPL_TAG52  53
+#define _CCCL_PP_INC_IMPL_TAG53  54
+#define _CCCL_PP_INC_IMPL_TAG54  55
+#define _CCCL_PP_INC_IMPL_TAG55  56
+#define _CCCL_PP_INC_IMPL_TAG56  57
+#define _CCCL_PP_INC_IMPL_TAG57  58
+#define _CCCL_PP_INC_IMPL_TAG58  59
+#define _CCCL_PP_INC_IMPL_TAG59  60
+#define _CCCL_PP_INC_IMPL_TAG60  61
+#define _CCCL_PP_INC_IMPL_TAG61  62
+#define _CCCL_PP_INC_IMPL_TAG62  63
+#define _CCCL_PP_INC_IMPL_TAG63  64
+#define _CCCL_PP_INC_IMPL_TAG64  65
+#define _CCCL_PP_INC_IMPL_TAG65  66
+#define _CCCL_PP_INC_IMPL_TAG66  67
+#define _CCCL_PP_INC_IMPL_TAG67  68
+#define _CCCL_PP_INC_IMPL_TAG68  69
+#define _CCCL_PP_INC_IMPL_TAG69  70
+#define _CCCL_PP_INC_IMPL_TAG70  71
+#define _CCCL_PP_INC_IMPL_TAG71  72
+#define _CCCL_PP_INC_IMPL_TAG72  73
+#define _CCCL_PP_INC_IMPL_TAG73  74
+#define _CCCL_PP_INC_IMPL_TAG74  75
+#define _CCCL_PP_INC_IMPL_TAG75  76
+#define _CCCL_PP_INC_IMPL_TAG76  77
+#define _CCCL_PP_INC_IMPL_TAG77  78
+#define _CCCL_PP_INC_IMPL_TAG78  79
+#define _CCCL_PP_INC_IMPL_TAG79  80
+#define _CCCL_PP_INC_IMPL_TAG80  81
+#define _CCCL_PP_INC_IMPL_TAG81  82
+#define _CCCL_PP_INC_IMPL_TAG82  83
+#define _CCCL_PP_INC_IMPL_TAG83  84
+#define _CCCL_PP_INC_IMPL_TAG84  85
+#define _CCCL_PP_INC_IMPL_TAG85  86
+#define _CCCL_PP_INC_IMPL_TAG86  87
+#define _CCCL_PP_INC_IMPL_TAG87  88
+#define _CCCL_PP_INC_IMPL_TAG88  89
+#define _CCCL_PP_INC_IMPL_TAG89  90
+#define _CCCL_PP_INC_IMPL_TAG90  91
+#define _CCCL_PP_INC_IMPL_TAG91  92
+#define _CCCL_PP_INC_IMPL_TAG92  93
+#define _CCCL_PP_INC_IMPL_TAG93  94
+#define _CCCL_PP_INC_IMPL_TAG94  95
+#define _CCCL_PP_INC_IMPL_TAG95  96
+#define _CCCL_PP_INC_IMPL_TAG96  97
+#define _CCCL_PP_INC_IMPL_TAG97  98
+#define _CCCL_PP_INC_IMPL_TAG98  99
+#define _CCCL_PP_INC_IMPL_TAG99  100
+#define _CCCL_PP_INC_IMPL_TAG100 101
+#define _CCCL_PP_INC_IMPL_TAG101 102
+#define _CCCL_PP_INC_IMPL_TAG102 103
+#define _CCCL_PP_INC_IMPL_TAG103 104
+#define _CCCL_PP_INC_IMPL_TAG104 105
+#define _CCCL_PP_INC_IMPL_TAG105 106
+#define _CCCL_PP_INC_IMPL_TAG106 107
+#define _CCCL_PP_INC_IMPL_TAG107 108
+#define _CCCL_PP_INC_IMPL_TAG108 109
+#define _CCCL_PP_INC_IMPL_TAG109 110
+#define _CCCL_PP_INC_IMPL_TAG110 111
+#define _CCCL_PP_INC_IMPL_TAG111 112
+#define _CCCL_PP_INC_IMPL_TAG112 113
+#define _CCCL_PP_INC_IMPL_TAG113 114
+#define _CCCL_PP_INC_IMPL_TAG114 115
+#define _CCCL_PP_INC_IMPL_TAG115 116
+#define _CCCL_PP_INC_IMPL_TAG116 117
+#define _CCCL_PP_INC_IMPL_TAG117 118
+#define _CCCL_PP_INC_IMPL_TAG118 119
+#define _CCCL_PP_INC_IMPL_TAG119 120
+#define _CCCL_PP_INC_IMPL_TAG120 121
+#define _CCCL_PP_INC_IMPL_TAG121 122
+#define _CCCL_PP_INC_IMPL_TAG122 123
+#define _CCCL_PP_INC_IMPL_TAG123 124
+#define _CCCL_PP_INC_IMPL_TAG124 125
+#define _CCCL_PP_INC_IMPL_TAG125 126
+#define _CCCL_PP_INC_IMPL_TAG126 127
+#define _CCCL_PP_INC_IMPL_TAG127 128
+#define _CCCL_PP_INC_IMPL_TAG128 129
+#define _CCCL_PP_INC_IMPL_TAG129 130
+#define _CCCL_PP_INC_IMPL_TAG130 131
+#define _CCCL_PP_INC_IMPL_TAG131 132
+#define _CCCL_PP_INC_IMPL_TAG132 133
+#define _CCCL_PP_INC_IMPL_TAG133 134
+#define _CCCL_PP_INC_IMPL_TAG134 135
+#define _CCCL_PP_INC_IMPL_TAG135 136
+#define _CCCL_PP_INC_IMPL_TAG136 137
+#define _CCCL_PP_INC_IMPL_TAG137 138
+#define _CCCL_PP_INC_IMPL_TAG138 139
+#define _CCCL_PP_INC_IMPL_TAG139 140
+#define _CCCL_PP_INC_IMPL_TAG140 141
+#define _CCCL_PP_INC_IMPL_TAG141 142
+#define _CCCL_PP_INC_IMPL_TAG142 143
+#define _CCCL_PP_INC_IMPL_TAG143 144
+#define _CCCL_PP_INC_IMPL_TAG144 145
+#define _CCCL_PP_INC_IMPL_TAG145 146
+#define _CCCL_PP_INC_IMPL_TAG146 147
+#define _CCCL_PP_INC_IMPL_TAG147 148
+#define _CCCL_PP_INC_IMPL_TAG148 149
+#define _CCCL_PP_INC_IMPL_TAG149 150
+#define _CCCL_PP_INC_IMPL_TAG150 151
+#define _CCCL_PP_INC_IMPL_TAG151 152
+#define _CCCL_PP_INC_IMPL_TAG152 153
+#define _CCCL_PP_INC_IMPL_TAG153 154
+#define _CCCL_PP_INC_IMPL_TAG154 155
+#define _CCCL_PP_INC_IMPL_TAG155 156
+#define _CCCL_PP_INC_IMPL_TAG156 157
+#define _CCCL_PP_INC_IMPL_TAG157 158
+#define _CCCL_PP_INC_IMPL_TAG158 159
+#define _CCCL_PP_INC_IMPL_TAG159 160
+#define _CCCL_PP_INC_IMPL_TAG160 161
+#define _CCCL_PP_INC_IMPL_TAG161 162
+#define _CCCL_PP_INC_IMPL_TAG162 163
+#define _CCCL_PP_INC_IMPL_TAG163 164
+#define _CCCL_PP_INC_IMPL_TAG164 165
+#define _CCCL_PP_INC_IMPL_TAG165 166
+#define _CCCL_PP_INC_IMPL_TAG166 167
+#define _CCCL_PP_INC_IMPL_TAG167 168
+#define _CCCL_PP_INC_IMPL_TAG168 169
+#define _CCCL_PP_INC_IMPL_TAG169 170
+#define _CCCL_PP_INC_IMPL_TAG170 171
+#define _CCCL_PP_INC_IMPL_TAG171 172
+#define _CCCL_PP_INC_IMPL_TAG172 173
+#define _CCCL_PP_INC_IMPL_TAG173 174
+#define _CCCL_PP_INC_IMPL_TAG174 175
+#define _CCCL_PP_INC_IMPL_TAG175 176
+#define _CCCL_PP_INC_IMPL_TAG176 177
+#define _CCCL_PP_INC_IMPL_TAG177 178
+#define _CCCL_PP_INC_IMPL_TAG178 179
+#define _CCCL_PP_INC_IMPL_TAG179 180
+#define _CCCL_PP_INC_IMPL_TAG180 181
+#define _CCCL_PP_INC_IMPL_TAG181 182
+#define _CCCL_PP_INC_IMPL_TAG182 183
+#define _CCCL_PP_INC_IMPL_TAG183 184
+#define _CCCL_PP_INC_IMPL_TAG184 185
+#define _CCCL_PP_INC_IMPL_TAG185 186
+#define _CCCL_PP_INC_IMPL_TAG186 187
+#define _CCCL_PP_INC_IMPL_TAG187 188
+#define _CCCL_PP_INC_IMPL_TAG188 189
+#define _CCCL_PP_INC_IMPL_TAG189 190
+#define _CCCL_PP_INC_IMPL_TAG190 191
+#define _CCCL_PP_INC_IMPL_TAG191 192
+#define _CCCL_PP_INC_IMPL_TAG192 193
+#define _CCCL_PP_INC_IMPL_TAG193 194
+#define _CCCL_PP_INC_IMPL_TAG194 195
+#define _CCCL_PP_INC_IMPL_TAG195 196
+#define _CCCL_PP_INC_IMPL_TAG196 197
+#define _CCCL_PP_INC_IMPL_TAG197 198
+#define _CCCL_PP_INC_IMPL_TAG198 199
+#define _CCCL_PP_INC_IMPL_TAG199 200
+#define _CCCL_PP_INC_IMPL_TAG200 201
+#define _CCCL_PP_INC_IMPL_TAG201 202
+#define _CCCL_PP_INC_IMPL_TAG202 203
+#define _CCCL_PP_INC_IMPL_TAG203 204
+#define _CCCL_PP_INC_IMPL_TAG204 205
+#define _CCCL_PP_INC_IMPL_TAG205 206
+#define _CCCL_PP_INC_IMPL_TAG206 207
+#define _CCCL_PP_INC_IMPL_TAG207 208
+#define _CCCL_PP_INC_IMPL_TAG208 209
+#define _CCCL_PP_INC_IMPL_TAG209 210
+#define _CCCL_PP_INC_IMPL_TAG210 211
+#define _CCCL_PP_INC_IMPL_TAG211 212
+#define _CCCL_PP_INC_IMPL_TAG212 213
+#define _CCCL_PP_INC_IMPL_TAG213 214
+#define _CCCL_PP_INC_IMPL_TAG214 215
+#define _CCCL_PP_INC_IMPL_TAG215 216
+#define _CCCL_PP_INC_IMPL_TAG216 217
+#define _CCCL_PP_INC_IMPL_TAG217 218
+#define _CCCL_PP_INC_IMPL_TAG218 219
+#define _CCCL_PP_INC_IMPL_TAG219 220
+#define _CCCL_PP_INC_IMPL_TAG220 221
+#define _CCCL_PP_INC_IMPL_TAG221 222
+#define _CCCL_PP_INC_IMPL_TAG222 223
+#define _CCCL_PP_INC_IMPL_TAG223 224
+#define _CCCL_PP_INC_IMPL_TAG224 225
+#define _CCCL_PP_INC_IMPL_TAG225 226
+#define _CCCL_PP_INC_IMPL_TAG226 227
+#define _CCCL_PP_INC_IMPL_TAG227 228
+#define _CCCL_PP_INC_IMPL_TAG228 229
+#define _CCCL_PP_INC_IMPL_TAG229 230
+#define _CCCL_PP_INC_IMPL_TAG230 231
+#define _CCCL_PP_INC_IMPL_TAG231 232
+#define _CCCL_PP_INC_IMPL_TAG232 233
+#define _CCCL_PP_INC_IMPL_TAG233 234
+#define _CCCL_PP_INC_IMPL_TAG234 235
+#define _CCCL_PP_INC_IMPL_TAG235 236
+#define _CCCL_PP_INC_IMPL_TAG236 237
+#define _CCCL_PP_INC_IMPL_TAG237 238
+#define _CCCL_PP_INC_IMPL_TAG238 239
+#define _CCCL_PP_INC_IMPL_TAG239 240
+#define _CCCL_PP_INC_IMPL_TAG240 241
+#define _CCCL_PP_INC_IMPL_TAG241 242
+#define _CCCL_PP_INC_IMPL_TAG242 243
+#define _CCCL_PP_INC_IMPL_TAG243 244
+#define _CCCL_PP_INC_IMPL_TAG244 245
+#define _CCCL_PP_INC_IMPL_TAG245 246
+#define _CCCL_PP_INC_IMPL_TAG246 247
+#define _CCCL_PP_INC_IMPL_TAG247 248
+#define _CCCL_PP_INC_IMPL_TAG248 249
+#define _CCCL_PP_INC_IMPL_TAG249 250
+#define _CCCL_PP_INC_IMPL_TAG250 251
+#define _CCCL_PP_INC_IMPL_TAG251 252
+#define _CCCL_PP_INC_IMPL_TAG252 253
+#define _CCCL_PP_INC_IMPL_TAG253 254
+#define _CCCL_PP_INC_IMPL_TAG254 255
+#define _CCCL_PP_INC_IMPL_TAG255 256
+#define _CCCL_PP_INC_IMPL_TAG256 257
+
+#define _CCCL_PP_DEC(_X) _CCCL_PP_DEC_IMPL0(_X)
+
+#define _CCCL_PP_DEC_IMPL0(_X) _CCCL_PP_CAT(_CCCL_PP_DEC_IMPL_TAG, _X)
+
+#define _CCCL_PP_DEC_IMPL_TAG0   ~##~ // This will generate a syntax error
+#define _CCCL_PP_DEC_IMPL_TAG1   0
+#define _CCCL_PP_DEC_IMPL_TAG2   1
+#define _CCCL_PP_DEC_IMPL_TAG3   2
+#define _CCCL_PP_DEC_IMPL_TAG4   3
+#define _CCCL_PP_DEC_IMPL_TAG5   4
+#define _CCCL_PP_DEC_IMPL_TAG6   5
+#define _CCCL_PP_DEC_IMPL_TAG7   6
+#define _CCCL_PP_DEC_IMPL_TAG8   7
+#define _CCCL_PP_DEC_IMPL_TAG9   8
+#define _CCCL_PP_DEC_IMPL_TAG10  9
+#define _CCCL_PP_DEC_IMPL_TAG11  10
+#define _CCCL_PP_DEC_IMPL_TAG12  11
+#define _CCCL_PP_DEC_IMPL_TAG13  12
+#define _CCCL_PP_DEC_IMPL_TAG14  13
+#define _CCCL_PP_DEC_IMPL_TAG15  14
+#define _CCCL_PP_DEC_IMPL_TAG16  15
+#define _CCCL_PP_DEC_IMPL_TAG17  16
+#define _CCCL_PP_DEC_IMPL_TAG18  17
+#define _CCCL_PP_DEC_IMPL_TAG19  18
+#define _CCCL_PP_DEC_IMPL_TAG20  19
+#define _CCCL_PP_DEC_IMPL_TAG21  20
+#define _CCCL_PP_DEC_IMPL_TAG22  21
+#define _CCCL_PP_DEC_IMPL_TAG23  22
+#define _CCCL_PP_DEC_IMPL_TAG24  23
+#define _CCCL_PP_DEC_IMPL_TAG25  24
+#define _CCCL_PP_DEC_IMPL_TAG26  25
+#define _CCCL_PP_DEC_IMPL_TAG27  26
+#define _CCCL_PP_DEC_IMPL_TAG28  27
+#define _CCCL_PP_DEC_IMPL_TAG29  28
+#define _CCCL_PP_DEC_IMPL_TAG30  29
+#define _CCCL_PP_DEC_IMPL_TAG31  30
+#define _CCCL_PP_DEC_IMPL_TAG32  31
+#define _CCCL_PP_DEC_IMPL_TAG33  32
+#define _CCCL_PP_DEC_IMPL_TAG34  33
+#define _CCCL_PP_DEC_IMPL_TAG35  34
+#define _CCCL_PP_DEC_IMPL_TAG36  35
+#define _CCCL_PP_DEC_IMPL_TAG37  36
+#define _CCCL_PP_DEC_IMPL_TAG38  37
+#define _CCCL_PP_DEC_IMPL_TAG39  38
+#define _CCCL_PP_DEC_IMPL_TAG40  39
+#define _CCCL_PP_DEC_IMPL_TAG41  40
+#define _CCCL_PP_DEC_IMPL_TAG42  41
+#define _CCCL_PP_DEC_IMPL_TAG43  42
+#define _CCCL_PP_DEC_IMPL_TAG44  43
+#define _CCCL_PP_DEC_IMPL_TAG45  44
+#define _CCCL_PP_DEC_IMPL_TAG46  45
+#define _CCCL_PP_DEC_IMPL_TAG47  46
+#define _CCCL_PP_DEC_IMPL_TAG48  47
+#define _CCCL_PP_DEC_IMPL_TAG49  48
+#define _CCCL_PP_DEC_IMPL_TAG50  49
+#define _CCCL_PP_DEC_IMPL_TAG51  50
+#define _CCCL_PP_DEC_IMPL_TAG52  51
+#define _CCCL_PP_DEC_IMPL_TAG53  52
+#define _CCCL_PP_DEC_IMPL_TAG54  53
+#define _CCCL_PP_DEC_IMPL_TAG55  54
+#define _CCCL_PP_DEC_IMPL_TAG56  55
+#define _CCCL_PP_DEC_IMPL_TAG57  56
+#define _CCCL_PP_DEC_IMPL_TAG58  57
+#define _CCCL_PP_DEC_IMPL_TAG59  58
+#define _CCCL_PP_DEC_IMPL_TAG60  59
+#define _CCCL_PP_DEC_IMPL_TAG61  60
+#define _CCCL_PP_DEC_IMPL_TAG62  61
+#define _CCCL_PP_DEC_IMPL_TAG63  62
+#define _CCCL_PP_DEC_IMPL_TAG64  63
+#define _CCCL_PP_DEC_IMPL_TAG65  64
+#define _CCCL_PP_DEC_IMPL_TAG66  65
+#define _CCCL_PP_DEC_IMPL_TAG67  66
+#define _CCCL_PP_DEC_IMPL_TAG68  67
+#define _CCCL_PP_DEC_IMPL_TAG69  68
+#define _CCCL_PP_DEC_IMPL_TAG70  69
+#define _CCCL_PP_DEC_IMPL_TAG71  70
+#define _CCCL_PP_DEC_IMPL_TAG72  71
+#define _CCCL_PP_DEC_IMPL_TAG73  72
+#define _CCCL_PP_DEC_IMPL_TAG74  73
+#define _CCCL_PP_DEC_IMPL_TAG75  74
+#define _CCCL_PP_DEC_IMPL_TAG76  75
+#define _CCCL_PP_DEC_IMPL_TAG77  76
+#define _CCCL_PP_DEC_IMPL_TAG78  77
+#define _CCCL_PP_DEC_IMPL_TAG79  78
+#define _CCCL_PP_DEC_IMPL_TAG80  79
+#define _CCCL_PP_DEC_IMPL_TAG81  80
+#define _CCCL_PP_DEC_IMPL_TAG82  81
+#define _CCCL_PP_DEC_IMPL_TAG83  82
+#define _CCCL_PP_DEC_IMPL_TAG84  83
+#define _CCCL_PP_DEC_IMPL_TAG85  84
+#define _CCCL_PP_DEC_IMPL_TAG86  85
+#define _CCCL_PP_DEC_IMPL_TAG87  86
+#define _CCCL_PP_DEC_IMPL_TAG88  87
+#define _CCCL_PP_DEC_IMPL_TAG89  88
+#define _CCCL_PP_DEC_IMPL_TAG90  89
+#define _CCCL_PP_DEC_IMPL_TAG91  90
+#define _CCCL_PP_DEC_IMPL_TAG92  91
+#define _CCCL_PP_DEC_IMPL_TAG93  92
+#define _CCCL_PP_DEC_IMPL_TAG94  93
+#define _CCCL_PP_DEC_IMPL_TAG95  94
+#define _CCCL_PP_DEC_IMPL_TAG96  95
+#define _CCCL_PP_DEC_IMPL_TAG97  96
+#define _CCCL_PP_DEC_IMPL_TAG98  97
+#define _CCCL_PP_DEC_IMPL_TAG99  98
+#define _CCCL_PP_DEC_IMPL_TAG100 99
+#define _CCCL_PP_DEC_IMPL_TAG101 100
+#define _CCCL_PP_DEC_IMPL_TAG102 101
+#define _CCCL_PP_DEC_IMPL_TAG103 102
+#define _CCCL_PP_DEC_IMPL_TAG104 103
+#define _CCCL_PP_DEC_IMPL_TAG105 104
+#define _CCCL_PP_DEC_IMPL_TAG106 105
+#define _CCCL_PP_DEC_IMPL_TAG107 106
+#define _CCCL_PP_DEC_IMPL_TAG108 107
+#define _CCCL_PP_DEC_IMPL_TAG109 108
+#define _CCCL_PP_DEC_IMPL_TAG110 109
+#define _CCCL_PP_DEC_IMPL_TAG111 110
+#define _CCCL_PP_DEC_IMPL_TAG112 111
+#define _CCCL_PP_DEC_IMPL_TAG113 112
+#define _CCCL_PP_DEC_IMPL_TAG114 113
+#define _CCCL_PP_DEC_IMPL_TAG115 114
+#define _CCCL_PP_DEC_IMPL_TAG116 115
+#define _CCCL_PP_DEC_IMPL_TAG117 116
+#define _CCCL_PP_DEC_IMPL_TAG118 117
+#define _CCCL_PP_DEC_IMPL_TAG119 118
+#define _CCCL_PP_DEC_IMPL_TAG120 119
+#define _CCCL_PP_DEC_IMPL_TAG121 120
+#define _CCCL_PP_DEC_IMPL_TAG122 121
+#define _CCCL_PP_DEC_IMPL_TAG123 122
+#define _CCCL_PP_DEC_IMPL_TAG124 123
+#define _CCCL_PP_DEC_IMPL_TAG125 124
+#define _CCCL_PP_DEC_IMPL_TAG126 125
+#define _CCCL_PP_DEC_IMPL_TAG127 126
+#define _CCCL_PP_DEC_IMPL_TAG128 127
+#define _CCCL_PP_DEC_IMPL_TAG129 128
+#define _CCCL_PP_DEC_IMPL_TAG130 129
+#define _CCCL_PP_DEC_IMPL_TAG131 130
+#define _CCCL_PP_DEC_IMPL_TAG132 131
+#define _CCCL_PP_DEC_IMPL_TAG133 132
+#define _CCCL_PP_DEC_IMPL_TAG134 133
+#define _CCCL_PP_DEC_IMPL_TAG135 134
+#define _CCCL_PP_DEC_IMPL_TAG136 135
+#define _CCCL_PP_DEC_IMPL_TAG137 136
+#define _CCCL_PP_DEC_IMPL_TAG138 137
+#define _CCCL_PP_DEC_IMPL_TAG139 138
+#define _CCCL_PP_DEC_IMPL_TAG140 139
+#define _CCCL_PP_DEC_IMPL_TAG141 140
+#define _CCCL_PP_DEC_IMPL_TAG142 141
+#define _CCCL_PP_DEC_IMPL_TAG143 142
+#define _CCCL_PP_DEC_IMPL_TAG144 143
+#define _CCCL_PP_DEC_IMPL_TAG145 144
+#define _CCCL_PP_DEC_IMPL_TAG146 145
+#define _CCCL_PP_DEC_IMPL_TAG147 146
+#define _CCCL_PP_DEC_IMPL_TAG148 147
+#define _CCCL_PP_DEC_IMPL_TAG149 148
+#define _CCCL_PP_DEC_IMPL_TAG150 149
+#define _CCCL_PP_DEC_IMPL_TAG151 150
+#define _CCCL_PP_DEC_IMPL_TAG152 151
+#define _CCCL_PP_DEC_IMPL_TAG153 152
+#define _CCCL_PP_DEC_IMPL_TAG154 153
+#define _CCCL_PP_DEC_IMPL_TAG155 154
+#define _CCCL_PP_DEC_IMPL_TAG156 155
+#define _CCCL_PP_DEC_IMPL_TAG157 156
+#define _CCCL_PP_DEC_IMPL_TAG158 157
+#define _CCCL_PP_DEC_IMPL_TAG159 158
+#define _CCCL_PP_DEC_IMPL_TAG160 159
+#define _CCCL_PP_DEC_IMPL_TAG161 160
+#define _CCCL_PP_DEC_IMPL_TAG162 161
+#define _CCCL_PP_DEC_IMPL_TAG163 162
+#define _CCCL_PP_DEC_IMPL_TAG164 163
+#define _CCCL_PP_DEC_IMPL_TAG165 164
+#define _CCCL_PP_DEC_IMPL_TAG166 165
+#define _CCCL_PP_DEC_IMPL_TAG167 166
+#define _CCCL_PP_DEC_IMPL_TAG168 167
+#define _CCCL_PP_DEC_IMPL_TAG169 168
+#define _CCCL_PP_DEC_IMPL_TAG170 169
+#define _CCCL_PP_DEC_IMPL_TAG171 170
+#define _CCCL_PP_DEC_IMPL_TAG172 171
+#define _CCCL_PP_DEC_IMPL_TAG173 172
+#define _CCCL_PP_DEC_IMPL_TAG174 173
+#define _CCCL_PP_DEC_IMPL_TAG175 174
+#define _CCCL_PP_DEC_IMPL_TAG176 175
+#define _CCCL_PP_DEC_IMPL_TAG177 176
+#define _CCCL_PP_DEC_IMPL_TAG178 177
+#define _CCCL_PP_DEC_IMPL_TAG179 178
+#define _CCCL_PP_DEC_IMPL_TAG180 179
+#define _CCCL_PP_DEC_IMPL_TAG181 180
+#define _CCCL_PP_DEC_IMPL_TAG182 181
+#define _CCCL_PP_DEC_IMPL_TAG183 182
+#define _CCCL_PP_DEC_IMPL_TAG184 183
+#define _CCCL_PP_DEC_IMPL_TAG185 184
+#define _CCCL_PP_DEC_IMPL_TAG186 185
+#define _CCCL_PP_DEC_IMPL_TAG187 186
+#define _CCCL_PP_DEC_IMPL_TAG188 187
+#define _CCCL_PP_DEC_IMPL_TAG189 188
+#define _CCCL_PP_DEC_IMPL_TAG190 189
+#define _CCCL_PP_DEC_IMPL_TAG191 190
+#define _CCCL_PP_DEC_IMPL_TAG192 191
+#define _CCCL_PP_DEC_IMPL_TAG193 192
+#define _CCCL_PP_DEC_IMPL_TAG194 193
+#define _CCCL_PP_DEC_IMPL_TAG195 194
+#define _CCCL_PP_DEC_IMPL_TAG196 195
+#define _CCCL_PP_DEC_IMPL_TAG197 196
+#define _CCCL_PP_DEC_IMPL_TAG198 197
+#define _CCCL_PP_DEC_IMPL_TAG199 198
+#define _CCCL_PP_DEC_IMPL_TAG200 199
+#define _CCCL_PP_DEC_IMPL_TAG201 200
+#define _CCCL_PP_DEC_IMPL_TAG202 201
+#define _CCCL_PP_DEC_IMPL_TAG203 202
+#define _CCCL_PP_DEC_IMPL_TAG204 203
+#define _CCCL_PP_DEC_IMPL_TAG205 204
+#define _CCCL_PP_DEC_IMPL_TAG206 205
+#define _CCCL_PP_DEC_IMPL_TAG207 206
+#define _CCCL_PP_DEC_IMPL_TAG208 207
+#define _CCCL_PP_DEC_IMPL_TAG209 208
+#define _CCCL_PP_DEC_IMPL_TAG210 209
+#define _CCCL_PP_DEC_IMPL_TAG211 210
+#define _CCCL_PP_DEC_IMPL_TAG212 211
+#define _CCCL_PP_DEC_IMPL_TAG213 212
+#define _CCCL_PP_DEC_IMPL_TAG214 213
+#define _CCCL_PP_DEC_IMPL_TAG215 214
+#define _CCCL_PP_DEC_IMPL_TAG216 215
+#define _CCCL_PP_DEC_IMPL_TAG217 216
+#define _CCCL_PP_DEC_IMPL_TAG218 217
+#define _CCCL_PP_DEC_IMPL_TAG219 218
+#define _CCCL_PP_DEC_IMPL_TAG220 219
+#define _CCCL_PP_DEC_IMPL_TAG221 220
+#define _CCCL_PP_DEC_IMPL_TAG222 221
+#define _CCCL_PP_DEC_IMPL_TAG223 222
+#define _CCCL_PP_DEC_IMPL_TAG224 223
+#define _CCCL_PP_DEC_IMPL_TAG225 224
+#define _CCCL_PP_DEC_IMPL_TAG226 225
+#define _CCCL_PP_DEC_IMPL_TAG227 226
+#define _CCCL_PP_DEC_IMPL_TAG228 227
+#define _CCCL_PP_DEC_IMPL_TAG229 228
+#define _CCCL_PP_DEC_IMPL_TAG230 229
+#define _CCCL_PP_DEC_IMPL_TAG231 230
+#define _CCCL_PP_DEC_IMPL_TAG232 231
+#define _CCCL_PP_DEC_IMPL_TAG233 232
+#define _CCCL_PP_DEC_IMPL_TAG234 233
+#define _CCCL_PP_DEC_IMPL_TAG235 234
+#define _CCCL_PP_DEC_IMPL_TAG236 235
+#define _CCCL_PP_DEC_IMPL_TAG237 236
+#define _CCCL_PP_DEC_IMPL_TAG238 237
+#define _CCCL_PP_DEC_IMPL_TAG239 238
+#define _CCCL_PP_DEC_IMPL_TAG240 239
+#define _CCCL_PP_DEC_IMPL_TAG241 240
+#define _CCCL_PP_DEC_IMPL_TAG242 241
+#define _CCCL_PP_DEC_IMPL_TAG243 242
+#define _CCCL_PP_DEC_IMPL_TAG244 243
+#define _CCCL_PP_DEC_IMPL_TAG245 244
+#define _CCCL_PP_DEC_IMPL_TAG246 245
+#define _CCCL_PP_DEC_IMPL_TAG247 246
+#define _CCCL_PP_DEC_IMPL_TAG248 247
+#define _CCCL_PP_DEC_IMPL_TAG249 248
+#define _CCCL_PP_DEC_IMPL_TAG250 249
+#define _CCCL_PP_DEC_IMPL_TAG251 250
+#define _CCCL_PP_DEC_IMPL_TAG252 251
+#define _CCCL_PP_DEC_IMPL_TAG253 252
+#define _CCCL_PP_DEC_IMPL_TAG254 253
+#define _CCCL_PP_DEC_IMPL_TAG255 254
+#define _CCCL_PP_DEC_IMPL_TAG256 255
+#define _CCCL_PP_DEC_IMPL_TAG257 256
+
+////////////////////////////////////////////////////////////////////////////////
+
+// _CCCL_PP_REPEAT(COUNT, MACRO, STATE, INCREMENT)
+//
+// Expands to: MACRO(STATE) MACRO(INCREMENT(STATE)) ... MACRO(INCREMENT(INCREMENT(INCREMENT(...))))
+// STATE defaults to 0, INCREMENT defaults to _CCCL_PP_INC
+#define _CCCL_PP_REPEAT_AUX1(_N, _M)         _CCCL_PP_CAT(_CCCL_PP_REPEAT, _N)(_M, 0, _CCCL_PP_INC)
+#define _CCCL_PP_REPEAT_AUX2(_N, _M, _S)     _CCCL_PP_CAT(_CCCL_PP_REPEAT, _N)(_M, _S, _CCCL_PP_INC)
+#define _CCCL_PP_REPEAT_AUX3(_N, _M, _S, _F) _CCCL_PP_CAT(_CCCL_PP_REPEAT, _N)(_M, _S, _F)
+
+#define _CCCL_PP_REPEAT_AUX(_C, _N, ...) _CCCL_PP_CAT(_CCCL_PP_REPEAT_AUX, _C)(_N, __VA_ARGS__)
+#define _CCCL_PP_REPEAT(_N, ...)         _CCCL_PP_REPEAT_AUX(_CCCL_PP_COUNT(__VA_ARGS__), _N, __VA_ARGS__)
+
+#define _CCCL_PP_REPEAT0(_M, _S, _F)
+#define _CCCL_PP_REPEAT1(_M, _S, _F)   _M(_S)
+#define _CCCL_PP_REPEAT2(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT1(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT3(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT2(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT4(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT3(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT5(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT4(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT6(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT5(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT7(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT6(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT8(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT7(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT9(_M, _S, _F)   _M(_S) _CCCL_PP_REPEAT8(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT10(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT9(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT11(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT10(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT12(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT11(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT13(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT12(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT14(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT13(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT15(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT14(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT16(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT15(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT17(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT16(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT18(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT17(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT19(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT18(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT20(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT19(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT21(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT20(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT22(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT21(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT23(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT22(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT24(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT23(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT25(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT24(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT26(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT25(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT27(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT26(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT28(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT27(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT29(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT28(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT30(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT29(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT31(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT30(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT32(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT31(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT33(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT32(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT34(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT33(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT35(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT34(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT36(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT35(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT37(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT36(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT38(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT37(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT39(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT38(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT40(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT39(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT41(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT40(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT42(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT41(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT43(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT42(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT44(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT43(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT45(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT44(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT46(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT45(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT47(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT46(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT48(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT47(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT49(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT48(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT50(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT49(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT51(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT50(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT52(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT51(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT53(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT52(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT54(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT53(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT55(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT54(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT56(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT55(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT57(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT56(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT58(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT57(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT59(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT58(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT60(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT59(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT61(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT60(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT62(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT61(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT63(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT62(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT64(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT63(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT65(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT64(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT66(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT65(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT67(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT66(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT68(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT67(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT69(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT68(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT70(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT69(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT71(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT70(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT72(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT71(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT73(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT72(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT74(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT73(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT75(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT74(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT76(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT75(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT77(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT76(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT78(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT77(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT79(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT78(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT80(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT79(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT81(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT80(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT82(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT81(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT83(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT82(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT84(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT83(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT85(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT84(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT86(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT85(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT87(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT86(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT88(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT87(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT89(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT88(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT90(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT89(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT91(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT90(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT92(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT91(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT93(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT92(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT94(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT93(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT95(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT94(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT96(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT95(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT97(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT96(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT98(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT97(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT99(_M, _S, _F)  _M(_S) _CCCL_PP_REPEAT98(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT100(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT99(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT101(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT100(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT102(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT101(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT103(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT102(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT104(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT103(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT105(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT104(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT106(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT105(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT107(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT106(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT108(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT107(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT109(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT108(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT110(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT109(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT111(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT110(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT112(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT111(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT113(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT112(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT114(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT113(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT115(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT114(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT116(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT115(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT117(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT116(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT118(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT117(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT119(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT118(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT120(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT119(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT121(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT120(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT122(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT121(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT123(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT122(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT124(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT123(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT125(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT124(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT126(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT125(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT127(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT126(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT128(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT127(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT129(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT128(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT130(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT129(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT131(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT130(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT132(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT131(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT133(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT132(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT134(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT133(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT135(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT134(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT136(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT135(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT137(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT136(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT138(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT137(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT139(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT138(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT140(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT139(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT141(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT140(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT142(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT141(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT143(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT142(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT144(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT143(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT145(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT144(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT146(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT145(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT147(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT146(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT148(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT147(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT149(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT148(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT150(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT149(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT151(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT150(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT152(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT151(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT153(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT152(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT154(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT153(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT155(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT154(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT156(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT155(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT157(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT156(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT158(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT157(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT159(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT158(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT160(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT159(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT161(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT160(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT162(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT161(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT163(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT162(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT164(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT163(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT165(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT164(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT166(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT165(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT167(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT166(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT168(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT167(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT169(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT168(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT170(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT169(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT171(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT170(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT172(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT171(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT173(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT172(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT174(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT173(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT175(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT174(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT176(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT175(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT177(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT176(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT178(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT177(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT179(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT178(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT180(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT179(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT181(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT180(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT182(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT181(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT183(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT182(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT184(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT183(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT185(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT184(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT186(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT185(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT187(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT186(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT188(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT187(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT189(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT188(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT190(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT189(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT191(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT190(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT192(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT191(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT193(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT192(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT194(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT193(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT195(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT194(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT196(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT195(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT197(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT196(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT198(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT197(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT199(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT198(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT200(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT199(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT201(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT200(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT202(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT201(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT203(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT202(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT204(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT203(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT205(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT204(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT206(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT205(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT207(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT206(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT208(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT207(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT209(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT208(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT210(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT209(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT211(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT210(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT212(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT211(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT213(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT212(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT214(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT213(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT215(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT214(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT216(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT215(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT217(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT216(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT218(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT217(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT219(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT218(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT220(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT219(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT221(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT220(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT222(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT221(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT223(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT222(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT224(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT223(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT225(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT224(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT226(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT225(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT227(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT226(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT228(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT227(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT229(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT228(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT230(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT229(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT231(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT230(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT232(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT231(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT233(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT232(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT234(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT233(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT235(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT234(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT236(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT235(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT237(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT236(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT238(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT237(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT239(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT238(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT240(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT239(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT241(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT240(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT242(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT241(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT243(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT242(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT244(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT243(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT245(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT244(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT246(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT245(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT247(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT246(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT248(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT247(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT249(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT248(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT250(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT249(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT251(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT250(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT252(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT251(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT253(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT252(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT254(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT253(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT255(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT254(_M, _F(_S), _F)
+#define _CCCL_PP_REPEAT256(_M, _S, _F) _M(_S) _CCCL_PP_REPEAT255(_M, _F(_S), _F)
+
+////////////////////////////////////////////////////////////////////////////////
+
+// _CCCL_PP_REPEAT_REVERSE(COUNT, MACRO, STATE, INCREMENT)
+//
+// Expands to: MACRO(INCREMENT(INCREMENT(INCREMENT(...)))) ... MACRO(INCREMENT(STATE)) MACRO(STATE)
+// STATE defaults to 0, INCREMENT defaults to _CCCL_PP_INC
+#define _CCCL_PP_REPEAT_REVERSE_AUX1(_N, _M)         _CCCL_PP_CAT(_CCCL_PP_REPEAT_REVERSE, _N)(_M, 0, _CCCL_PP_INC)
+#define _CCCL_PP_REPEAT_REVERSE_AUX2(_N, _M, _S)     _CCCL_PP_CAT(_CCCL_PP_REPEAT_REVERSE, _N)(_M, _S, _CCCL_PP_INC)
+#define _CCCL_PP_REPEAT_REVERSE_AUX3(_N, _M, _S, _F) _CCCL_PP_CAT(_CCCL_PP_REPEAT_REVERSE, _N)(_M, _S, _F)
+
+#define _CCCL_PP_REPEAT_REVERSE_AUX(_C, _N, ...) _CCCL_PP_CAT(_CCCL_PP_REPEAT_REVERSE_AUX, _C)(_N, __VA_ARGS__)
+#define _CCCL_PP_REPEAT_REVERSE(_N, ...)         _CCCL_PP_REPEAT_REVERSE_AUX(_CCCL_PP_COUNT(__VA_ARGS__), _N, __VA_ARGS__)
+
+#define _CCCL_PP_REPEAT_REVERSE0(_M, _S, _F)
+#define _CCCL_PP_REPEAT_REVERSE1(_M, _S, _F)   _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE2(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE1(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE3(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE2(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE4(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE3(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE5(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE4(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE6(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE5(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE7(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE6(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE8(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE7(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE9(_M, _S, _F)   _CCCL_PP_REPEAT_REVERSE8(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE10(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE9(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE11(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE10(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE12(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE11(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE13(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE12(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE14(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE13(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE15(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE14(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE16(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE15(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE17(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE16(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE18(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE17(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE19(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE18(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE20(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE19(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE21(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE20(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE22(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE21(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE23(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE22(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE24(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE23(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE25(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE24(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE26(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE25(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE27(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE26(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE28(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE27(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE29(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE28(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE30(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE29(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE31(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE30(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE32(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE31(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE33(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE32(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE34(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE33(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE35(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE34(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE36(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE35(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE37(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE36(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE38(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE37(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE39(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE38(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE40(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE39(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE41(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE40(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE42(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE41(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE43(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE42(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE44(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE43(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE45(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE44(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE46(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE45(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE47(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE46(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE48(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE47(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE49(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE48(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE50(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE49(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE51(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE50(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE52(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE51(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE53(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE52(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE54(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE53(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE55(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE54(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE56(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE55(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE57(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE56(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE58(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE57(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE59(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE58(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE60(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE59(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE61(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE60(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE62(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE61(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE63(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE62(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE64(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE63(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE65(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE64(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE66(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE65(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE67(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE66(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE68(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE67(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE69(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE68(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE70(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE69(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE71(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE70(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE72(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE71(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE73(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE72(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE74(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE73(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE75(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE74(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE76(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE75(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE77(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE76(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE78(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE77(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE79(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE78(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE80(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE79(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE81(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE80(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE82(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE81(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE83(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE82(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE84(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE83(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE85(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE84(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE86(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE85(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE87(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE86(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE88(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE87(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE89(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE88(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE90(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE89(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE91(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE90(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE92(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE91(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE93(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE92(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE94(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE93(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE95(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE94(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE96(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE95(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE97(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE96(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE98(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE97(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE99(_M, _S, _F)  _CCCL_PP_REPEAT_REVERSE98(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE100(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE99(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE101(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE100(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE102(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE101(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE103(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE102(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE104(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE103(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE105(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE104(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE106(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE105(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE107(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE106(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE108(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE107(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE109(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE108(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE110(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE109(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE111(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE110(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE112(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE111(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE113(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE112(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE114(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE113(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE115(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE114(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE116(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE115(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE117(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE116(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE118(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE117(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE119(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE118(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE120(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE119(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE121(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE120(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE122(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE121(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE123(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE122(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE124(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE123(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE125(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE124(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE126(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE125(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE127(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE126(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE128(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE127(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE129(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE128(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE130(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE129(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE131(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE130(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE132(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE131(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE133(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE132(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE134(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE133(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE135(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE134(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE136(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE135(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE137(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE136(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE138(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE137(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE139(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE138(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE140(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE139(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE141(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE140(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE142(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE141(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE143(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE142(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE144(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE143(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE145(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE144(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE146(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE145(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE147(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE146(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE148(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE147(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE149(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE148(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE150(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE149(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE151(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE150(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE152(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE151(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE153(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE152(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE154(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE153(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE155(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE154(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE156(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE155(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE157(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE156(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE158(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE157(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE159(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE158(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE160(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE159(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE161(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE160(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE162(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE161(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE163(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE162(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE164(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE163(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE165(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE164(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE166(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE165(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE167(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE166(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE168(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE167(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE169(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE168(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE170(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE169(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE171(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE170(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE172(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE171(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE173(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE172(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE174(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE173(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE175(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE174(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE176(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE175(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE177(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE176(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE178(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE177(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE179(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE178(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE180(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE179(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE181(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE180(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE182(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE181(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE183(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE182(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE184(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE183(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE185(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE184(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE186(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE185(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE187(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE186(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE188(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE187(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE189(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE188(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE190(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE189(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE191(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE190(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE192(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE191(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE193(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE192(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE194(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE193(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE195(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE194(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE196(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE195(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE197(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE196(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE198(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE197(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE199(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE198(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE200(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE199(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE201(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE200(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE202(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE201(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE203(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE202(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE204(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE203(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE205(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE204(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE206(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE205(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE207(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE206(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE208(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE207(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE209(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE208(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE210(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE209(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE211(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE210(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE212(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE211(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE213(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE212(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE214(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE213(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE215(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE214(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE216(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE215(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE217(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE216(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE218(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE217(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE219(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE218(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE220(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE219(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE221(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE220(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE222(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE221(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE223(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE222(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE224(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE223(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE225(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE224(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE226(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE225(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE227(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE226(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE228(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE227(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE229(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE228(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE230(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE229(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE231(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE230(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE232(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE231(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE233(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE232(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE234(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE233(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE235(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE234(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE236(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE235(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE237(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE236(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE238(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE237(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE239(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE238(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE240(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE239(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE241(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE240(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE242(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE241(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE243(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE242(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE244(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE243(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE245(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE244(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE246(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE245(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE247(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE246(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE248(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE247(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE249(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE248(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE250(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE249(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE251(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE250(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE252(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE251(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE253(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE252(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE254(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE253(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE255(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE254(_M, _F(_S), _F) _M(_S)
+#define _CCCL_PP_REPEAT_REVERSE256(_M, _S, _F) _CCCL_PP_REPEAT_REVERSE255(_M, _F(_S), _F) _M(_S)
+
+#define _CCCL_PP_SPLICE_WITH_IMPL1(_SEP, _P1) _P1
+#define _CCCL_PP_SPLICE_WITH_IMPL2(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL1(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL3(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL2(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL4(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL3(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL5(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL4(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL6(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL5(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL7(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL6(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL8(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL7(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL9(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL8(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL10(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL9(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL11(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL10(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL12(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL11(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL13(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL12(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL14(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL13(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL15(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL14(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL16(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL15(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL17(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL16(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL18(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL17(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL19(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL18(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL20(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL19(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL21(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL20(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL22(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL21(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL23(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL22(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL24(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL23(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL25(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL24(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL26(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL25(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL27(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL26(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL28(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL27(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL29(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL28(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL30(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL29(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL31(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL30(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL32(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL31(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL33(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL32(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL34(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL33(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL35(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL34(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL36(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL35(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL37(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL36(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL38(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL37(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL39(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL38(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL40(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL39(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL41(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL40(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL42(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL41(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL43(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL42(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL44(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL43(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL45(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL44(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL46(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL45(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL47(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL46(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL48(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL47(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL49(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL48(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL50(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL49(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL51(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL50(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL52(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL51(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL53(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL52(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL54(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL53(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL55(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL54(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL56(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL55(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL57(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL56(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL58(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL57(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL59(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL58(_SEP, __VA_ARGS__))
+#define _CCCL_PP_SPLICE_WITH_IMPL60(_SEP, _P1, ...) \
+  _CCCL_PP_CAT(_P1##_SEP, _CCCL_PP_SPLICE_WITH_IMPL59(_SEP, __VA_ARGS__))
+
+#define _CCCL_PP_SPLICE_WITH_IMPL_DISPATCH(_NUM) _CCCL_PP_SPLICE_WITH_IMPL##_NUM
+
+// Splices a pack of arguments into a single token, separated by _SEP
+// E.g., _CCCL_PP_SPLICE_WITH(_, A, B, C) will evaluate to A_B_C
+#define _CCCL_PP_SPLICE_WITH(_SEP, ...) \
+  _CCCL_PP_EXPAND(_CCCL_PP_EVAL(_CCCL_PP_SPLICE_WITH_IMPL_DISPATCH, _CCCL_PP_COUNT(__VA_ARGS__))(_SEP, __VA_ARGS__))
+
+#endif // __CCCL_PREPROCESSOR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/prologue.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/prologue.h
new file mode 100644
index 00000000..da21a9cc
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/prologue.h
@@ -0,0 +1,348 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// !!! DO NOT EDIT THIS FILE !!! This file is generated by utils/generate_prologue_epilogue.py.
+
+// NO include guards here (this file is included multiple times)
+
+#if defined(_CCCL_PROLOGUE_INCLUDED)
+#  error \
+    "cccl internal error: <cuda/std/__cccl/epilogue.h> must be included before next <cuda/std/__cccl/prologue.h> is reincluded"
+#endif
+#define _CCCL_PROLOGUE_INCLUDED() 1
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/diagnostic.h>
+#include <cuda/std/__cccl/dialect.h>
+
+// __declspec modifiers
+
+#if defined(align)
+#  pragma push_macro("align")
+#  undef align
+#  define _CCCL_POP_MACRO_align
+#endif // defined(align)
+
+#if defined(allocate)
+#  pragma push_macro("allocate")
+#  undef allocate
+#  define _CCCL_POP_MACRO_allocate
+#endif // defined(allocate)
+
+#if defined(allocator)
+#  pragma push_macro("allocator")
+#  undef allocator
+#  define _CCCL_POP_MACRO_allocator
+#endif // defined(allocator)
+
+#if defined(appdomain)
+#  pragma push_macro("appdomain")
+#  undef appdomain
+#  define _CCCL_POP_MACRO_appdomain
+#endif // defined(appdomain)
+
+#if defined(code_seg)
+#  pragma push_macro("code_seg")
+#  undef code_seg
+#  define _CCCL_POP_MACRO_code_seg
+#endif // defined(code_seg)
+
+#if defined(deprecated)
+#  pragma push_macro("deprecated")
+#  undef deprecated
+#  define _CCCL_POP_MACRO_deprecated
+#endif // defined(deprecated)
+
+#if defined(dllimport)
+#  pragma push_macro("dllimport")
+#  undef dllimport
+#  define _CCCL_POP_MACRO_dllimport
+#endif // defined(dllimport)
+
+#if defined(dllexport)
+#  pragma push_macro("dllexport")
+#  undef dllexport
+#  define _CCCL_POP_MACRO_dllexport
+#endif // defined(dllexport)
+
+#if defined(empty_bases)
+#  pragma push_macro("empty_bases")
+#  undef empty_bases
+#  define _CCCL_POP_MACRO_empty_bases
+#endif // defined(empty_bases)
+
+#if defined(hybrid_patchable)
+#  pragma push_macro("hybrid_patchable")
+#  undef hybrid_patchable
+#  define _CCCL_POP_MACRO_hybrid_patchable
+#endif // defined(hybrid_patchable)
+
+#if defined(jitintrinsic)
+#  pragma push_macro("jitintrinsic")
+#  undef jitintrinsic
+#  define _CCCL_POP_MACRO_jitintrinsic
+#endif // defined(jitintrinsic)
+
+#if defined(lifetimebound)
+#  pragma push_macro("lifetimebound")
+#  undef lifetimebound
+#  define _CCCL_POP_MACRO_lifetimebound
+#endif // defined(lifetimebound)
+
+#if defined(naked)
+#  pragma push_macro("naked")
+#  undef naked
+#  define _CCCL_POP_MACRO_naked
+#endif // defined(naked)
+
+#if defined(noalias)
+#  pragma push_macro("noalias")
+#  undef noalias
+#  define _CCCL_POP_MACRO_noalias
+#endif // defined(noalias)
+
+#if defined(noinline)
+#  pragma push_macro("noinline")
+#  undef noinline
+#  define _CCCL_POP_MACRO_noinline
+#endif // defined(noinline)
+
+#if defined(noreturn)
+#  pragma push_macro("noreturn")
+#  undef noreturn
+#  define _CCCL_POP_MACRO_noreturn
+#endif // defined(noreturn)
+
+#if defined(nothrow)
+#  pragma push_macro("nothrow")
+#  undef nothrow
+#  define _CCCL_POP_MACRO_nothrow
+#endif // defined(nothrow)
+
+#if defined(novtable)
+#  pragma push_macro("novtable")
+#  undef novtable
+#  define _CCCL_POP_MACRO_novtable
+#endif // defined(novtable)
+
+#if defined(no_sanitize_address)
+#  pragma push_macro("no_sanitize_address")
+#  undef no_sanitize_address
+#  define _CCCL_POP_MACRO_no_sanitize_address
+#endif // defined(no_sanitize_address)
+
+#if defined(process)
+#  pragma push_macro("process")
+#  undef process
+#  define _CCCL_POP_MACRO_process
+#endif // defined(process)
+
+#if defined(property)
+#  pragma push_macro("property")
+#  undef property
+#  define _CCCL_POP_MACRO_property
+#endif // defined(property)
+
+#if defined(restrict)
+#  pragma push_macro("restrict")
+#  undef restrict
+#  define _CCCL_POP_MACRO_restrict
+#endif // defined(restrict)
+
+#if defined(safebuffers)
+#  pragma push_macro("safebuffers")
+#  undef safebuffers
+#  define _CCCL_POP_MACRO_safebuffers
+#endif // defined(safebuffers)
+
+#if defined(selectany)
+#  pragma push_macro("selectany")
+#  undef selectany
+#  define _CCCL_POP_MACRO_selectany
+#endif // defined(selectany)
+
+#if defined(spectre)
+#  pragma push_macro("spectre")
+#  undef spectre
+#  define _CCCL_POP_MACRO_spectre
+#endif // defined(spectre)
+
+#if defined(thread)
+#  pragma push_macro("thread")
+#  undef thread
+#  define _CCCL_POP_MACRO_thread
+#endif // defined(thread)
+
+#if defined(uuid)
+#  pragma push_macro("uuid")
+#  undef uuid
+#  define _CCCL_POP_MACRO_uuid
+#endif // defined(uuid)
+
+// [[msvc::attribute]] attributes
+
+#if defined(msvc)
+#  pragma push_macro("msvc")
+#  undef msvc
+#  define _CCCL_POP_MACRO_msvc
+#endif // defined(msvc)
+
+#if defined(flatten)
+#  pragma push_macro("flatten")
+#  undef flatten
+#  define _CCCL_POP_MACRO_flatten
+#endif // defined(flatten)
+
+#if defined(forceinline)
+#  pragma push_macro("forceinline")
+#  undef forceinline
+#  define _CCCL_POP_MACRO_forceinline
+#endif // defined(forceinline)
+
+#if defined(forceinline_calls)
+#  pragma push_macro("forceinline_calls")
+#  undef forceinline_calls
+#  define _CCCL_POP_MACRO_forceinline_calls
+#endif // defined(forceinline_calls)
+
+#if defined(intrinsic)
+#  pragma push_macro("intrinsic")
+#  undef intrinsic
+#  define _CCCL_POP_MACRO_intrinsic
+#endif // defined(intrinsic)
+
+#if defined(noinline)
+#  pragma push_macro("noinline")
+#  undef noinline
+#  define _CCCL_POP_MACRO_noinline
+#endif // defined(noinline)
+
+#if defined(noinline_calls)
+#  pragma push_macro("noinline_calls")
+#  undef noinline_calls
+#  define _CCCL_POP_MACRO_noinline_calls
+#endif // defined(noinline_calls)
+
+#if defined(no_tls_guard)
+#  pragma push_macro("no_tls_guard")
+#  undef no_tls_guard
+#  define _CCCL_POP_MACRO_no_tls_guard
+#endif // defined(no_tls_guard)
+
+// Windows nasty macros
+
+#if defined(min)
+#  pragma push_macro("min")
+#  undef min
+#  define _CCCL_POP_MACRO_min
+#endif // defined(min)
+
+#if defined(max)
+#  pragma push_macro("max")
+#  undef max
+#  define _CCCL_POP_MACRO_max
+#endif // defined(max)
+
+#if defined(interface)
+#  pragma push_macro("interface")
+#  undef interface
+#  define _CCCL_POP_MACRO_interface
+#endif // defined(interface)
+
+// sal.h on Windows
+
+#if defined(__valid)
+#  pragma push_macro("__valid")
+#  undef __valid
+#  define _CCCL_POP_MACRO___valid
+#endif // defined(__valid)
+
+#if defined(__callback)
+#  pragma push_macro("__callback")
+#  undef __callback
+#  define _CCCL_POP_MACRO___callback
+#endif // defined(__callback)
+
+// other macros
+
+#if defined(clang)
+#  pragma push_macro("clang")
+#  undef clang
+#  define _CCCL_POP_MACRO_clang
+#endif // defined(clang)
+
+// sys/sysmacros.h on linux
+
+#if defined(major)
+#  pragma push_macro("major")
+#  undef major
+#  define _CCCL_POP_MACRO_major
+#endif // defined(major)
+
+#if defined(minor)
+#  pragma push_macro("minor")
+#  undef minor
+#  define _CCCL_POP_MACRO_minor
+#endif // defined(minor)
+
+#if defined(makedev)
+#  pragma push_macro("makedev")
+#  undef makedev
+#  define _CCCL_POP_MACRO_makedev
+#endif // defined(makedev)
+
+_CCCL_DIAG_PUSH
+_CCCL_NV_DIAG_PUSH()
+
+// disable some msvc warnings
+// https://github.com/microsoft/STL/blob/master/stl/inc/yvals_core.h#L353
+// warning C4100: 'quack': unreferenced formal parameter
+// warning C4127: conditional expression is constant
+// warning C4180: qualifier applied to function type has no meaning; ignored
+// warning C4197: 'purr': top-level volatile in cast is ignored
+// warning C4324: 'roar': structure was padded due to alignment specifier
+// warning C4455: literal suffix identifiers that do not start with an underscore are reserved
+// warning C4503: 'hum': decorated name length exceeded, name was truncated
+// warning C4522: 'woof' : multiple assignment operators specified
+// warning C4668: 'meow' is not defined as a preprocessor macro, replacing with '0' for '#if/#elif'
+// warning C4800: 'boo': forcing value to bool 'true' or 'false' (performance warning)
+// warning C4996: 'meow': was declared deprecated
+_CCCL_DIAG_SUPPRESS_MSVC(4100 4127 4180 4197 4296 4324 4455 4503 4522 4668 4800 4996)
+
+// Suppress compiler warnings about C++ extensions.
+
+#if _CCCL_COMPILER(GCC, >=, 12)
+_CCCL_DIAG_SUPPRESS_GCC("-Wc++20-extensions")
+_CCCL_DIAG_SUPPRESS_GCC("-Wc++23-extensions")
+#endif // _CCCL_COMPILER(GCC, >=, 12)
+#if _CCCL_COMPILER(GCC, >=, 14)
+_CCCL_DIAG_SUPPRESS_GCC("-Wc++26-extensions")
+#endif // _CCCL_COMPILER(GCC, >=, 14)
+
+_CCCL_DIAG_SUPPRESS_CLANG("-Wc++20-extensions")
+#if _CCCL_COMPILER(CLANG, >=, 17)
+_CCCL_DIAG_SUPPRESS_CLANG("-Wc++23-extensions")
+_CCCL_DIAG_SUPPRESS_CLANG("-Wc++26-extensions")
+#else // ^^^ _CCCL_COMPILER(CLANG, >=, 17) ^^^ / vvv _CCCL_COMPILER(CLANG, <, 17) vvv
+_CCCL_DIAG_SUPPRESS_CLANG("-Wc++2b-extensions")
+#endif // ^^^ _CCCL_COMPILER(CLANG, <, 17) ^^^
+
+// Suppress `if consteval`-related warnings.
+
+_CCCL_DIAG_SUPPRESS_NVHPC(if_consteval_nonstandard)
+_CCCL_DIAG_SUPPRESS_NVHPC(is_constant_evaluated_in_nonconstexpr_context)
+_CCCL_DIAG_SUPPRESS_NVHPC(if_consteval_in_nonconstexpr_function)
+
+_CCCL_DIAG_SUPPRESS_NVCC(3215) // "if consteval" and "if not consteval" are not standard in this mode
+_CCCL_DIAG_SUPPRESS_NVCC(3206) // "if consteval" and "if not consteval" are meaningless in a non-constexpr function
+_CCCL_DIAG_SUPPRESS_NVCC(3060) // call to __builtin_is_constant_evaluated appearing in a non-constexpr function always
+                               // produces "false"
+
+// NO include guards here (this file is included multiple times)
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/ptx_isa.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/ptx_isa.h
new file mode 100644
index 00000000..209ea812
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/ptx_isa.h
@@ -0,0 +1,369 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_PTX_ISA_H_
+#define __CCCL_PTX_ISA_H_
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <nv/target> // __CUDA_MINIMUM_ARCH__ and friends
+
+/*
+ * Targeting macros
+ *
+ * Information from:
+ * https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#release-notes
+ */
+
+// The first define is for future major versions of CUDACC.
+// We make sure that these get the highest known PTX ISA version.
+// For clang cuda check
+// https://github.com/llvm/llvm-project/blob/release/<VER>.x/clang/lib/Driver/ToolChains/Cuda.cpp getNVPTXTargetFeatures
+#if _CCCL_CUDACC_AT_LEAST(14, 0) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 940ULL
+// PTX ISA 9.4 is available from CUDA 13.4
+#elif _CCCL_CUDACC_AT_LEAST(13, 4) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 940ULL
+// PTX ISA 9.3 is available from CUDA 13.3
+#elif _CCCL_CUDACC_AT_LEAST(13, 3) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 930ULL
+// PTX ISA 9.2 is available from CUDA 13.2
+#elif _CCCL_CUDACC_AT_LEAST(13, 2) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 920ULL
+// PTX ISA 9.1 is available from CUDA 13.1
+#elif _CCCL_CUDACC_AT_LEAST(13, 1) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 910ULL
+// PTX ISA 9.0 is available from CUDA 13.0, driver r580
+#elif _CCCL_CUDACC_AT_LEAST(13, 0) && !_CCCL_CUDA_COMPILER(CLANG)
+#  define __cccl_ptx_isa 900ULL
+// PTX ISA 8.8 is available from CUDA 12.9, driver r575
+#elif _CCCL_CUDACC_AT_LEAST(12, 9) && !_CCCL_CUDA_COMPILER(CLANG, <, 22)
+#  define __cccl_ptx_isa 880ULL
+// PTX ISA 8.7 is available from CUDA 12.8, driver r570
+#elif _CCCL_CUDACC_AT_LEAST(12, 8) && !_CCCL_CUDA_COMPILER(CLANG, <, 20)
+#  define __cccl_ptx_isa 870ULL
+// PTX ISA 8.5 is available from CUDA 12.5, driver r555
+#elif _CCCL_CUDACC_AT_LEAST(12, 5) && !_CCCL_CUDA_COMPILER(CLANG, <, 19)
+#  define __cccl_ptx_isa 850ULL
+// PTX ISA 8.4 is available from CUDA 12.4, driver r550
+#elif _CCCL_CUDACC_AT_LEAST(12, 4) && !_CCCL_CUDA_COMPILER(CLANG, <, 19)
+#  define __cccl_ptx_isa 840ULL
+// PTX ISA 8.3 is available from CUDA 12.3, driver r545
+#elif _CCCL_CUDACC_AT_LEAST(12, 3) && !_CCCL_CUDA_COMPILER(CLANG, <, 18)
+#  define __cccl_ptx_isa 830ULL
+// PTX ISA 8.2 is available from CUDA 12.2, driver r535
+#elif _CCCL_CUDACC_AT_LEAST(12, 2) && !_CCCL_CUDA_COMPILER(CLANG, <, 18)
+#  define __cccl_ptx_isa 820ULL
+// PTX ISA 8.1 is available from CUDA 12.1, driver r530
+#elif _CCCL_CUDACC_AT_LEAST(12, 1) && !_CCCL_CUDA_COMPILER(CLANG, <, 17)
+#  define __cccl_ptx_isa 810ULL
+// PTX ISA 8.0 is available from CUDA 12.0, driver r525
+#elif _CCCL_CUDACC_AT_LEAST(12, 0) && !_CCCL_CUDA_COMPILER(CLANG, <, 17)
+#  define __cccl_ptx_isa 800ULL
+// PTX ISA 7.8 is available from CUDA 11.8, driver r520
+#elif _CCCL_CUDACC_AT_LEAST(11, 8) && !_CCCL_CUDA_COMPILER(CLANG, <, 16)
+#  define __cccl_ptx_isa 780ULL
+// PTX ISA 7.7 is available from CUDA 11.7, driver r515
+#elif _CCCL_CUDACC_AT_LEAST(11, 7) && !_CCCL_CUDA_COMPILER(CLANG, <, 16)
+#  define __cccl_ptx_isa 770ULL
+// PTX ISA 7.6 is available from CUDA 11.6, driver r510
+#elif _CCCL_CUDACC_AT_LEAST(11, 6) && !_CCCL_CUDA_COMPILER(CLANG, <, 16)
+#  define __cccl_ptx_isa 760ULL
+// PTX ISA 7.5 is available from CUDA 11.5, driver r495
+#elif _CCCL_CUDACC_AT_LEAST(11, 5) && !_CCCL_CUDA_COMPILER(CLANG, <, 14)
+#  define __cccl_ptx_isa 750ULL
+// PTX ISA 7.4 is available from CUDA 11.4, driver r470
+#elif _CCCL_CUDACC_AT_LEAST(11, 4) && !_CCCL_CUDA_COMPILER(CLANG, <, 14)
+#  define __cccl_ptx_isa 740ULL
+// PTX ISA 7.3 is available from CUDA 11.3, driver r465
+#elif _CCCL_CUDACC_AT_LEAST(11, 3) && !_CCCL_CUDA_COMPILER(CLANG, <, 14)
+#  define __cccl_ptx_isa 730ULL
+// PTX ISA 7.2 is available from CUDA 11.2, driver r460
+#elif _CCCL_CUDACC_AT_LEAST(11, 2) && !_CCCL_CUDA_COMPILER(CLANG, <, 13)
+#  define __cccl_ptx_isa 720ULL
+// PTX ISA 7.1 is available from CUDA 11.1, driver r455
+#elif _CCCL_CUDACC_AT_LEAST(11, 1) && !_CCCL_CUDA_COMPILER(CLANG, <, 13)
+#  define __cccl_ptx_isa 710ULL
+// PTX ISA 7.0 is available from CUDA 11.0, driver r445
+#elif _CCCL_CUDACC_AT_LEAST(11, 0) && !_CCCL_CUDA_COMPILER(CLANG, <, 12)
+#  define __cccl_ptx_isa 700ULL
+// Fallback case. Define the ISA version to be zero. This ensures that the macro is always defined.
+#else
+#  define __cccl_ptx_isa 0ULL
+#endif
+
+// We define certain feature test macros depending on availability. When
+// __CUDA_MINIMUM_ARCH__ is not available, we define the following features
+// depending on PTX ISA. This permits checking for the feature in host code.
+// When __CUDA_MINIMUM_ARCH__ is available, we only enable the feature when the
+// hardware supports it.
+#if __cccl_ptx_isa >= 800
+#  if (!defined(__CUDA_MINIMUM_ARCH__)) || (defined(__CUDA_MINIMUM_ARCH__) && 900 <= __CUDA_MINIMUM_ARCH__)
+#    define __cccl_lib_local_barrier_arrive_tx
+#    define __cccl_lib_experimental_ctk12_cp_async_exposure
+#  endif
+#endif // __cccl_ptx_isa >= 800
+
+// NVRTC ships a built-in copy of <nv/detail/__target_macros>, so including CCCL's version of this header will omit the
+// content since the header guards are already defined. To make older NVRTC versions have a few newer feature macros
+// required for the PTX tests, we define them here outside the header guards.
+// TODO(bgruber): limit this workaround to NVRTC versions older than the first one shipping those macros
+#if _CCCL_COMPILER(NVRTC)
+
+// missing SM_88
+#  if !defined(NV_PROVIDES_SM_88)
+#    define _NV_TARGET_VAL_SM_88 880
+#    define NV_PROVIDES_SM_88    __NV_PROVIDES_SM_88
+#    define NV_IS_EXACTLY_SM_88  __NV_IS_EXACTLY_SM_88
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_88)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_88 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_88      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_88 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_88      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_88)
+#      define _NV_TARGET___NV_PROVIDES_SM_88      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_88 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_88      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_88 0
+#    endif
+#  endif // !NV_PROVIDES_SM_88
+
+// missing SM_90a
+#  ifndef NV_HAS_FEATURE_SM_90a
+#    define NV_HAS_FEATURE_SM_90a __NV_HAS_FEATURE_SM_90a
+#    if defined(__CUDA_ARCH_FEAT_SM90_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 900))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_90a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_90a 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_90a
+
+// missing SM_100
+#  ifndef NV_PROVIDES_SM_100
+#    define _NV_TARGET_VAL_SM_100 1000
+#    define NV_PROVIDES_SM_100    __NV_PROVIDES_SM_100
+#    define NV_IS_EXACTLY_SM_100  __NV_IS_EXACTLY_SM_100
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_100)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_100      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_100      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_100)
+#      define _NV_TARGET___NV_PROVIDES_SM_100      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_100      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 0
+#    endif
+#  endif // !NV_PROVIDES_SM_100
+
+// missing SM_100a
+#  ifndef NV_HAS_FEATURE_SM_100a
+#    define NV_HAS_FEATURE_SM_100a __NV_HAS_FEATURE_SM_100a
+#    if defined(__CUDA_ARCH_FEAT_SM100_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 1000))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100a 0
+#    endif
+#  endif // !NV_HAS_FEATURE_SM_100a
+
+// missing SM_103
+#  ifndef NV_PROVIDES_SM_103
+#    define _NV_TARGET_VAL_SM_103 1030
+#    define NV_PROVIDES_SM_103    __NV_PROVIDES_SM_103
+#    define NV_IS_EXACTLY_SM_103  __NV_IS_EXACTLY_SM_103
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_103)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_103 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_103      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_103 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_103      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_103)
+#      define _NV_TARGET___NV_PROVIDES_SM_103      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_103 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_103      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_103 0
+#    endif
+#  endif // !NV_PROVIDES_SM_103
+
+// missing SM_103
+#  ifndef NV_HAS_FEATURE_SM_103a
+#    define NV_HAS_FEATURE_SM_103a __NV_HAS_FEATURE_SM_103a
+#    if defined(__CUDA_ARCH_FEAT_SM103_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 1030))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103a 0
+#    endif
+#  endif // !NV_HAS_FEATURE_SM_103a
+
+// missing SM_110
+#  ifndef NV_PROVIDES_SM_110
+#    define _NV_TARGET_VAL_SM_110 1100
+#    define NV_PROVIDES_SM_110    __NV_PROVIDES_SM_110
+#    define NV_IS_EXACTLY_SM_110  __NV_IS_EXACTLY_SM_110
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_110)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_110 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_110      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_110 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_110      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_110)
+#      define _NV_TARGET___NV_PROVIDES_SM_110      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_110 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_110      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_110 0
+#    endif
+#  endif // !NV_PROVIDES_SM_110
+
+// missing SM_110a
+#  ifndef NV_HAS_FEATURE_SM_110a
+#    define NV_HAS_FEATURE_SM_110a __NV_HAS_FEATURE_SM_110a
+#    if defined(__CUDA_ARCH_FEAT_SM110_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 1100))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110a 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_110a
+
+// missing SM_120
+#  ifndef NV_PROVIDES_SM_120
+#    define _NV_TARGET_VAL_SM_120 1200
+#    define NV_PROVIDES_SM_120    __NV_PROVIDES_SM_120
+#    define NV_IS_EXACTLY_SM_120  __NV_IS_EXACTLY_SM_120
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_120)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_120 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_120      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_120 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_120      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_120)
+#      define _NV_TARGET___NV_PROVIDES_SM_120      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_120 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_120      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_120 0
+#    endif
+#  endif // !NV_PROVIDES_SM_120
+
+// missing SM_120a
+#  ifndef NV_HAS_FEATURE_SM_120a
+#    define NV_HAS_FEATURE_SM_120a __NV_HAS_FEATURE_SM_120a
+#    if defined(__CUDA_ARCH_FEAT_SM120_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 1200))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120a 0
+#    endif
+#  endif // _CCCL_COMPILER(NVRTC)
+
+// missing SM_121
+#  if !defined(NV_PROVIDES_SM_121)
+#    define _NV_TARGET_VAL_SM_121 1210
+#    define NV_PROVIDES_SM_121    __NV_PROVIDES_SM_121
+#    define NV_IS_EXACTLY_SM_121  __NV_IS_EXACTLY_SM_121
+#    if (__CUDA_ARCH__ == _NV_TARGET_VAL_SM_121)
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_121 1
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_121      1
+#    else
+#      define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_121 0
+#      define _NV_TARGET___NV_IS_EXACTLY_SM_121      0
+#    endif
+#    if (__CUDA_ARCH__ >= _NV_TARGET_VAL_SM_121)
+#      define _NV_TARGET___NV_PROVIDES_SM_121      1
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_121 1
+#    else
+#      define _NV_TARGET___NV_PROVIDES_SM_121      0
+#      define _NV_TARGET_BOOL___NV_PROVIDES_SM_121 0
+#    endif
+#  endif // !NV_PROVIDES_SM_121
+
+// missing SM_121a
+#  ifndef NV_HAS_FEATURE_SM_121a
+#    define NV_HAS_FEATURE_SM_121a __NV_HAS_FEATURE_SM_121a
+#    if defined(__CUDA_ARCH_FEAT_SM121_ALL) || (defined(__CUDA_ARCH_SPECIFIC__) && (__CUDA_ARCH_SPECIFIC__ == 1210))
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121a 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121a 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_121a
+
+//----------------------------------------------------------------------------------------------------------------------
+// family-specific SM versions
+
+// missing SM_100f
+#  ifndef NV_HAS_FEATURE_SM_100f
+#    define NV_HAS_FEATURE_SM_100f __NV_HAS_FEATURE_SM_100f
+#    if defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ == 1000)
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100f 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100f 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_100
+
+// missing SM_103f
+#  ifndef NV_HAS_FEATURE_SM_103f
+#    define NV_HAS_FEATURE_SM_103f __NV_HAS_FEATURE_SM_103f
+#    if defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ == 1030)
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103f 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103f 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_103f
+
+// missing SM_110f
+#  ifndef NV_HAS_FEATURE_SM_110f
+#    define NV_HAS_FEATURE_SM_110f __NV_HAS_FEATURE_SM_110f
+#    if defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ == 1100)
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110f 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110f 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_110f
+
+// missing SM_120f
+#  ifndef NV_HAS_FEATURE_SM_120f
+#    define NV_HAS_FEATURE_SM_120f __NV_HAS_FEATURE_SM_120f
+#    if defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ == 1200)
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120f 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120f 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_120f
+
+// missing SM_121f
+#  ifndef NV_HAS_FEATURE_SM_121f
+#    define NV_HAS_FEATURE_SM_121f __NV_HAS_FEATURE_SM_121f
+#    if defined(__CUDA_ARCH_FAMILY_SPECIFIC__) && (__CUDA_ARCH_FAMILY_SPECIFIC__ == 1210)
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121f 1
+#    else
+#      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121f 0
+#    endif
+#  endif // NV_HAS_FEATURE_SM_121f
+
+#endif // _CCCL_COMPILER(NVRTC)
+#endif // __CCCL_PTX_ISA_H_
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/rtti.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/rtti.h
new file mode 100644
index 00000000..14ce5dc5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/rtti.h
@@ -0,0 +1,72 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_RTTI_H
+#define __CCCL_RTTI_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/builtin.h>
+
+// NOTE: some compilers support the `typeid` feature but not the `dynamic_cast`
+// feature. This is why we have separate macros for each.
+
+#ifndef _CCCL_NO_RTTI
+#  if defined(CCCL_DISABLE_RTTI) // Escape hatch for users to manually disable RTTI
+#    define _CCCL_NO_RTTI
+#  elif defined(__CUDA_ARCH__)
+#    define _CCCL_NO_RTTI // No RTTI in CUDA device code
+#  elif _CCCL_COMPILER(NVRTC)
+#    define _CCCL_NO_RTTI
+#  elif _CCCL_COMPILER(MSVC)
+#    if _CPPRTTI == 0
+#      define _CCCL_NO_RTTI
+#    endif
+#  elif _CCCL_COMPILER(CLANG)
+#    if !_CCCL_HAS_FEATURE(cxx_rtti)
+#      define _CCCL_NO_RTTI
+#    endif
+#  else
+#    if __GXX_RTTI == 0 && __cpp_rtti == 0
+#      define _CCCL_NO_RTTI
+#    endif
+#  endif
+#endif // !_CCCL_NO_RTTI
+
+#ifndef _CCCL_NO_TYPEID
+#  if defined(CCCL_DISABLE_RTTI) // CCCL_DISABLE_RTTI disables typeid also
+#    define _CCCL_NO_TYPEID
+#  elif defined(__CUDA_ARCH__)
+#    define _CCCL_NO_TYPEID // No typeid in CUDA device code
+#  elif _CCCL_COMPILER(NVRTC)
+#    define _CCCL_NO_TYPEID
+#  elif _CCCL_COMPILER(MSVC)
+// No-op, MSVC always supports typeid even when RTTI is disabled
+#  elif _CCCL_COMPILER(CLANG)
+#    if !_CCCL_HAS_FEATURE(cxx_rtti)
+#      define _CCCL_NO_TYPEID
+#    endif
+#  else
+#    if __GXX_RTTI == 0 && __cpp_rtti == 0
+#      define _CCCL_NO_TYPEID
+#    endif
+#  endif
+#endif // !_CCCL_NO_TYPEID
+
+#endif // __CCCL_RTTI_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/sequence_access.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/sequence_access.h
new file mode 100644
index 00000000..574b44a8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/sequence_access.h
@@ -0,0 +1,83 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_SEQUENCE_ACCESS_H
+#define __CCCL_SEQUENCE_ACCESS_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// We need to define hidden friends for {cr,r,}{begin,end} of our containers as we will otherwise encounter ambigouities
+#define _CCCL_SYNTHESIZE_SEQUENCE_ACCESS(_ClassName, _ConstIter)                                                       \
+  [[nodiscard]] _CCCL_API friend iterator begin(_ClassName& __sequence) noexcept(noexcept(__sequence.begin()))         \
+  {                                                                                                                    \
+    return __sequence.begin();                                                                                         \
+  }                                                                                                                    \
+  [[nodiscard]] _CCCL_API friend _ConstIter begin(const _ClassName& __sequence) noexcept(noexcept(__sequence.begin())) \
+  {                                                                                                                    \
+    return __sequence.begin();                                                                                         \
+  }                                                                                                                    \
+  [[nodiscard]] _CCCL_API friend iterator end(_ClassName& __sequence) noexcept(noexcept(__sequence.end()))             \
+  {                                                                                                                    \
+    return __sequence.end();                                                                                           \
+  }                                                                                                                    \
+  [[nodiscard]] _CCCL_API friend _ConstIter end(const _ClassName& __sequence) noexcept(noexcept(__sequence.end()))     \
+  {                                                                                                                    \
+    return __sequence.end();                                                                                           \
+  }                                                                                                                    \
+  [[nodiscard]] _CCCL_API friend _ConstIter cbegin(const _ClassName& __sequence) noexcept(                             \
+    noexcept(__sequence.begin()))                                                                                      \
+  {                                                                                                                    \
+    return __sequence.begin();                                                                                         \
+  }                                                                                                                    \
+  [[nodiscard]] _CCCL_API friend _ConstIter cend(const _ClassName& __sequence) noexcept(noexcept(__sequence.end()))    \
+  {                                                                                                                    \
+    return __sequence.end();                                                                                           \
+  }
+#define _CCCL_SYNTHESIZE_SEQUENCE_REVERSE_ACCESS(_ClassName, _ConstRevIter)                                          \
+  [[nodiscard]] _CCCL_API friend reverse_iterator rbegin(_ClassName& __sequence) noexcept(                           \
+    noexcept(__sequence.rbegin()))                                                                                   \
+  {                                                                                                                  \
+    return __sequence.rbegin();                                                                                      \
+  }                                                                                                                  \
+  [[nodiscard]] _CCCL_API friend _ConstRevIter rbegin(const _ClassName& __sequence) noexcept(                        \
+    noexcept(__sequence.rbegin()))                                                                                   \
+  {                                                                                                                  \
+    return __sequence.rbegin();                                                                                      \
+  }                                                                                                                  \
+  [[nodiscard]] _CCCL_API friend reverse_iterator rend(_ClassName& __sequence) noexcept(noexcept(__sequence.rend())) \
+  {                                                                                                                  \
+    return __sequence.rend();                                                                                        \
+  }                                                                                                                  \
+  [[nodiscard]] _CCCL_API friend _ConstRevIter rend(const _ClassName& __sequence) noexcept(                          \
+    noexcept(__sequence.rend()))                                                                                     \
+  {                                                                                                                  \
+    return __sequence.rend();                                                                                        \
+  }                                                                                                                  \
+  [[nodiscard]] _CCCL_API friend _ConstRevIter crbegin(const _ClassName& __sequence) noexcept(                       \
+    noexcept(__sequence.rbegin()))                                                                                   \
+  {                                                                                                                  \
+    return __sequence.rbegin();                                                                                      \
+  }                                                                                                                  \
+  [[nodiscard]] _CCCL_API friend _ConstRevIter crend(const _ClassName& __sequence) noexcept(                         \
+    noexcept(__sequence.rend()))                                                                                     \
+  {                                                                                                                  \
+    return __sequence.rend();                                                                                        \
+  }
+
+#endif // __CCCL_SEQUENCE_ACCESS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/system_header.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/system_header.h
new file mode 100644
index 00000000..3d2d1ac6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/system_header.h
@@ -0,0 +1,38 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_SYSTEM_HEADER_H
+#define __CCCL_SYSTEM_HEADER_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/is_non_narrowing_convertible.h> // IWYU pragma: export
+
+// Enforce that cccl headers are treated as system headers
+#if _CCCL_COMPILER(GCC) || _CCCL_COMPILER(NVHPC)
+#  define _CCCL_FORCE_SYSTEM_HEADER_GCC
+#elif _CCCL_COMPILER(CLANG)
+#  define _CCCL_FORCE_SYSTEM_HEADER_CLANG
+#elif _CCCL_COMPILER(MSVC)
+#  define _CCCL_FORCE_SYSTEM_HEADER_MSVC
+#endif // other compilers
+
+// Potentially enable that cccl headers are treated as system headers
+#if !defined(_CCCL_NO_SYSTEM_HEADER) && !(_CCCL_COMPILER(MSVC) && defined(_LIBCUDACXX_DISABLE_PRAGMA_MSVC_WARNING)) \
+  && !_CCCL_COMPILER(NVRTC) && !defined(_LIBCUDACXX_DISABLE_PRAGMA_GCC_SYSTEM_HEADER)
+#  if _CCCL_COMPILER(GCC) || _CCCL_COMPILER(NVHPC)
+#    define _CCCL_IMPLICIT_SYSTEM_HEADER_GCC
+#  elif _CCCL_COMPILER(CLANG)
+#    define _CCCL_IMPLICIT_SYSTEM_HEADER_CLANG
+#  elif _CCCL_COMPILER(MSVC)
+#    define _CCCL_IMPLICIT_SYSTEM_HEADER_MSVC
+#  endif // other compilers
+#endif // Use system header
+
+#endif // __CCCL_SYSTEM_HEADER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/unreachable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/unreachable.h
new file mode 100644
index 00000000..2c2dba1a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/unreachable.h
@@ -0,0 +1,31 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_UNREACHABLE_H
+#define __CCCL_UNREACHABLE_H
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_COMPILER(MSVC) && !_CCCL_DEVICE_COMPILATION()
+#  define _CCCL_UNREACHABLE() __assume(0)
+#else
+#  define _CCCL_UNREACHABLE() __builtin_unreachable()
+#endif
+
+#endif // __CCCL_UNREACHABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/version.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/version.h
new file mode 100644
index 00000000..781c7125
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/version.h
@@ -0,0 +1,26 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+// This file is somewhat automatically generated. Disable clang-format.
+// clang-format off
+
+
+#ifndef __CCCL_VERSION_H
+#define __CCCL_VERSION_H
+
+#define CCCL_VERSION 3005000
+#define CCCL_MAJOR_VERSION (CCCL_VERSION / 1000000)
+#define CCCL_MINOR_VERSION (((CCCL_VERSION / 1000) % 1000))
+#define CCCL_PATCH_VERSION (CCCL_VERSION % 1000)
+
+#if CCCL_PATCH_VERSION > 99
+#  error "CCCL patch version cannot be greater than 99 for compatibility with Thrust/CUB's MMMmmmpp format."
+#endif
+
+#endif // __CCCL_VERSION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/visibility.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/visibility.h
new file mode 100644
index 00000000..3cfba491
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cccl/visibility.h
@@ -0,0 +1,198 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __CCCL_VISIBILITY_H
+#define __CCCL_VISIBILITY_H
+
+#ifndef _CUDA__CCCL_CONFIG
+#  error "<__cccl/visibility.h> should only be included in from <cuda/__cccl_config>"
+#endif // _CUDA__CCCL_CONFIG
+
+#include <cuda/std/__cccl/compiler.h>
+#include <cuda/std/__cccl/system_header.h>
+
+// We want to ensure that all warning emitting from this header are suppressed
+#if defined(_CCCL_FORCE_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_FORCE_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_FORCE_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/attributes.h>
+#include <cuda/std/__cccl/cuda_capabilities.h>
+#include <cuda/std/__cccl/execution_space.h>
+#include <cuda/std/__cccl/os.h>
+
+// For unknown reasons, nvc++ need to selectively disable this warning
+// We do not want to use our usual macro because that would have push / pop semantics
+#if _CCCL_COMPILER(NVHPC)
+#  pragma nv_diag_suppress 1407
+#endif // _CCCL_COMPILER(NVHPC)
+
+// Enable us to hide kernels
+#if _CCCL_OS(WINDOWS) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_VISIBILITY_HIDDEN
+#else // ^^^ _CCCL_COMPILER(NVRTC) ^^^ / vvv _CCCL_COMPILER(NVRTC) vvv
+#  define _CCCL_VISIBILITY_HIDDEN __attribute__((__visibility__("hidden")))
+#endif // !_CCCL_COMPILER(NVRTC)
+
+#if _CCCL_COMPILER(NVRTC)
+#  define _CCCL_VISIBILITY_DEFAULT
+#elif _CCCL_OS(WINDOWS)
+#  define _CCCL_VISIBILITY_DEFAULT __declspec(dllimport)
+#else // ^^^ _CCCL_COMPILER(NVRTC) ^^^ / vvv !_CCCL_COMPILER(NVRTC) vvv
+#  define _CCCL_VISIBILITY_DEFAULT __attribute__((__visibility__("default")))
+#endif // !_CCCL_COMPILER(NVRTC)
+
+#if _CCCL_COMPILER(NVRTC)
+#  define _CCCL_VISIBILITY_EXPORT
+#elif _CCCL_OS(WINDOWS)
+#  define _CCCL_VISIBILITY_EXPORT __declspec(dllexport)
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_VISIBILITY_EXPORT _CCCL_VISIBILITY_DEFAULT
+#endif // !_CCCL_COMPILER(MSVC)
+
+#if _CCCL_OS(WINDOWS) || _CCCL_COMPILER(NVRTC)
+#  define _CCCL_TYPE_VISIBILITY_DEFAULT
+#  define _CCCL_TYPE_VISIBILITY_HIDDEN
+#elif _CCCL_HAS_ATTRIBUTE(__type_visibility__)
+#  define _CCCL_TYPE_VISIBILITY_DEFAULT __attribute__((__type_visibility__("default")))
+#  define _CCCL_TYPE_VISIBILITY_HIDDEN  __attribute__((__type_visibility__("hidden")))
+#else // ^^^ _CCCL_HAS_ATTRIBUTE(__type_visibility__) ^^^ / vvv !_CCCL_HAS_ATTRIBUTE(__type_visibility__) vvv
+#  define _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_VISIBILITY_DEFAULT
+#  define _CCCL_TYPE_VISIBILITY_HIDDEN  _CCCL_VISIBILITY_HIDDEN
+#endif // !_CCCL_COMPILER(NVRTC)
+
+#if _CCCL_COMPILER(MSVC)
+#  define _CCCL_FORCEINLINE __forceinline
+#  define _CCCL_FORCEINLINE_LAMBDA
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define _CCCL_FORCEINLINE        __inline__ __attribute__((__always_inline__))
+#  define _CCCL_FORCEINLINE_LAMBDA __attribute__((__always_inline__))
+#endif // ^^^ !_CCCL_COMPILER(MSVC) ^^^
+
+#if _CCCL_COMPILER(NVRTC)
+#  define _CCCL_NOINLINE __attribute__((noinline))
+#elif _CCCL_OS(WINDOWS)
+#  define _CCCL_NOINLINE __declspec(noinline)
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv _CCCL_COMPILER(MSVC) vvv
+// We can't use __noinline__ here because of CTK defining this macro.
+#  define _CCCL_NOINLINE __attribute__((noinline))
+#endif // ^^^ !_CCCL_COMPILER(MSVC) ^^^
+
+#if _CCCL_DEVICE_COMPILATION()
+#  define _CCCL_NOINLINE_DEVICE _CCCL_NOINLINE
+#else // ^^^ _CCCL_DEVICE_COMPILATION() ^^^ / vvv !_CCCL_DEVICE_COMPILATION() vvv
+#  define _CCCL_NOINLINE_DEVICE
+#endif // ^^^ !_CCCL_DEVICE_COMPILATION() ^^^
+
+#if _CCCL_HAS_ATTRIBUTE(__exclude_from_explicit_instantiation__)
+#  define _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION __attribute__((__exclude_from_explicit_instantiation__))
+#else // ^^^ exclude_from_explicit_instantiation ^^^ / vvv !exclude_from_explicit_instantiation vvv
+// NVCC complains mightily about being unable to inline functions if we use _CCCL_FORCEINLINE here
+#  define _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#endif // !exclude_from_explicit_instantiation
+
+#if _CCCL_COMPILER(NVHPC) // NVHPC has issues with visibility attributes on symbols with internal linkage
+#  define _CCCL_HIDE_FROM_ABI inline
+#else // ^^^ _CCCL_COMPILER(NVHPC) ^^^ / vvv !_CCCL_COMPILER(NVHPC) vvv
+#  define _CCCL_HIDE_FROM_ABI _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION inline
+#endif // !_CCCL_COMPILER(NVHPC)
+
+// Note: we will allow the user to redefine _CCCL_KERNEL_ATTRIBUTES until CCCL 4.0, since they may have
+// redefined CUB_DETAIL_KERNEL_ATTRIBUTES or THRUST_DETAIL_KERNEL_ATTRIBUTES.
+#if !defined(_CCCL_KERNEL_ATTRIBUTES)
+#  define _CCCL_KERNEL_ATTRIBUTES __global__ _CCCL_VISIBILITY_HIDDEN
+#endif // !_CCCL_KERNEL_ATTRIBUTES
+
+#if defined(CUB_DETAIL_KERNEL_ATTRIBUTES) || defined(THRUST_DETAIL_KERNEL_ATTRIBUTES)
+#  error \
+    "Redefining CCCL's kernel attributes via CUB_DETAIL_KERNEL_ATTRIBUTES or THRUST_DETAIL_KERNEL_ATTRIBUTES is not allowed. If you absolutely rely on this, you can override them by defining _CCCL_KERNEL_ATTRIBUTES, but this will be disallowed in CCCL 4.0."
+#endif // !_CCCL_KERNEL_ATTRIBUTES
+
+//! @brief \c _CCCL_HIDE_FROM_ABI and \c _CCCL_FORCEINLINE cannot be used together because
+//! they both try to add `inline` to the function declaration. The following macros slice
+//! the function attributes differently to avoid this problem:
+//! - \c _CCCL_API declares the function host/device and hides the symbol from the ABI
+//! - \c _CCCL_NODEBUG_API does the same while also hiding the function from
+//!   debuggers and marking the function as \c inline.
+//! - \c _CCCL_TRIVIAL_API does the same as \c _CCCL_NODEBUG_API while also force-inlining
+//!   the function.
+#if _CCCL_COMPILER(NVHPC) // NVHPC has issues with visibility attributes on symbols with internal linkage
+#  define _CCCL_API             _CCCL_HOST_DEVICE
+#  define _CCCL_HOST_DEVICE_API _CCCL_HOST_DEVICE
+#  define _CCCL_HOST_API        _CCCL_HOST
+#  define _CCCL_DEVICE_API      _CCCL_DEVICE
+#  define _CCCL_TILE_API        _CCCL_TILE
+#else // ^^^ _CCCL_COMPILER(NVHPC) ^^^ / vvv !_CCCL_COMPILER(NVHPC) vvv
+#  define _CCCL_API             _CCCL_TILE _CCCL_HOST_DEVICE _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#  define _CCCL_HOST_DEVICE_API _CCCL_HOST_DEVICE _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#  define _CCCL_HOST_API        _CCCL_HOST _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#  define _CCCL_DEVICE_API      _CCCL_DEVICE _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#  define _CCCL_TILE_API        _CCCL_TILE _CCCL_VISIBILITY_HIDDEN _CCCL_EXCLUDE_FROM_EXPLICIT_INSTANTIATION
+#endif // !_CCCL_COMPILER(NVHPC)
+
+//! @brief \c _CCCL_NODEBUG_API marks a function's visibility as hidden and causes
+//! debuggers to skip it. This is useful for functions like \c cuda::std::move that
+//! debuggers should not step into. If a \c _CCCL_NODEBUG_API function \c F calls a normal
+//! function \c G, stepping into \c F in a debugger will skip over \c F and step directly
+//! into \c G. In a stacktrace, \c F will still be shone, but you will not be able to
+//! set the debugger's active frame to \c F.
+#define _CCCL_NODEBUG_API        _CCCL_API _CCCL_ARTIFICIAL _CCCL_NODEBUG inline
+#define _CCCL_NODEBUG_HOST_API   _CCCL_HOST_API _CCCL_ARTIFICIAL _CCCL_NODEBUG inline
+#define _CCCL_NODEBUG_DEVICE_API _CCCL_DEVICE_API _CCCL_ARTIFICIAL _CCCL_NODEBUG inline
+
+//! @brief \c _CCCL_TRIVIAL_API force-inlines a function, marks its visibility as hidden,
+//! and causes debuggers to skip it. This is useful for trivial internal functions that do
+//! dispatching or other plumbing work. It is particularly useful in the definition of
+//! customization point objects.
+#define _CCCL_TRIVIAL_API        _CCCL_API _CCCL_ARTIFICIAL _CCCL_NODEBUG _CCCL_FORCEINLINE
+#define _CCCL_TRIVIAL_HOST_API   _CCCL_HOST_API _CCCL_ARTIFICIAL _CCCL_NODEBUG _CCCL_FORCEINLINE
+#define _CCCL_TRIVIAL_DEVICE_API _CCCL_DEVICE_API _CCCL_ARTIFICIAL _CCCL_NODEBUG _CCCL_FORCEINLINE
+
+// Some functions have their addresses appear in public types (e.g., in
+// `cuda::__overrides_for` specializations). If the function is declared
+// `__attribute__((visibility("hidden")))`, and if the address appears, say, in the type
+// of a member of a class that is declared `__attribute__((visibility("default")))`, GCC
+// complains bitterly. So we avoid declaring those functions `hidden`. Instead of the
+// typical `_CCCL_API` macro, we use `_CCCL_PUBLIC_API` for those functions.
+#if _CCCL_OS(WINDOWS)
+#  define _CCCL_PUBLIC_API        _CCCL_HOST_DEVICE
+#  define _CCCL_PUBLIC_HOST_API   _CCCL_HOST
+#  define _CCCL_PUBLIC_DEVICE_API _CCCL_DEVICE
+#else // ^^^ _CCCL_OS(WINDOWS) ^^^ / vvv !_CCCL_OS(WINDOWS) vvv
+#  define _CCCL_PUBLIC_API        _CCCL_HOST_DEVICE _CCCL_VISIBILITY_DEFAULT
+#  define _CCCL_PUBLIC_HOST_API   _CCCL_HOST _CCCL_VISIBILITY_DEFAULT
+#  define _CCCL_PUBLIC_DEVICE_API _CCCL_DEVICE _CCCL_VISIBILITY_DEFAULT
+#endif // !_CCCL_OS(WINDOWS)
+
+#ifdef _CCCL_DOXYGEN_INVOKED // Only for documentation
+//! If defined, usage of CUDA Dynamic Parallelism is disabled and APIs launching kernels can only be called from the
+//! host
+#  define CCCL_DISABLE_CDP
+#endif // _CCCL_DOXYGEN_INVOKED
+
+#if _CCCL_HAS_CDP()
+// We have CDP, so host and device APIs can call kernels
+#  define _CCCL_CDP_API _CCCL_API
+#else // ^^^ _CCCL_HAS_CDP() ^^^ / vvv !_CCCL_HAS_CDP() vvv
+// We don't have CDP, only host APIs can call kernels
+#  define _CCCL_CDP_API _CCCL_HOST_API
+#endif // ^^^ !_CCCL_HAS_CDP() ^^^
+
+//! _LIBCUDACXX_HIDE_FROM_ABI is for backwards compatibility for external projects.
+//! _CCCL_API and its variants are the preferred way to declare functions
+//! that should be hidden from the ABI.
+//! Defined here to suppress any warnings from the definition
+#define _LIBCUDACXX_HIDE_FROM_ABI _CCCL_API inline
+
+#endif // __CCCL_VISIBILITY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/arithmetic.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/arithmetic.h
new file mode 100644
index 00000000..403b9031
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/arithmetic.h
@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_ARITHMETIC_H
+#define _CUDA_STD___CONCEPTS_ARITHMETIC_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/is_arithmetic.h>
+#include <cuda/std/__type_traits/is_floating_point.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_signed.h>
+#include <cuda/std/__type_traits/is_signed_integer.h>
+#include <cuda/std/__type_traits/is_unsigned_integer.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// [concepts.arithmetic], arithmetic concepts
+
+template <class _Tp>
+_CCCL_CONCEPT integral = is_integral_v<_Tp>;
+
+template <class _Tp>
+_CCCL_CONCEPT signed_integral = integral<_Tp> && is_signed_v<_Tp>;
+
+template <class _Tp>
+_CCCL_CONCEPT unsigned_integral = integral<_Tp> && !signed_integral<_Tp>;
+
+template <class _Tp>
+_CCCL_CONCEPT floating_point = is_floating_point_v<_Tp>;
+
+template <class _Tp>
+_CCCL_CONCEPT __cccl_signed_integer = __cccl_is_signed_integer_v<_Tp>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_ARITHMETIC_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/assignable.h
new file mode 100644
index 00000000..84994a88
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/assignable.h
@@ -0,0 +1,64 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_ASSIGNABLE_H
+#define _CUDA_STD___CONCEPTS_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/common_reference_with.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/make_const_lvalue_ref.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.assignable]
+
+template <class _Lhs, class _Rhs>
+concept assignable_from =
+  is_lvalue_reference_v<_Lhs> && common_reference_with<__make_const_lvalue_ref<_Lhs>, __make_const_lvalue_ref<_Rhs>>
+  && requires(_Lhs __lhs, _Rhs&& __rhs) {
+       { __lhs = ::cuda::std::forward<_Rhs>(__rhs) } -> same_as<_Lhs>;
+     };
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Lhs, class _Rhs>
+_CCCL_CONCEPT_FRAGMENT(
+  __assignable_from_,
+  requires(_Lhs __lhs,
+           _Rhs&& __rhs)(requires(is_lvalue_reference_v<_Lhs>),
+                         requires(common_reference_with<__make_const_lvalue_ref<_Lhs>, __make_const_lvalue_ref<_Rhs>>),
+                         requires(same_as<_Lhs, decltype(__lhs = ::cuda::std::forward<_Rhs>(__rhs))>)));
+
+template <class _Lhs, class _Rhs>
+_CCCL_CONCEPT assignable_from = _CCCL_FRAGMENT(__assignable_from_, _Lhs, _Rhs);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/boolean_testable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/boolean_testable.h
new file mode 100644
index 00000000..7b3df15a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/boolean_testable.h
@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_BOOLEAN_TESTABLE_H
+#define _CUDA_STD___CONCEPTS_BOOLEAN_TESTABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concepts.booleantestable]
+
+template <class _Tp>
+concept __boolean_testable_impl = convertible_to<_Tp, bool>;
+
+template <class _Tp>
+concept __boolean_testable = __boolean_testable_impl<_Tp> && requires(_Tp&& __t) {
+  { !::cuda::std::forward<_Tp>(__t) } -> __boolean_testable_impl;
+};
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp>
+_CCCL_CONCEPT __boolean_testable_impl = convertible_to<_Tp, bool>;
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __boolean_testable_,
+  requires(_Tp&& __t)(requires(__boolean_testable_impl<_Tp>),
+                      requires(__boolean_testable_impl<decltype(!::cuda::std::forward<_Tp>(__t))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __boolean_testable = _CCCL_FRAGMENT(__boolean_testable_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_BOOLEAN_TESTABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/class_or_enum.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/class_or_enum.h
new file mode 100644
index 00000000..526b1e09
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/class_or_enum.h
@@ -0,0 +1,45 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_CLASS_OR_ENUM_H
+#define _CUDA_STD___CONCEPTS_CLASS_OR_ENUM_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/is_class.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_union.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+_CCCL_CONCEPT __class_or_enum = is_class_v<_Tp> || is_union_v<_Tp> || is_enum_v<_Tp>;
+
+// Work around Clang bug https://llvm.org/PR52970
+// TODO: remove this workaround once libc++ no longer has to support Clang 13 (it was fixed in Clang 14).
+template <class _Tp>
+_CCCL_CONCEPT __workaround_52970 = is_class_v<remove_cvref_t<_Tp>> || is_union_v<remove_cvref_t<_Tp>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_CLASS_OR_ENUM_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/common_reference_with.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/common_reference_with.h
new file mode 100644
index 00000000..d680e346
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/common_reference_with.h
@@ -0,0 +1,69 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_COMMON_REFERENCE_WITH_H
+#define _CUDA_STD___CONCEPTS_COMMON_REFERENCE_WITH_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__type_traits/common_reference.h>
+#include <cuda/std/__type_traits/copy_cv.h>
+#include <cuda/std/__type_traits/copy_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.commonref]
+
+template <class _Tp, class _Up>
+concept common_reference_with =
+  same_as<common_reference_t<_Tp, _Up>, common_reference_t<_Up, _Tp>>
+  && convertible_to<_Tp, common_reference_t<_Tp, _Up>> && convertible_to<_Up, common_reference_t<_Tp, _Up>>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(__common_reference_exists_,
+                       requires()(typename(common_reference_t<_Tp, _Up>), typename(common_reference_t<_Up, _Tp>)));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT _Common_reference_exists = _CCCL_FRAGMENT(__common_reference_exists_, _Tp, _Up);
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __common_reference_with_,
+  requires()(requires(_Common_reference_exists<_Tp, _Up>),
+             requires(same_as<common_reference_t<_Tp, _Up>, common_reference_t<_Up, _Tp>>),
+             requires(convertible_to<_Tp, common_reference_t<_Tp, _Up>>),
+             requires(convertible_to<_Up, common_reference_t<_Tp, _Up>>)));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT common_reference_with = _CCCL_FRAGMENT(__common_reference_with_, _Tp, _Up);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_COMMON_REFERENCE_WITH_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/concept_macros.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/concept_macros.h
new file mode 100644
index 00000000..c821e022
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/concept_macros.h
@@ -0,0 +1,389 @@
+//===----------------------------------------------------------------------===//
+//
+// Copyright (c) Facebook, Inc. and its affiliates.
+// Copyright (c) 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___CONCEPTS_CONCEPT_MACROS_H
+#define _CUDA___CONCEPTS_CONCEPT_MACROS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+////////////////////////////////////////////////////////////////////////////////
+// _CCCL_TEMPLATE
+// Usage:
+//   _CCCL_TEMPLATE(typename A, typename _Bp)
+//     _CCCL_REQUIRES( Concept1<A> _CCCL_AND Concept2<_Bp>)
+//   void foo(A a, _Bp b)
+//   {}
+
+// Barebones enable if implementation to use outside of cuda::std
+template <bool>
+struct __cccl_select
+{};
+
+template <>
+struct __cccl_select<true>
+{
+  template <class _Tp>
+  using type = _Tp;
+};
+
+template <bool _Bp, class _Tp = void>
+using __cccl_enable_if_t = typename __cccl_select<_Bp>::template type<_Tp>;
+
+template <class _Tp, bool _Bp>
+using __cccl_requires_t = typename __cccl_select<_Bp>::template type<_Tp>;
+
+#if _CCCL_HAS_CONCEPTS() || defined(_CCCL_DOXYGEN_INVOKED)
+#  define _CCCL_TEMPLATE(...)                template <__VA_ARGS__>
+#  define _CCCL_REQUIRES(...)                requires __VA_ARGS__
+#  define _CCCL_AND                          &&
+#  define _CCCL_TRAILING_REQUIRES_IMPL_(...) requires __VA_ARGS__
+#  define _CCCL_TRAILING_REQUIRES(...)       ->__VA_ARGS__ _CCCL_TRAILING_REQUIRES_IMPL_
+#  define _CCCL_CONCEPT                      concept
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+#  define _CCCL_TEMPLATE(...)                template <__VA_ARGS__
+#  define _CCCL_REQUIRES(...)                , bool __cccl_true_ = true, __cccl_enable_if_t < __VA_ARGS__ && __cccl_true_, int > = 0 >
+#  define _CCCL_AND                          &&__cccl_true_, int > = 0, __cccl_enable_if_t <
+#  define _CCCL_TRAILING_REQUIRES(...)       ->__cccl_requires_t < __VA_ARGS__ _CCCL_TRAILING_REQUIRES_IMPL_
+#  define _CCCL_TRAILING_REQUIRES_IMPL_(...) , __VA_ARGS__ >
+#  define _CCCL_CONCEPT                      inline constexpr bool
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+// The following concepts emulation macros need variable template support
+
+template <class...>
+struct __cccl_tag;
+
+template <class>
+_CCCL_API constexpr bool __cccl_is_true()
+{
+  return true;
+}
+
+#if _CCCL_COMPILER(MSVC)
+template <bool _Bp>
+_CCCL_API inline __cccl_enable_if_t<_Bp> __cccl_requires()
+{}
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+template <bool _Bp, __cccl_enable_if_t<_Bp, int> = 0>
+inline constexpr int __cccl_requires = 0;
+#endif // !_CCCL_COMPILER(MSVC)
+
+template <class _Tp, class... _Args>
+extern _Tp __cccl_make_dependent;
+
+template <class _Impl, class... _Args>
+using __cccl_requires_expr_impl = decltype(__cccl_make_dependent<_Impl, _Args...>);
+
+template <typename _Tp>
+_CCCL_API constexpr void __cccl_unused(_Tp&&) noexcept
+{}
+
+// So that we can refer to the ::cuda::std namespace below
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// We put an alias for ::cuda::std here because of a bug in nvcc <12.2
+// where a requirement such as:
+//
+//  { expression } -> ::concept<type>
+//
+// where ::concept is a fully qualified name, would not compile. The
+// ::cuda::std macro is fully qualified.
+namespace __cccl_unqualified_cuda_std = ::cuda::std; // NOLINT(misc-unused-alias-decls)
+
+#if _CCCL_CUDACC_BELOW(12, 2)
+#  define _CCCL_CONCEPT_VSTD __cccl_unqualified_cuda_std // must not be fully qualified
+#else
+#  define _CCCL_CONCEPT_VSTD ::cuda::std
+#endif
+
+// GCC < 14 can't mangle noexcept expressions. See
+// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70790.
+#if _CCCL_COMPILER(GCC, <, 14)
+#  define _CCCL_HAS_NOEXCEPT_MANGLING() 0
+#else
+#  define _CCCL_HAS_NOEXCEPT_MANGLING() 1
+#endif
+
+// We use this macro to ignore the result of required expressions. It is needed because
+// gcc < 10 complains about ignored [[nodiscard]] expressions when emulating concepts.
+#if _CCCL_COMPILER(GCC, <, 10)
+#  define _CCCL_CONCEPT_IGNORE_RESULT_(...) static_cast<void>(__VA_ARGS__)
+#else
+#  define _CCCL_CONCEPT_IGNORE_RESULT_(...) __VA_ARGS__
+#endif
+
+// The "0" or "1" suffixes indicate whether _REQ is parenthesized or not.
+#define _CCCL_CONCEPT_REQUIREMENT_0(_REQ) _CCCL_PP_SWITCH(_CCCL_CONCEPT_REQUIREMENT, _REQ)
+#define _CCCL_CONCEPT_REQUIREMENT_1(_REQ) _CCCL_CONCEPT_IGNORE_RESULT_ _REQ
+
+// Permissible requirements are of the form (where ... indicates that the pattern can
+// contain commas):
+//
+// - EXPR
+// - (EXPR...)
+// - noexcept(EXPR...)
+// - requires(BOOL-EXPR...)
+// - typename(TYPE...)
+// - _Same_as(TYPE...) EXPR...
+// - _Satisfies(CONCEPT...) EXPR...
+//
+// The last 4 are handled below:
+#define _CCCL_CONCEPT_REQUIREMENT_SWITCH_requires   _CCCL_PP_CASE(_CCCL_SWITCH_REQUIRES)
+#define _CCCL_CONCEPT_REQUIREMENT_SWITCH_noexcept   _CCCL_PP_CASE(_CCCL_SWITCH_NOEXCEPT)
+#define _CCCL_CONCEPT_REQUIREMENT_SWITCH_typename   _CCCL_PP_CASE(_CCCL_SWITCH_TYPENAME)
+#define _CCCL_CONCEPT_REQUIREMENT_SWITCH__Same_as   _CCCL_PP_CASE(_CCCL_SWITCH_SAME_AS)
+#define _CCCL_CONCEPT_REQUIREMENT_SWITCH__Satisfies _CCCL_PP_CASE(_CCCL_SWITCH_SATISFIES)
+
+// Converts "requires(ARGS...)" to "ARGS..."
+#define _CCCL_CONCEPT_EAT_REQUIRES_(...)         _CCCL_PP_CAT(_CCCL_CONCEPT_EAT_REQUIRES_, __VA_ARGS__)
+#define _CCCL_CONCEPT_EAT_REQUIRES_requires(...) __VA_ARGS__
+
+// Converts "noexcept(ARGS...)" to "ARGS..."
+#define _CCCL_CONCEPT_EAT_NOEXCEPT_(...)         _CCCL_PP_CAT(_CCCL_CONCEPT_EAT_NOEXCEPT_, __VA_ARGS__)
+#define _CCCL_CONCEPT_EAT_NOEXCEPT_noexcept(...) __VA_ARGS__
+
+// Converts "typename(TYPE...)" to "TYPE..."
+#define _CCCL_CONCEPT_EAT_TYPENAME_(_REQ)        _CCCL_PP_CAT2(_CCCL_CONCEPT_EAT_TYPENAME_, _REQ)
+#define _CCCL_CONCEPT_EAT_TYPENAME_typename(...) __VA_ARGS__
+
+// Converts "[typename]opt TYPE..." to "typename TYPE..."
+#define _CCCL_CONCEPT_TRY_ADD_TYPENAME_(...)                           _CCCL_PP_SWITCH2(_CCCL_CONCEPT_TRY_ADD_TYPENAME, __VA_ARGS__)
+#define _CCCL_CONCEPT_TRY_ADD_TYPENAME_SWITCH_typename                 _CCCL_PP_CASE(_CCCL_SWITCH_TYPENAME)
+#define _CCCL_CONCEPT_TRY_ADD_TYPENAME_CASE__CCCL_SWITCH_DEFAULT(...)  typename __VA_ARGS__
+#define _CCCL_CONCEPT_TRY_ADD_TYPENAME_CASE__CCCL_SWITCH_TYPENAME(...) __VA_ARGS__
+
+// Converts "_Same_as(TYPE) EXPR..." to "EXPR..."
+#define _CCCL_CONCEPT_EAT_SAME_AS_(...) _CCCL_PP_CAT(_CCCL_CONCEPT_EAT_SAME_AS_, __VA_ARGS__)
+#define _CCCL_CONCEPT_EAT_SAME_AS__Same_as(...)
+
+// Converts "_Same_as(TYPE) EXPR..." to "TYPE" (The ridiculous concatenation of _CCCL with
+// _PP_EXPAND(__VA_ARGS__) is the only way to get MSVC's broken preprocessor to do macro
+// expansion here.)
+#define _CCCL_CONCEPT_GET_TYPE_FROM_SAME_AS_(...) \
+  _CCCL_PP_CAT(_CCCL, _CCCL_PP_EVAL(_CCCL_PP_FIRST, _CCCL_PP_CAT(_CCCL_CONCEPT_GET_TYPE_FROM_SAME_AS_, __VA_ARGS__)))
+#define _CCCL_CONCEPT_GET_TYPE_FROM_SAME_AS__Same_as(...) _PP_EXPAND(__VA_ARGS__),
+
+// Converts "_Satisfies(TYPE) EXPR..." to "EXPR..."
+#define _CCCL_CONCEPT_EAT_SATISFIES_(...) _CCCL_PP_CAT(_CCCL_CONCEPT_EAT_SATISFIES_, __VA_ARGS__)
+#define _CCCL_CONCEPT_EAT_SATISFIES__Satisfies(...)
+
+// Converts "_Satisfies(TYPE) EXPR..." to "TYPE" (The ridiculous concatenation of _CCCL
+// with _PP_EXPAND(__VA_ARGS__) is the only way to get MSVC's broken preprocessor to do macro
+// expansion here.)
+#define _CCCL_CONCEPT_GET_CONCEPT_FROM_SATISFIES_(...) \
+  _CCCL_PP_CAT(_CCCL,                                  \
+               _CCCL_PP_EVAL(_CCCL_PP_FIRST, _CCCL_PP_CAT(_CCCL_CONCEPT_GET_CONCEPT_FROM_SATISFIES_, __VA_ARGS__)))
+#define _CCCL_CONCEPT_GET_CONCEPT_FROM_SATISFIES__Satisfies(...) _PP_EXPAND(__VA_ARGS__),
+
+// Here are the implementations of the internal macros, first for when concepts
+// are available, and then for when they're not.
+#if _CCCL_HAS_CONCEPTS() || defined(_CCCL_DOXYGEN_INVOKED)
+
+// "_CCCL_CONCEPT_FRAGMENT(NAME, ARGS...)(REQS...)" expands into
+// "concept NAME = requires(ARGS...) { _CCCL_CONCEPT_REQUIREMENT_(REQS)... }"
+#  define _CCCL_CONCEPT_FRAGMENT(_NAME, ...)                concept _NAME = _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_##__VA_ARGS__
+#  define _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_requires(...) requires(__VA_ARGS__) _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_
+#  define _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_(...)         {_CCCL_PP_FOR_EACH(_CCCL_CONCEPT_REQUIREMENT_, __VA_ARGS__)}
+
+// Converts "EXPR" to "_CCCL_CONCEPT_REQUIREMENT_0(EXPR)", and
+// "(EXPR)" to "_CCCL_CONCEPT_REQUIREMENT_1((EXPR))"
+#  define _CCCL_CONCEPT_REQUIREMENT_(_REQ)                            \
+    _CCCL_PP_CAT(_CCCL_CONCEPT_REQUIREMENT_, _CCCL_PP_IS_PAREN(_REQ)) \
+    (_REQ);
+
+// The following macros handle the various special forms of requirements:
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_DEFAULT(_REQ)  _REQ
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_REQUIRES(_REQ) requires _CCCL_CONCEPT_EAT_REQUIRES_(_REQ)
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_NOEXCEPT(_REQ) \
+    _CCCL_PP_EXPAND({ _CCCL_CONCEPT_EAT_NOEXCEPT_(_REQ) } noexcept)
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_TYPENAME(_REQ) \
+    _CCCL_CONCEPT_TRY_ADD_TYPENAME_(_CCCL_CONCEPT_EAT_TYPENAME_(_REQ))
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_SAME_AS(_REQ) \
+    {_CCCL_CONCEPT_EAT_SAME_AS_(_REQ)}->_CCCL_CONCEPT_VSTD::same_as<_CCCL_CONCEPT_GET_TYPE_FROM_SAME_AS_(_REQ)>
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_SATISFIES(_REQ) \
+    {_CCCL_CONCEPT_EAT_SATISFIES_(_REQ)}->_CCCL_CONCEPT_GET_CONCEPT_FROM_SATISFIES_(_REQ)
+
+#  define _CCCL_FRAGMENT(_NAME, ...) _NAME<__VA_ARGS__>
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// "_CCCL_CONCEPT_FRAGMENT(Foo, ARGS...)(REQS...)" expands into:
+//
+// _CCCL_API inline auto Foo_CCCL_CONCEPT_FRAGMENT_impl_(ARGS...)
+//   -> __cccl_enable_if_t<
+//        ::__cccl_is_true<decltype(_CCCL_CONCEPT_REQUIREMENT_(REQS)..., void())>()>
+// {}
+//
+// template <class... As>
+// _CCCL_API inline auto Foo_CCCL_CONCEPT_FRAGMENT_(::__cccl_tag<As...>*,
+//                                                  decltype(&Foo_CCCL_CONCEPT_FRAGMENT_impl_<As...>))
+//   -> char(&)[1];
+//
+// template <class... As>
+// _CCCL_API inline auto Foo_CCCL_CONCEPT_FRAGMENT_(...)
+//   -> char(&)[2]
+//
+#  define _CCCL_CONCEPT_FRAGMENT(_NAME, ...)                                                                        \
+    _CCCL_API inline auto _NAME##_CCCL_CONCEPT_FRAGMENT_impl_ _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_##__VA_ARGS__> {} \
+    template <class... _As>                                                                                         \
+    _CCCL_API inline auto _NAME##_CCCL_CONCEPT_FRAGMENT_(                                                           \
+      ::__cccl_tag<_As...>*, decltype(&_NAME##_CCCL_CONCEPT_FRAGMENT_impl_<_As...>)) -> char (&)[1];                \
+    _CCCL_API inline auto _NAME##_CCCL_CONCEPT_FRAGMENT_(...) -> char (&)[2]
+#  define _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_requires(...) \
+    (__VA_ARGS__)->__cccl_enable_if_t < _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_IMPL_
+#  define _CCCL_CONCEPT_FRAGMENT_REQUIREMENTS_IMPL_(...) \
+    ::__cccl_is_true<decltype(_CCCL_PP_FOR_EACH(_CCCL_CONCEPT_REQUIREMENT_, __VA_ARGS__) void())>()
+
+// Called with each individual requirement in the list of requirements
+#  define _CCCL_CONCEPT_REQUIREMENT_(_REQ) \
+    void(), _CCCL_PP_CAT(_CCCL_CONCEPT_REQUIREMENT_, _CCCL_PP_IS_PAREN(_REQ))(_REQ),
+
+// The following macros handle the various special forms of requirements:
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_DEFAULT(_REQ) _CCCL_CONCEPT_IGNORE_RESULT_(_REQ)
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_REQUIRES(_REQ) \
+    ::__cccl_requires<_CCCL_CONCEPT_EAT_REQUIRES_(_REQ)>
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_NOEXCEPT(_REQ) _CCCL_CONCEPT_NOEXCEPT_REQUIREMENT_(_REQ)
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_TYPENAME(_REQ) \
+    static_cast<::__cccl_tag<_CCCL_CONCEPT_EAT_TYPENAME_(_REQ)>*>(nullptr)
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_SAME_AS(_REQ) \
+    ::__cccl_requires<::cuda::std::same_as<_CCCL_CONCEPT_SAME_AS_REQUIREMENT_(_REQ)>>
+#  define _CCCL_CONCEPT_REQUIREMENT_CASE__CCCL_SWITCH_SATISFIES(_REQ)                                                  \
+    ::__cccl_requires < _CCCL_CONCEPT_GET_CONCEPT_FROM_SATISFIES_(_REQ) < decltype(_CCCL_CONCEPT_EAT_SATISFIES_(_REQ)) \
+      >>
+
+// Converts "_Same_as(TYPE) EXPR..." to "TYPE, decltype(EXPR...)"
+#  define _CCCL_CONCEPT_SAME_AS_REQUIREMENT_(_REQ) \
+    _CCCL_CONCEPT_GET_TYPE_FROM_SAME_AS_(_REQ), decltype(_CCCL_CONCEPT_EAT_SAME_AS_(_REQ))
+
+#  if _CCCL_HAS_NOEXCEPT_MANGLING()
+// Converts "noexcept(EXPR)" to "::__cccl_requires<noexcept(EXPR)>"
+#    define _CCCL_CONCEPT_NOEXCEPT_REQUIREMENT_(_REQ) ::__cccl_requires<_REQ>
+#  else
+// If the compiler cannot mangle noexcept expressions, just check that the expression is
+// well-formed. This converts "noexcept(EXPR)" to "static_cast<void>(EXPR)"
+#    define _CCCL_CONCEPT_NOEXCEPT_REQUIREMENT_(_REQ) _CCCL_CONCEPT_IGNORE_RESULT_(_CCCL_CONCEPT_EAT_NOEXCEPT_(_REQ))
+#  endif
+
+// "_CCCL_FRAGMENT(Foo, Args...)" expands to
+// "(1 == sizeof(Foo_CCCL_CONCEPT_FRAGMENT_(static_cast<::__cccl_tag<Args...>*>(nullptr), nullptr)))"
+#  define _CCCL_FRAGMENT(_NAME, ...) \
+    (1 == sizeof(_NAME##_CCCL_CONCEPT_FRAGMENT_(static_cast<::__cccl_tag<__VA_ARGS__>*>(nullptr), nullptr)))
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+////////////////////////////////////////////////////////////////////////////////
+// _CCCL_REQUIRES_EXPR
+// Usage:
+//   template <typename T>
+//   _CCCL_CONCEPT equality_comparable =
+//     _CCCL_REQUIRES_EXPR((T), T const& lhs, T const& rhs) (
+//       lhs == rhs,
+//       lhs != rhs
+//     );
+//
+// Can only be used as the last requirement in a concept definition.
+#if _CCCL_HAS_CONCEPTS() || defined(_CCCL_DOXYGEN_INVOKED)
+
+#  define _CCCL_REQUIRES_EXPR(_TY, ...)  requires(__VA_ARGS__) _CCCL_REQUIRES_EXPR_IMPL_
+#  define _CCCL_REQUIRES_EXPR_IMPL_(...) {_CCCL_PP_FOR_EACH(_CCCL_CONCEPT_REQUIREMENT_, __VA_ARGS__)}
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+#  define _CCCL_REQUIRES_EXPR(_TY, ...) _CCCL_REQUIRES_EXPR_IMPL(_TY, _CCCL_REQUIRES_EXPR_ID(_TY), __VA_ARGS__)
+#  define _CCCL_REQUIRES_EXPR_IMPL(_TY, _ID, ...)                                                                    \
+    ::__cccl_requires_expr_impl<                                                                                     \
+      struct _CCCL_PP_CAT(__cccl_requires_expr_detail_, _ID) _CCCL_REQUIRES_EXPR_TPARAM_REFS                         \
+        _TY>::__cccl_is_satisfied(static_cast<::__cccl_tag<void _CCCL_REQUIRES_EXPR_TPARAM_REFS _TY>*>(nullptr), 0); \
+    struct _CCCL_PP_CAT(__cccl_requires_expr_detail_, _ID)                                                           \
+    {                                                                                                                \
+      using __cccl_self_t = _CCCL_PP_CAT(__cccl_requires_expr_detail_, _ID);                                         \
+      template <class _CCCL_REQUIRES_EXPR_TPARAM_DEFNS _TY>                                                          \
+      _CCCL_API inline static auto __cccl_well_formed(__VA_ARGS__) _CCCL_REQUIRES_EXPR_REQUIREMENTS_
+
+// Expands "T1, T2, variadic T3" to ", class T1, class T2, class... T3"
+#  define _CCCL_REQUIRES_EXPR_TPARAM_DEFNS(...)                          _CCCL_PP_FOR_EACH(_CCCL_REQUIRES_EXPR_TPARAM_DEFN, __VA_ARGS__)
+
+// Expands "TY" to ", class TY" and "variadic TY" to ", class... TY"
+#  define _CCCL_REQUIRES_EXPR_TPARAM_DEFN(_TY)                           , _CCCL_PP_SWITCH2(_CCCL_REQUIRES_EXPR_TPARAM_DEFN, _TY)
+#  define _CCCL_REQUIRES_EXPR_TPARAM_DEFN_SWITCH_variadic                _CCCL_PP_CASE(_CCCL_SWITCH_VARIADIC)
+#  define _CCCL_REQUIRES_EXPR_TPARAM_DEFN_CASE__CCCL_SWITCH_DEFAULT(_TY) class _TY
+#  define _CCCL_REQUIRES_EXPR_TPARAM_DEFN_CASE__CCCL_SWITCH_VARIADIC(_TY) \
+    class... _CCCL_PP_CAT(_CCCL_REQUIRES_EXPR_EAT_VARIADIC_, _TY)
+
+// Expands "T1, T2, variadic T3" to ", T1, T2, T3..."
+#  define _CCCL_REQUIRES_EXPR_TPARAM_REFS(...)                          _CCCL_PP_FOR_EACH(_CCCL_REQUIRES_EXPR_TPARAM_REF, __VA_ARGS__)
+
+// Expands "TY" to ", TY" and "variadic TY" to ", TY..."
+#  define _CCCL_REQUIRES_EXPR_TPARAM_REF(_TY)                           , _CCCL_PP_SWITCH2(_CCCL_REQUIRES_EXPR_TPARAM_REF, _TY)
+#  define _CCCL_REQUIRES_EXPR_TPARAM_REF_SWITCH_variadic                _CCCL_PP_CASE(_CCCL_SWITCH_VARIADIC)
+#  define _CCCL_REQUIRES_EXPR_TPARAM_REF_CASE__CCCL_SWITCH_DEFAULT(_TY) _TY
+#  define _CCCL_REQUIRES_EXPR_TPARAM_REF_CASE__CCCL_SWITCH_VARIADIC(_TY) \
+    _CCCL_PP_CAT(_CCCL_REQUIRES_EXPR_EAT_VARIADIC_, _TY)...
+
+// NVRTC does not support __COUNTER__ so we need a better way of defining unique identifiers
+#  if _CCCL_COMPILER(NVRTC)
+
+// Expands ((Ty...), Ty...) into _CCCL_REQUIRES_EXPR_ID_NO_PAREN(Ty...)
+#    define _CCCL_REQUIRES_EXPR_ID(_TY, ...) _CCCL_REQUIRES_EXPR_ID_NO_PAREN _TY
+
+// Expands "T1, T2, variadic T3" to "T1_T2_T3_##__LINE__"
+#    define _CCCL_REQUIRES_EXPR_ID_NO_PAREN(...) \
+      _CCCL_REQUIRES_EXPR_ID_CONCAT_ALL(_CCCL_PP_FOR_EACH(_CCCL_REQUIRES_EXPR_ID_IMPL, __VA_ARGS__), _CCCL_COUNTER())
+
+// Expands "T1, T2, T3" to "T1T2T3"
+#    define _CCCL_REQUIRES_EXPR_ID_CONCAT_ALL_IMPL(_0, _1, _2, _3, _4, _5, _6, _7, _8, _9, ...) \
+      _0##_1##_2##_3##_4##_5##_6##_7##_8##_9
+#    define _CCCL_REQUIRES_EXPR_ID_CONCAT_ALL(...) \
+      _CCCL_PP_EVAL(_CCCL_REQUIRES_EXPR_ID_CONCAT_ALL_IMPL, __VA_ARGS__, , , , , , , , , )
+
+// Expands "TY" to "TY" and "variadic TY" to "TY"
+#    define _CCCL_REQUIRES_EXPR_ID_IMPL(_TY)                           , _CCCL_PP_SWITCH2(_CCCL_REQUIRES_EXPR_ID_IMPL, _TY)
+#    define _CCCL_REQUIRES_EXPR_ID_IMPL_SWITCH_variadic                _CCCL_PP_CASE(_CCCL_SWITCH_VARIADIC)
+#    define _CCCL_REQUIRES_EXPR_ID_IMPL_CASE__CCCL_SWITCH_DEFAULT(_TY) _TY
+#    define _CCCL_REQUIRES_EXPR_ID_IMPL_CASE__CCCL_SWITCH_VARIADIC(_TY) \
+      _CCCL_PP_CAT(_CCCL_REQUIRES_EXPR_EAT_VARIADIC_, _TY)
+
+#  else // ^^^ _CCCL_COMPILER(NVRTC) ^^^^/ vvv !_CCCL_COMPILER(NVRTC)
+#    define _CCCL_REQUIRES_EXPR_ID(...) _CCCL_COUNTER()
+#  endif // !_CCCL_COMPILER(NVRTC)
+
+#  define _CCCL_REQUIRES_EXPR_EAT_VARIADIC_variadic
+
+#  define _CCCL_REQUIRES_EXPR_REQUIREMENTS_(...)                                              \
+    ->decltype(_CCCL_PP_FOR_EACH(_CCCL_CONCEPT_REQUIREMENT_, __VA_ARGS__) void()) {}          \
+    template <class... _Args, class = decltype(&__cccl_self_t::__cccl_well_formed<_Args...>)> \
+    _CCCL_API static constexpr bool __cccl_is_satisfied(::__cccl_tag<_Args...>*, int)         \
+    {                                                                                         \
+      return true;                                                                            \
+    }                                                                                         \
+    _CCCL_API static constexpr bool __cccl_is_satisfied(void*, long)                          \
+    {                                                                                         \
+      return false;                                                                           \
+    }                                                                                         \
+    }
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif //_CUDA___CONCEPTS_CONCEPT_MACROS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/constructible.h
new file mode 100644
index 00000000..8eb98baf
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/constructible.h
@@ -0,0 +1,174 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_CONSTRUCTIBLE_H
+#define _CUDA_STD___CONCEPTS_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__concepts/destructible.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/is_callable.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.constructible]
+template <class _Tp, class... _Args>
+concept constructible_from = destructible<_Tp> && is_constructible_v<_Tp, _Args...>;
+
+// [concept.default.init]
+template <class _Tp>
+concept __default_initializable = requires { ::new _Tp; };
+
+template <class _Tp>
+concept default_initializable = constructible_from<_Tp> && requires { _Tp{}; } && __default_initializable<_Tp>;
+
+// [concept.moveconstructible]
+template <class _Tp>
+concept move_constructible = constructible_from<_Tp, _Tp> && convertible_to<_Tp, _Tp>;
+
+// [concept.copyconstructible]
+template <class _Tp>
+concept copy_constructible =
+  move_constructible<_Tp> && constructible_from<_Tp, _Tp&> && convertible_to<_Tp&, _Tp>
+  && constructible_from<_Tp, const _Tp&> && convertible_to<const _Tp&, _Tp> && constructible_from<_Tp, const _Tp>
+  && convertible_to<const _Tp, _Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT_FRAGMENT(__constructible_from_,
+                       requires()(requires(destructible<_Tp>), requires(is_constructible_v<_Tp, _Args...>)));
+
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT constructible_from = _CCCL_FRAGMENT(__constructible_from_, _Tp, _Args...);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__default_initializable_, requires()((::new _Tp)));
+
+template <class _Tp>
+_CCCL_CONCEPT __default_initializable = _CCCL_FRAGMENT(__default_initializable_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(_Default_initializable_,
+                       requires(_Tp = _Tp{})(requires(constructible_from<_Tp>), requires(__default_initializable<_Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT default_initializable = _CCCL_FRAGMENT(_Default_initializable_, _Tp);
+
+// [concept.moveconstructible]
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__move_constructible_,
+                       requires()(requires(constructible_from<_Tp, _Tp>), requires(convertible_to<_Tp, _Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT move_constructible = _CCCL_FRAGMENT(__move_constructible_, _Tp);
+
+// [concept.copyconstructible]
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __copy_constructible_,
+  requires()(
+    requires(move_constructible<_Tp>),
+    requires(constructible_from<_Tp, add_lvalue_reference_t<_Tp>>&& convertible_to<add_lvalue_reference_t<_Tp>, _Tp>),
+    requires(constructible_from<_Tp, const add_lvalue_reference_t<_Tp>>&&
+               convertible_to<const add_lvalue_reference_t<_Tp>, _Tp>),
+    requires(constructible_from<_Tp, const _Tp>&& convertible_to<const _Tp, _Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT copy_constructible = _CCCL_FRAGMENT(__copy_constructible_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+//! The code below provides the following concepts in the ::cuda:: namespace:
+//!
+//! - `__list_initializable_from`
+//! - `__nothrow_list_initializable_from`
+//! - `__initializable_from`
+//! - `__nothrow_initializable_from`
+//! - `__emplaceable_from`
+//! - `__nothrow_emplaceable_from`
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+// constructible_from using list initialization syntax.
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT __list_initializable_from =
+  _CCCL_REQUIRES_EXPR((_Tp, variadic _Args), _Args&&... __args)(_Tp{static_cast<_Args&&>(__args)...});
+
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT __nothrow_list_initializable_from =
+  _CCCL_REQUIRES_EXPR((_Tp, variadic _Args), _Args&&... __args)(noexcept(_Tp{static_cast<_Args&&>(__args)...}));
+
+//! Constructible from arguments using either direct non-list initialization or direct
+//! list initialization.
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT __initializable_from =
+  ::cuda::std::constructible_from<_Tp, _Args...> || __list_initializable_from<_Tp, _Args...>;
+
+template <class _Tp, class... _Args>
+_CCCL_CONCEPT __nothrow_initializable_from =
+  __initializable_from<_Tp, _Args...>
+  && (::cuda::std::constructible_from<_Tp, _Args...>
+        ? ::cuda::std::is_nothrow_constructible_v<_Tp, _Args...>
+        : __nothrow_list_initializable_from<_Tp, _Args...>);
+
+#if !_CCCL_COMPILER(MSVC) && !_CCCL_CUDA_COMPILER(NVCC, <, 12, 9)
+
+//! Constructible with direct non-list initialization syntax from the result of
+//! a function call expression (often useful for immovable types).
+template <class _Tp, class _Fn, class... _Args>
+_CCCL_CONCEPT __emplaceable_from = _CCCL_REQUIRES_EXPR((_Tp, _Fn, variadic _Args), _Fn&& __fn, _Args&&... __args)(
+  _Tp(static_cast<_Fn&&>(__fn)(static_cast<_Args&&>(__args)...)));
+
+template <class _Tp, class _Fn, class... _Args>
+_CCCL_CONCEPT __nothrow_emplaceable_from =
+  _CCCL_REQUIRES_EXPR((_Tp, _Fn, variadic _Args), _Fn&& __fn, _Args&&... __args)(
+    noexcept(_Tp(static_cast<_Fn&&>(__fn)(static_cast<_Args&&>(__args)...))));
+
+#else // ^^^ !_CCCL_COMPILER(MSVC) ^^^ / vvv _CCCL_COMPILER(MSVC) vvv
+
+//! Constructible with direct non-list initialization syntax from the result of
+//! a function call expression (often useful for immovable types). MSVC cannot
+//! use the above formulation because it has poor support for deferred materialization
+//! of temporary object (aka, guaranteed copy elision).
+template <class _Tp, class _Fn, class... _Args>
+_CCCL_CONCEPT __emplaceable_from = _CCCL_REQUIRES_EXPR((_Tp, _Fn, variadic _Args), _Fn&& __fn, _Args&&... __args)(
+  _Same_as(_Tp) static_cast<_Fn&&>(__fn)(static_cast<_Args&&>(__args)...));
+
+template <class _Tp, class _Fn, class... _Args>
+_CCCL_CONCEPT __nothrow_emplaceable_from =
+  __emplaceable_from<_Tp, _Fn, _Args...> && ::cuda::std::__is_nothrow_callable_v<_Fn, _Args...>;
+
+#endif // ^^^ _CCCL_COMPILER(MSVC) ^^^
+
+_CCCL_END_NAMESPACE_CUDA
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/convertible_to.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/convertible_to.h
new file mode 100644
index 00000000..4e41115a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/convertible_to.h
@@ -0,0 +1,70 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_CONVERTIBLE_TO_H
+#define _CUDA_STD___CONCEPTS_CONVERTIBLE_TO_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// [concept.convertible]
+
+#if _CCCL_HAS_CONCEPTS()
+
+template <class _From, class _To>
+concept convertible_to = is_convertible_v<_From, _To> && requires { static_cast<_To>(::cuda::std::declval<_From>()); };
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+#  if _CCCL_COMPILER(MSVC)
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(1211) // nonstandard cast to array type ignored
+#  endif // _CCCL_COMPILER(MSVC)
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(171) // invalid type conversion, e.g. [with _From=int **, _To=const int *const *]
+
+// We cannot put this conversion check with the other constraint, as types with deleted operator will break here
+template <class _From, class _To>
+_CCCL_CONCEPT_FRAGMENT(__test_conversion_, requires()(static_cast<_To>(::cuda::std::declval<_From>())));
+
+template <class _From, class _To>
+_CCCL_CONCEPT __test_conversion = _CCCL_FRAGMENT(__test_conversion_, _From, _To);
+
+template <class _From, class _To>
+_CCCL_CONCEPT_FRAGMENT(__convertible_to_,
+                       requires()(requires(is_convertible_v<_From, _To>), requires(__test_conversion<_From, _To>)));
+
+template <class _From, class _To>
+_CCCL_CONCEPT convertible_to = _CCCL_FRAGMENT(__convertible_to_, _From, _To);
+
+#  if _CCCL_COMPILER(MSVC)
+_CCCL_END_NV_DIAG_SUPPRESS() // nonstandard cast to array type ignored
+#  endif // _CCCL_COMPILER(MSVC)
+_CCCL_END_NV_DIAG_SUPPRESS() // invalid type conversion, e.g. [with _From=int **, _To=const int *const *]
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_CONVERTIBLE_TO_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/copyable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/copyable.h
new file mode 100644
index 00000000..b1a08668
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/copyable.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_COPYABLE_H
+#define _CUDA_STD___CONCEPTS_COPYABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/assignable.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/movable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concepts.object]
+
+template <class _Tp>
+concept copyable = copy_constructible<_Tp> && movable<_Tp> && assignable_from<_Tp&, _Tp&>
+                && assignable_from<_Tp&, const _Tp&> && assignable_from<_Tp&, const _Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __copyable_,
+  requires()(requires(copy_constructible<_Tp>),
+             requires(movable<_Tp>),
+             requires(assignable_from<_Tp&, _Tp&>),
+             requires(assignable_from<_Tp&, const _Tp&>),
+             requires(assignable_from<_Tp&, const _Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT copyable = _CCCL_FRAGMENT(__copyable_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_COPYABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/derived_from.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/derived_from.h
new file mode 100644
index 00000000..562e35d0
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/derived_from.h
@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_DERIVED_FROM_H
+#define _CUDA_STD___CONCEPTS_DERIVED_FROM_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/add_pointer.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.derived]
+
+template <class _Dp, class _Bp>
+concept derived_from = is_base_of_v<_Bp, _Dp> && is_convertible_v<const volatile _Dp*, const volatile _Bp*>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Dp, class _Bp>
+_CCCL_CONCEPT_FRAGMENT(
+  __derived_from_,
+  requires()(requires(is_base_of_v<_Bp, _Dp>),
+             requires(is_convertible_v<add_pointer_t<const volatile _Dp>, add_pointer_t<const volatile _Bp>>)));
+
+template <class _Dp, class _Bp>
+_CCCL_CONCEPT derived_from = _CCCL_FRAGMENT(__derived_from_, _Dp, _Bp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_DERIVED_FROM_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/destructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/destructible.h
new file mode 100644
index 00000000..9299f48c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/destructible.h
@@ -0,0 +1,76 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_DESTRUCTIBLE_H
+#define _CUDA_STD___CONCEPTS_DESTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_destructible.h>
+#include <cuda/std/__type_traits/is_nothrow_destructible.h>
+#include <cuda/std/__type_traits/is_object.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_COMPILER(MSVC)
+
+template <class _Tp>
+_CCCL_CONCEPT destructible = __is_nothrow_destructible(_Tp);
+
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+
+template <class _Tp, class = void, class = void>
+inline constexpr bool __destructible_impl = false;
+
+template <class _Tp>
+inline constexpr bool __destructible_impl<_Tp,
+                                          enable_if_t<is_object_v<_Tp>>,
+#  if _CCCL_COMPILER(GCC)
+                                          enable_if_t<is_destructible_v<_Tp>>>
+#  else // ^^^ _CCCL_COMPILER(GCC) ^^^ / vvv !_CCCL_COMPILER(GCC) vvv
+                                          void_t<decltype(::cuda::std::declval<_Tp>().~_Tp())>>
+#  endif // !_CCCL_COMPILER(GCC)
+  = noexcept(::cuda::std::declval<_Tp>().~_Tp());
+
+template <class _Tp>
+inline constexpr bool __destructible = __destructible_impl<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __destructible<_Tp&> = true;
+
+template <class _Tp>
+inline constexpr bool __destructible<_Tp&&> = true;
+
+template <class _Tp, size_t _Nm>
+inline constexpr bool __destructible<_Tp[_Nm]> = __destructible<_Tp>;
+
+template <class _Tp>
+_CCCL_CONCEPT destructible = __destructible<_Tp>;
+
+#endif // !_CCCL_COMPILER(MSVC)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_DESTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/equality_comparable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/equality_comparable.h
new file mode 100644
index 00000000..fb4c7a65
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/equality_comparable.h
@@ -0,0 +1,98 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_EQUALITY_COMPARABLE_H
+#define _CUDA_STD___CONCEPTS_EQUALITY_COMPARABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/boolean_testable.h>
+#include <cuda/std/__concepts/common_reference_with.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/common_reference.h>
+#include <cuda/std/__type_traits/is_comparable.h>
+#include <cuda/std/__type_traits/make_const_lvalue_ref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.equalitycomparable]
+
+template <class _Tp, class _Up>
+concept __weakly_equality_comparable_with =
+  requires(__make_const_lvalue_ref<_Tp> __t, __make_const_lvalue_ref<_Up> __u) {
+    { __t == __u } -> __boolean_testable;
+    { __t != __u } -> __boolean_testable;
+    { __u == __t } -> __boolean_testable;
+    { __u != __t } -> __boolean_testable;
+  };
+
+template <class _Tp>
+concept equality_comparable = __weakly_equality_comparable_with<_Tp, _Tp>;
+
+template <class _Tp, class _Up>
+concept equality_comparable_with =
+  equality_comparable<_Tp> && equality_comparable<_Up>
+  && common_reference_with<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>
+  && equality_comparable<common_reference_t<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>>
+  && __weakly_equality_comparable_with<_Tp, _Up>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp>
+_CCCL_CONCEPT _With_lvalue_reference = _CCCL_REQUIRES_EXPR((_Tp))(typename(__make_const_lvalue_ref<_Tp>));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __weakly_equality_comparable_with_,
+  requires(__make_const_lvalue_ref<_Tp> __t, __make_const_lvalue_ref<_Up> __u)(
+    requires(_With_lvalue_reference<_Tp>),
+    requires(_With_lvalue_reference<_Up>),
+    _Satisfies(__boolean_testable) __t == __u,
+    _Satisfies(__boolean_testable) __t != __u,
+    _Satisfies(__boolean_testable) __u == __t,
+    _Satisfies(__boolean_testable) __u != __t));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT __weakly_equality_comparable_with = _CCCL_FRAGMENT(__weakly_equality_comparable_with_, _Tp, _Up);
+
+template <class _Tp>
+_CCCL_CONCEPT equality_comparable = __weakly_equality_comparable_with<_Tp, _Tp>;
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __equality_comparable_with_,
+  requires()(
+    requires(equality_comparable<_Tp>),
+    requires(equality_comparable<_Up>),
+    requires(common_reference_with<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>),
+    requires(equality_comparable<common_reference_t<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>>),
+    requires(__weakly_equality_comparable_with<_Tp, _Up>)));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT equality_comparable_with = _CCCL_FRAGMENT(__equality_comparable_with_, _Tp, _Up);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_EQUALITY_COMPARABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/invocable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/invocable.h
new file mode 100644
index 00000000..889147f9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/invocable.h
@@ -0,0 +1,80 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_INVOCABLE_H
+#define _CUDA_STD___CONCEPTS_INVOCABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.invocable]
+
+template <class _Fn, class... _Args>
+concept invocable = requires(_Fn&& __fn, _Args&&... __args) {
+  ::cuda::std::invoke(::cuda::std::forward<_Fn>(__fn), ::cuda::std::forward<_Args>(__args)...); // not required to be
+                                                                                                // equality preserving
+};
+
+// [concept.regular.invocable]
+
+template <class _Fn, class... _Args>
+concept regular_invocable = invocable<_Fn, _Args...>;
+
+template <class _Fun, class... _Args>
+concept __invoke_constructible = requires(_Fun&& __fun, _Args&&... __args) {
+  static_cast<remove_cvref_t<invoke_result_t<_Fun, _Args...>>>(
+    ::cuda::std::invoke(::cuda::std::forward<_Fun>(__fun), ::cuda::std::forward<_Args>(__args)...));
+};
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Fn, class... _Args>
+_CCCL_CONCEPT_FRAGMENT(_Invocable_,
+                       requires(_Fn&& __fn, _Args&&... __args)((::cuda::std::invoke(
+                         ::cuda::std::forward<_Fn>(__fn), ::cuda::std::forward<_Args>(__args)...))));
+
+template <class _Fn, class... _Args>
+_CCCL_CONCEPT invocable = _CCCL_FRAGMENT(_Invocable_, _Fn, _Args...);
+
+template <class _Fn, class... _Args>
+_CCCL_CONCEPT regular_invocable = invocable<_Fn, _Args...>;
+
+template <class _Fun, class... _Args>
+_CCCL_CONCEPT_FRAGMENT(
+  __invoke_constructible_,
+  requires(_Fun&& __fun, _Args&&... __args)((static_cast<remove_cvref_t<invoke_result_t<_Fun, _Args...>>>(
+    ::cuda::std::invoke(::cuda::std::forward<_Fun>(__fun), ::cuda::std::forward<_Args>(__args)...)))));
+template <class _Fun, class... _Args>
+_CCCL_CONCEPT __invoke_constructible = _CCCL_FRAGMENT(__invoke_constructible_, _Fun, _Args...);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_INVOCABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/movable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/movable.h
new file mode 100644
index 00000000..fab06e6e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/movable.h
@@ -0,0 +1,58 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_MOVABLE_H
+#define _CUDA_STD___CONCEPTS_MOVABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/assignable.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/swappable.h>
+#include <cuda/std/__type_traits/is_object.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+template <class _Tp>
+concept movable = is_object_v<_Tp> && move_constructible<_Tp> && assignable_from<_Tp&, _Tp> && swappable<_Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [concepts.object]
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  _Movable_,
+  requires()(requires(is_object_v<_Tp>),
+             requires(move_constructible<_Tp>),
+             requires(assignable_from<_Tp&, _Tp>),
+             requires(swappable<_Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT movable = _CCCL_FRAGMENT(_Movable_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_MOVABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/predicate.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/predicate.h
new file mode 100644
index 00000000..a481790c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/predicate.h
@@ -0,0 +1,54 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_PREDICATE_H
+#define _CUDA_STD___CONCEPTS_PREDICATE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/boolean_testable.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/invocable.h>
+#include <cuda/std/__functional/invoke.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+template <class _Fn, class... _Args>
+concept predicate = regular_invocable<_Fn, _Args...> && __boolean_testable<invoke_result_t<_Fn, _Args...>>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [concept.predicate]
+template <class _Fn, class... _Args>
+_CCCL_CONCEPT_FRAGMENT(
+  _Predicate_,
+  requires()(requires(regular_invocable<_Fn, _Args...>), requires(__boolean_testable<invoke_result_t<_Fn, _Args...>>)));
+
+template <class _Fn, class... _Args>
+_CCCL_CONCEPT predicate = _CCCL_FRAGMENT(_Predicate_, _Fn, _Args...);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_PREDICATE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/regular.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/regular.h
new file mode 100644
index 00000000..c912eb73
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/regular.h
@@ -0,0 +1,54 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_REGULAR_H
+#define _CUDA_STD___CONCEPTS_REGULAR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/equality_comparable.h>
+#include <cuda/std/__concepts/semiregular.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.object]
+
+template <class _Tp>
+concept regular = semiregular<_Tp> && equality_comparable<_Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [concept.object]
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__regular_, requires()(requires(semiregular<_Tp>), requires(equality_comparable<_Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT regular = _CCCL_FRAGMENT(__regular_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_REGULAR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/relation.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/relation.h
new file mode 100644
index 00000000..fc2862d0
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/relation.h
@@ -0,0 +1,77 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_RELATION_H
+#define _CUDA_STD___CONCEPTS_RELATION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/predicate.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.relation]
+
+template <class _Rp, class _Tp, class _Up>
+concept relation =
+  predicate<_Rp, _Tp, _Tp> && predicate<_Rp, _Up, _Up> && predicate<_Rp, _Tp, _Up> && predicate<_Rp, _Up, _Tp>;
+
+// [concept.equiv]
+
+template <class _Rp, class _Tp, class _Up>
+concept equivalence_relation = relation<_Rp, _Tp, _Up>;
+
+// [concept.strictweakorder]
+
+template <class _Rp, class _Tp, class _Up>
+concept strict_weak_order = relation<_Rp, _Tp, _Up>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Rp, class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __relation_,
+  requires()(requires(predicate<_Rp, _Tp, _Tp>),
+             requires(predicate<_Rp, _Up, _Up>),
+             requires(predicate<_Rp, _Tp, _Up>),
+             requires(predicate<_Rp, _Up, _Tp>)));
+
+template <class _Rp, class _Tp, class _Up>
+_CCCL_CONCEPT relation = _CCCL_FRAGMENT(__relation_, _Rp, _Tp, _Up);
+
+// [concept.equiv]
+
+template <class _Rp, class _Tp, class _Up>
+_CCCL_CONCEPT equivalence_relation = relation<_Rp, _Tp, _Up>;
+
+// [concept.strictweakorder]
+
+template <class _Rp, class _Tp, class _Up>
+_CCCL_CONCEPT strict_weak_order = relation<_Rp, _Tp, _Up>;
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_RELATION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/same_as.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/same_as.h
new file mode 100644
index 00000000..30e09397
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/same_as.h
@@ -0,0 +1,39 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_SAME_AS_H
+#define _CUDA_STD___CONCEPTS_SAME_AS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/is_same.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// [concept.same]
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT same_as = is_same_v<_Tp, _Up> && is_same_v<_Up, _Tp>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_SAME_AS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/semiregular.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/semiregular.h
new file mode 100644
index 00000000..091c9561
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/semiregular.h
@@ -0,0 +1,54 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_SEMIREGULAR_H
+#define _CUDA_STD___CONCEPTS_SEMIREGULAR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/copyable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.object]
+
+template <class _Tp>
+concept semiregular = copyable<_Tp> && default_initializable<_Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [concept.object]
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__semiregular_, requires()(requires(copyable<_Tp>), requires(default_initializable<_Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT semiregular = _CCCL_FRAGMENT(__semiregular_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_SEMIREGULAR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/swappable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/swappable.h
new file mode 100644
index 00000000..76574c67
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/swappable.h
@@ -0,0 +1,209 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_SWAPPABLE_H
+#define _CUDA_STD___CONCEPTS_SWAPPABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/assignable.h>
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__concepts/common_reference_with.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__type_traits/extent.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_nothrow_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/type_identity.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/exchange.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_COMPILER(MSVC)
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(461) // nonstandard cast to array type ignored
+#endif // _CCCL_COMPILER(MSVC)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+// [concept.swappable]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__swap)
+
+template <class _Tp>
+void swap(_Tp&, _Tp&) = delete;
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp, class _Up>
+concept __unqualified_swappable_with =
+  (__class_or_enum<remove_cvref_t<_Tp>> || __class_or_enum<remove_cvref_t<_Up>>)
+  && requires(_Tp&& __t, _Up&& __u) { swap(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u)); };
+
+template <class _Tp>
+concept __exchangeable =
+  !__unqualified_swappable_with<_Tp&, _Tp&> && move_constructible<_Tp> && assignable_from<_Tp&, _Tp>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __unqualified_swappable_with_,
+  requires(_Tp&& __t, _Up&& __u)((swap(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u)))));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT __unqualified_swappable_with = _CCCL_FRAGMENT(__unqualified_swappable_with_, _Tp, _Up);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__exchangeable_,
+                       requires()(requires(!__unqualified_swappable_with<_Tp&, _Tp&>),
+                                  requires(move_constructible<_Tp>),
+                                  requires(assignable_from<_Tp&, _Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __exchangeable = _CCCL_FRAGMENT(__exchangeable_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+#if _CCCL_HAS_CONCEPTS() && !_CCCL_COMPILER(NVHPC) // nvbug4051640
+struct __fn;
+
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(2642)
+template <class _Tp, class _Up, size_t _Size>
+concept __swappable_arrays =
+  !__unqualified_swappable_with<_Tp (&)[_Size], _Up (&)[_Size]> && extent_v<_Tp> == extent_v<_Up>
+  && requires(_Tp (&__t)[_Size], _Up (&__u)[_Size], const __fn& __swap) { __swap(__t[0], __u[0]); };
+_CCCL_END_NV_DIAG_SUPPRESS()
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() && !_CCCL_COMPILER(NVHPC) ^^^ / vvv !_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVHPC) vvv
+template <class _Tp, class _Up, size_t _Size, class = void>
+inline constexpr bool __swappable_arrays = false;
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^ || _CCCL_COMPILER(NVHPC)
+
+template <class _Tp, class _Up, class = void>
+inline constexpr bool __noexcept_swappable_arrays = false;
+
+struct __fn
+{
+  // 2.1   `S` is `(void)swap(E1, E2)`* if `E1` or `E2` has class or enumeration type and...
+  // *The name `swap` is used here unqualified.
+  _CCCL_TEMPLATE(class _Tp, class _Up)
+  _CCCL_REQUIRES(__unqualified_swappable_with<_Tp, _Up>)
+  _CCCL_API constexpr void operator()(_Tp&& __t, _Up&& __u) const
+    noexcept(noexcept(swap(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u))))
+  {
+    swap(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u));
+  }
+
+  // 2.2   Otherwise, if `E1` and `E2` are lvalues of array types with equal extent and...
+  _CCCL_TEMPLATE(class _Tp, class _Up, size_t _Size)
+  _CCCL_REQUIRES(__swappable_arrays<_Tp, _Up, _Size>)
+  _CCCL_API constexpr void operator()(_Tp (&__t)[_Size], _Up (&__u)[_Size]) const
+    noexcept(__noexcept_swappable_arrays<_Tp, _Up>)
+  {
+    // TODO(cjdb): replace with `::cuda::std::ranges::swap_ranges`.
+    for (size_t __i = 0; __i < _Size; ++__i)
+    {
+      (*this)(__t[__i], __u[__i]);
+    }
+  }
+
+  // 2.3   Otherwise, if `E1` and `E2` are lvalues of the same type `T` that models...
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__exchangeable<_Tp>)
+  _CCCL_API constexpr void operator()(_Tp& __x, _Tp& __y) const
+    noexcept(is_nothrow_move_constructible_v<_Tp> && is_nothrow_move_assignable_v<_Tp>)
+  {
+    __y = ::cuda::std::exchange(__x, ::cuda::std::move(__y));
+  }
+};
+
+#if !_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVHPC)
+template <class _Tp, class _Up, class _Size>
+_CCCL_CONCEPT_FRAGMENT(
+  __swappable_arrays_,
+  requires(_Tp (&__t)[_Size::value], _Up (&__u)[_Size::value], const __fn& __swap)(
+    requires(!__unqualified_swappable_with<_Tp (&)[_Size::value], _Up (&)[_Size::value]>),
+    requires(extent_v<_Tp> == extent_v<_Up>),
+    (__swap(__t[0], __u[0]))));
+
+template <class _Tp, class _Up, size_t _Size>
+inline constexpr bool __swappable_arrays<_Tp, _Up, _Size, void_t<type_identity_t<_Tp>>> =
+  _CCCL_FRAGMENT(__swappable_arrays_, _Tp, _Up, ::cuda::std::integral_constant<size_t, _Size>);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^ || _CCCL_COMPILER(NVHPC)
+
+template <class _Tp, class _Up>
+inline constexpr bool __noexcept_swappable_arrays<_Tp, _Up, void_t<type_identity_t<_Tp>>> =
+  noexcept(__swap::__fn{}(::cuda::std::declval<_Tp&>(), ::cuda::std::declval<_Up&>()));
+
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto swap = __swap::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __swap_cpo = __swap::__fn;
+} // namespace __cpo
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept swappable = requires(_Tp& __a, _Tp& __b) { ::cuda::std::ranges::__swap_cpo{}(__a, __b); };
+
+template <class _Tp, class _Up>
+concept swappable_with = common_reference_with<_Tp, _Up> && requires(_Tp&& __t, _Up&& __u) {
+  ::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Tp>(__t));
+  ::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Up>(__u), ::cuda::std::forward<_Up>(__u));
+  ::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u));
+  ::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Up>(__u), ::cuda::std::forward<_Tp>(__t));
+};
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__swappable_, requires(_Tp& __a, _Tp& __b)((::cuda::std::ranges::__swap_cpo{}(__a, __b))));
+
+template <class _Tp>
+_CCCL_CONCEPT swappable = _CCCL_FRAGMENT(__swappable_, _Tp);
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __swappable_with_,
+  requires(_Tp&& __t, _Up&& __u)(
+    requires(common_reference_with<_Tp, _Up>),
+    (::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Tp>(__t))),
+    (::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Up>(__u), ::cuda::std::forward<_Up>(__u))),
+    (::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Tp>(__t), ::cuda::std::forward<_Up>(__u))),
+    (::cuda::std::ranges::__swap_cpo{}(::cuda::std::forward<_Up>(__u), ::cuda::std::forward<_Tp>(__t)))));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT swappable_with = _CCCL_FRAGMENT(__swappable_with_, _Tp, _Up);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#if _CCCL_COMPILER(MSVC)
+_CCCL_END_NV_DIAG_SUPPRESS() // nonstandard cast to array type ignored
+#endif // _CCCL_COMPILER(MSVC)
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_SWAPPABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/totally_ordered.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/totally_ordered.h
new file mode 100644
index 00000000..f7534699
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__concepts/totally_ordered.h
@@ -0,0 +1,101 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CONCEPTS_TOTALLY_ORDERED_H
+#define _CUDA_STD___CONCEPTS_TOTALLY_ORDERED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/boolean_testable.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/equality_comparable.h>
+#include <cuda/std/__type_traits/common_reference.h>
+#include <cuda/std/__type_traits/make_const_lvalue_ref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [concept.totallyordered]
+
+template <class _Tp, class _Up>
+concept __partially_ordered_with = requires(__make_const_lvalue_ref<_Tp> __t, __make_const_lvalue_ref<_Up> __u) {
+  { __t < __u } -> __boolean_testable;
+  { __t > __u } -> __boolean_testable;
+  { __t <= __u } -> __boolean_testable;
+  { __t >= __u } -> __boolean_testable;
+  { __u < __t } -> __boolean_testable;
+  { __u > __t } -> __boolean_testable;
+  { __u <= __t } -> __boolean_testable;
+  { __u >= __t } -> __boolean_testable;
+};
+
+template <class _Tp>
+concept totally_ordered = equality_comparable<_Tp> && __partially_ordered_with<_Tp, _Tp>;
+
+template <class _Tp, class _Up>
+concept totally_ordered_with =
+  totally_ordered<_Tp> && totally_ordered<_Up> && equality_comparable_with<_Tp, _Up>
+  && totally_ordered<common_reference_t<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>>
+  && __partially_ordered_with<_Tp, _Up>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __partially_ordered_with_,
+  requires(__make_const_lvalue_ref<_Tp> __t, __make_const_lvalue_ref<_Up> __u)(
+    _Satisfies(__boolean_testable)(__t < __u), //
+    _Satisfies(__boolean_testable)(__t > __u), //
+    _Satisfies(__boolean_testable)(__t <= __u), //
+    _Satisfies(__boolean_testable)(__t >= __u), //
+    _Satisfies(__boolean_testable)(__u < __t), //
+    _Satisfies(__boolean_testable)(__u > __t), //
+    _Satisfies(__boolean_testable)(__u <= __t), //
+    _Satisfies(__boolean_testable)(__u >= __t)));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT __partially_ordered_with = _CCCL_FRAGMENT(__partially_ordered_with_, _Tp, _Up);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__totally_ordered_,
+                       requires()(requires(equality_comparable<_Tp>), requires(__partially_ordered_with<_Tp, _Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT totally_ordered = _CCCL_FRAGMENT(__totally_ordered_, _Tp);
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT_FRAGMENT(
+  __totally_ordered_with_,
+  requires()(requires(totally_ordered<_Tp>),
+             requires(totally_ordered<_Up>),
+             requires(equality_comparable_with<_Tp, _Up>),
+             requires(totally_ordered<common_reference_t<__make_const_lvalue_ref<_Tp>, __make_const_lvalue_ref<_Up>>>),
+             requires(__partially_ordered_with<_Tp, _Up>)));
+
+template <class _Tp, class _Up>
+_CCCL_CONCEPT totally_ordered_with = _CCCL_FRAGMENT(__totally_ordered_with_, _Tp, _Up);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CONCEPTS_TOTALLY_ORDERED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/byte.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/byte.h
new file mode 100644
index 00000000..f5e90b57
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/byte.h
@@ -0,0 +1,113 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CSTDDEF_BYTE_H
+#define _CUDA_STD___CSTDDEF_BYTE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__type_traits/is_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_NOVERSION
+
+enum class byte : unsigned char
+{
+};
+
+_CCCL_API constexpr byte operator|(byte __lhs, byte __rhs) noexcept
+{
+  return static_cast<byte>(
+    static_cast<unsigned char>(static_cast<unsigned int>(__lhs) | static_cast<unsigned int>(__rhs)));
+}
+
+_CCCL_API constexpr byte& operator|=(byte& __lhs, byte __rhs) noexcept
+{
+  return __lhs = __lhs | __rhs;
+}
+
+_CCCL_API constexpr byte operator&(byte __lhs, byte __rhs) noexcept
+{
+  return static_cast<byte>(
+    static_cast<unsigned char>(static_cast<unsigned int>(__lhs) & static_cast<unsigned int>(__rhs)));
+}
+
+_CCCL_API constexpr byte& operator&=(byte& __lhs, byte __rhs) noexcept
+{
+  return __lhs = __lhs & __rhs;
+}
+
+_CCCL_API constexpr byte operator^(byte __lhs, byte __rhs) noexcept
+{
+  return static_cast<byte>(
+    static_cast<unsigned char>(static_cast<unsigned int>(__lhs) ^ static_cast<unsigned int>(__rhs)));
+}
+
+_CCCL_API constexpr byte& operator^=(byte& __lhs, byte __rhs) noexcept
+{
+  return __lhs = __lhs ^ __rhs;
+}
+
+_CCCL_API constexpr byte operator~(byte __b) noexcept
+{
+  return static_cast<byte>(static_cast<unsigned char>(~static_cast<unsigned int>(__b)));
+}
+
+_CCCL_TEMPLATE(class _Integer)
+_CCCL_REQUIRES(is_integral_v<_Integer>)
+_CCCL_API constexpr byte& operator<<=(byte& __lhs, _Integer __shift) noexcept
+{
+  return __lhs = __lhs << __shift;
+}
+
+_CCCL_TEMPLATE(class _Integer)
+_CCCL_REQUIRES(is_integral_v<_Integer>)
+_CCCL_API constexpr byte operator<<(byte __lhs, _Integer __shift) noexcept
+{
+  return static_cast<byte>(static_cast<unsigned char>(static_cast<unsigned int>(__lhs) << __shift));
+}
+
+_CCCL_TEMPLATE(class _Integer)
+_CCCL_REQUIRES(is_integral_v<_Integer>)
+_CCCL_API constexpr byte& operator>>=(byte& __lhs, _Integer __shift) noexcept
+{
+  return __lhs = __lhs >> __shift;
+}
+
+_CCCL_TEMPLATE(class _Integer)
+_CCCL_REQUIRES(is_integral_v<_Integer>)
+_CCCL_API constexpr byte operator>>(byte __lhs, _Integer __shift) noexcept
+{
+  return static_cast<byte>(static_cast<unsigned char>(static_cast<unsigned int>(__lhs) >> __shift));
+}
+
+_CCCL_TEMPLATE(class _Integer)
+_CCCL_REQUIRES(is_integral_v<_Integer>)
+_CCCL_API constexpr _Integer to_integer(byte __b) noexcept
+{
+  return static_cast<_Integer>(__b);
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD_NOVERSION
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CSTDDEF_BYTE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/types.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/types.h
new file mode 100644
index 00000000..57208e29
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstddef/types.h
@@ -0,0 +1,52 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CSTDDEF_TYPES_H
+#define _CUDA_STD___CSTDDEF_TYPES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <cstddef>
+#else // ^^^ _CCCL_HOSTED() ^^^ / vvv _CCCL_FREESTANDING() vvv
+#  if !defined(offsetof)
+#    define offsetof(type, member) (::size_t) ((char*) &(((type*) 0)->member) - (char*) 0)
+#  endif // !offsetof
+#endif // _CCCL_FREESTANDING()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_FREESTANDING()
+using max_align_t = long double;
+#else // ^^^ _CCCL_FREESTANDING() ^^^ / vvv _CCCL_HOSTED() vvv
+// Re-use the compiler's <stddef.h> max_align_t where possible.
+using ::max_align_t;
+#endif // _CCCL_HOSTED()
+
+using nullptr_t = decltype(nullptr);
+using ::ptrdiff_t;
+using ::size_t;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CSTDDEF_TYPES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstring/memcpy.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstring/memcpy.h
new file mode 100644
index 00000000..00bcc9e0
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__cstring/memcpy.h
@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___CSTRING_MEMCPY
+#define _CUDA_STD___CSTRING_MEMCPY
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__memory/check_address.h>
+
+#if _CCCL_HOSTED()
+#  include <cstring>
+#endif // _CCCL_HOSTED()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+using ::size_t;
+
+// old compilers still trigger the name conflict
+// nvcc 12.0 and 12.1 trigger segmentation fault
+#if _CCCL_COMPILER(GCC, <=, 9) || _CCCL_CUDA_COMPILER(NVCC, <=, 12, 1)
+
+using ::memcpy;
+
+#else // ^^^ _CCCL_COMPILER(GCC, <=, 9) ^^^ / vvv _CCCL_COMPILER(GCC, >, 9) vvv
+
+// The template parameter is used to avoid name ambiguity when external code calls 'memcpy' without namespace
+// qualification. Function templates have lower precedence than non-template functions for overload resolution.
+template <int = 0>
+_CCCL_API inline void* memcpy(void* __dest, const void* __src, size_t __count) noexcept
+{
+  _CCCL_ASSERT(::cuda::__is_valid_address_range(__src, __count), "memcpy: source range is invalid");
+  _CCCL_ASSERT(::cuda::__is_valid_address_range(__dest, __count), "memcpy: destination range is invalid");
+  _CCCL_ASSERT(!::cuda::__are_ptrs_overlapping(__src, __dest, __count), "memcpy: source and destination overlap");
+  return ::memcpy(__dest, __src, __count);
+}
+
+#endif // ^^^ _CCCL_COMPILER(GCC, <=, 9) ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___CSTRING_MEMCPY
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/exception_macros.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/exception_macros.h
new file mode 100644
index 00000000..02c8edf4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/exception_macros.h
@@ -0,0 +1,126 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___EXCEPTION_EXCEPTION_MACROS_H
+#define _CUDA_STD___EXCEPTION_EXCEPTION_MACROS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__exception/terminate.h>
+#include <cuda/std/__host_stdlib/cstdio>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct __cccl_catch_any_lvalue
+{
+  template <class _Tp>
+  _CCCL_API operator _Tp&() const noexcept;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// The following macros are used to conditionally compile exception handling code. They
+// are used in the same way as `try` and `catch`, but they allow for different behavior
+// based on whether exceptions are enabled or not, and whether the code is being compiled
+// for device or not.
+//
+// Usage:
+//   _CCCL_TRY
+//   {
+//     can_throw();               // Code that may throw an exception
+//   }
+//   _CCCL_CATCH (cuda_error& e)  // Handle CUDA exceptions
+//   {
+//     printf("CUDA error: %s\n", e.what());
+//   }
+//   _CCCL_CATCH_ALL              // Handle any other exceptions
+//   {
+//     printf("unknown error\n");
+//   }
+//
+// Notes:
+//   - the catch clause must always bind to a named variable
+
+// Expand to keywords only for host code when exceptions are enabled. nvc++ in CUDA mode traps when an exception is
+// thrown in device code.
+#if _CCCL_HAS_EXCEPTIONS() && _CCCL_HOST_COMPILATION()
+#  define _CCCL_TRY       try
+#  define _CCCL_CATCH     catch
+#  define _CCCL_CATCH_ALL catch (...)
+#  define _CCCL_CATCH_FALLTHROUGH
+
+// Even though nvc++ in CUDA mode replaces `throw` by `__trap()` call in device code, it instantiates the exception type
+// which can introduce some host only symbols to the nvvm ir (for example snprintf). So we need to wrap it by the
+// NV_IF_ELSE_TARGET macro.
+#  define _CCCL_THROW(_TYPE, ...)                                                             \
+    do                                                                                        \
+    {                                                                                         \
+      NV_IF_ELSE_TARGET(NV_IS_HOST, (throw _TYPE(__VA_ARGS__);), (::cuda::std::terminate();)) \
+    } while (0)
+#  define _CCCL_RETHROW throw
+#else // ^^^ use exceptions ^^^ / vvv no exceptions vvv
+#  define _CCCL_TRY     \
+    if constexpr (true) \
+    {
+#  define _CCCL_CATCH(...)    \
+    }                         \
+    else if constexpr (false) \
+    {                         \
+      for (__VA_ARGS__ = ::cuda::std::__cccl_catch_any_lvalue{}; false;)
+#  define _CCCL_CATCH_ALL \
+    }                     \
+    else
+#  define _CCCL_CATCH_FALLTHROUGH \
+    }                             \
+    else                          \
+    {                             \
+    }
+
+#  if _CCCL_HOSTJIT()
+#    define _CCCL_THROW(_TYPE, ...)                                              \
+      do                                                                         \
+      {                                                                          \
+        _CCCL_ASSERT(false, "An instance of class " #_TYPE " would be thrown."); \
+        ::cuda::std::terminate();                                                \
+      } while (0)
+#  else // ^^^ _CCCL_HOSTJIT() ^^^ / vvv !_CCCL_HOSTJIT() vvv
+#    define _CCCL_THROW(_TYPE, ...)                                                                                \
+      do                                                                                                           \
+      {                                                                                                            \
+        NV_IF_ELSE_TARGET(NV_IS_HOST,                                                                              \
+                          ({                                                                                       \
+                            ::fprintf(stderr,                                                                      \
+                                      "%s:%u: An instance of class %s would be thrown.\n  what():  %s\nAborted\n", \
+                                      __FILE__,                                                                    \
+                                      __LINE__,                                                                    \
+                                      #_TYPE,                                                                      \
+                                      (_TYPE(__VA_ARGS__)).what());                                                \
+                            ::fflush(stderr);                                                                      \
+                          }),                                                                                      \
+                          ({ _CCCL_ASSERT(false, "An instance of class " #_TYPE " would be thrown."); }))          \
+        ::cuda::std::terminate();                                                                                  \
+      } while (0)
+#  endif // !_CCCL_HOSTJIT()
+#  define _CCCL_RETHROW ::cuda::std::terminate()
+#endif // ^^^ no exceptions ^^^
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___EXCEPTION_EXCEPTION_MACROS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/terminate.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/terminate.h
new file mode 100644
index 00000000..6c31b483
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__exception/terminate.h
@@ -0,0 +1,82 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___EXCEPTION_TERMINATE_H
+#define _CUDA_STD___EXCEPTION_TERMINATE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_TILE_COMPILATION()
+#  include <cuda/std/cassert>
+#endif // !_CCCL_TILE_COMPILATION()
+
+#if _CCCL_HOSTED()
+#  include <stdlib.h>
+#endif // _CCCL_HOSTED()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4702) // unreachable code
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_NOVERSION // purposefully not using versioning namespace
+
+[[noreturn]] _CCCL_API inline void __cccl_terminate() noexcept
+{
+#if _CCCL_TILE_COMPILATION()
+  NV_IF_ELSE_TARGET(NV_IS_HOST, (::exit(-1);), (assert(false);))
+#else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION()
+  NV_IF_ELSE_TARGET(NV_IS_HOST, (::exit(-1);), (::__trap();))
+#endif // !_CCCL_TILE_COMPILATION()
+  _CCCL_UNREACHABLE();
+}
+
+#if 0 // Expose once atomic is universally available
+
+using terminate_handler = void (*)();
+
+#  ifdef __CUDA_ARCH__
+__device__
+#  endif // __CUDA_ARCH__
+  static _CCCL_CONSTINIT ::cuda::std::atomic<terminate_handler>
+    __cccl_terminate_handler{&__cccl_terminate};
+
+_CCCL_API inline  terminate_handler set_terminate(terminate_handler __func) noexcept
+{
+  return __cccl_terminate_handler.exchange(__func);
+}
+_CCCL_API inline  terminate_handler get_terminate() noexcept
+{
+  return __cccl_terminate_handler.load(__func);
+}
+
+#endif
+
+[[noreturn]] _CCCL_API inline void terminate() noexcept
+{
+  __cccl_terminate();
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD_NOVERSION
+
+_CCCL_DIAG_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___EXCEPTION_TERMINATE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/format.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/format.h
new file mode 100644
index 00000000..c90924ac
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/format.h
@@ -0,0 +1,157 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FLOATING_POINT_FORMAT_H
+#define _CUDA_STD___FLOATING_POINT_FORMAT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/fp.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/cfloat>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+enum class __fp_format
+{
+  __binary16, // IEEE 754 binary16
+  __binary32, // IEEE 754 binary32
+  __binary64, // IEEE 754 binary64
+  __binary128, // IEEE 754 binary128
+  __bfloat16, // Google's 16-bit brain float
+  __fp80_x86, // x86 80-bit extended precision
+  __fp8_nv_e4m3, // NVIDIA's __nv_fp8_e4m3
+  __fp8_nv_e5m2, // NVIDIA's __nv_fp8_e5m2
+  __fp8_nv_e8m0, // NVIDIA's __nv_fp8_e8m0
+  __fp6_nv_e2m3, // NVIDIA's __nv_fp6_e2m3
+  __fp6_nv_e3m2, // NVIDIA's __nv_fp6_e3m2
+  __fp4_nv_e2m1, // NVIDIA's __nv_fp4_e2m1
+
+  __invalid,
+};
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr __fp_format __fp_format_of_v_impl() noexcept
+{
+  if constexpr (is_same_v<_Tp, float>)
+  {
+    return __fp_format::__binary32;
+  }
+  else if constexpr (is_same_v<_Tp, double>)
+  {
+    return __fp_format::__binary64;
+  }
+#if _CCCL_HAS_LONG_DOUBLE()
+  else if constexpr (is_same_v<_Tp, long double>)
+  {
+#  if LDBL_MIN_EXP == -1021 && LDBL_MAX_EXP == 1024 && LDBL_MANT_DIG == 53
+    return __fp_format::__binary64;
+#  elif LDBL_MIN_EXP == -16381 && LDBL_MAX_EXP == 16384 && LDBL_MANT_DIG == 64
+    static_assert(sizeof(long double) == 16,
+                  "When the long double format is x86 80-bit extended floating point, CCCL requires the size of long "
+                  "double to be 16 bytes.");
+    return __fp_format::__fp80_x86;
+#  elif LDBL_MIN_EXP == -16381 && LDBL_MAX_EXP == 16384 && LDBL_MANT_DIG == 113
+    return __fp_format::__binary128;
+#  else
+#    error "Unknown long double format. Define CCCL_DISABLE_LONG_DOUBLE to disable long double support in CCCL."
+#  endif
+  }
+#endif // _CCCL_HAS_LONG_DOUBLE()
+#if _CCCL_HAS_NVFP16()
+  else if constexpr (is_same_v<_Tp, __half>)
+  {
+    return __fp_format::__binary16;
+  }
+#endif // _CCCL_HAS_NVFP16()
+#if _CCCL_HAS_NVBF16()
+  else if constexpr (is_same_v<_Tp, __nv_bfloat16>)
+  {
+    return __fp_format::__bfloat16;
+  }
+#endif // _CCCL_HAS_NVBF16()
+#if _CCCL_HAS_NVFP8_E4M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e4m3>)
+  {
+    return __fp_format::__fp8_nv_e4m3;
+  }
+#endif // _CCCL_HAS_NVFP8_E4M3()
+#if _CCCL_HAS_NVFP8_E5M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e5m2>)
+  {
+    return __fp_format::__fp8_nv_e5m2;
+  }
+#endif // _CCCL_HAS_NVFP8_E5M2()
+#if _CCCL_HAS_NVFP8_E8M0()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e8m0>)
+  {
+    return __fp_format::__fp8_nv_e8m0;
+  }
+#endif // _CCCL_HAS_NVFP8_E8M0()
+#if _CCCL_HAS_NVFP6_E2M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e2m3>)
+  {
+    return __fp_format::__fp6_nv_e2m3;
+  }
+#endif // _CCCL_HAS_NVFP6_E2M3()
+#if _CCCL_HAS_NVFP6_E3M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e3m2>)
+  {
+    return __fp_format::__fp6_nv_e3m2;
+  }
+#endif // _CCCL_HAS_NVFP6_E3M2()
+#if _CCCL_HAS_NVFP4_E2M1()
+  else if constexpr (is_same_v<_Tp, __nv_fp4_e2m1>)
+  {
+    return __fp_format::__fp4_nv_e2m1;
+  }
+#endif // _CCCL_HAS_NVFP4_E2M1()
+#if _CCCL_HAS_FLOAT128()
+  else if constexpr (is_same_v<_Tp, __float128>)
+  {
+    return __fp_format::__binary128;
+  }
+#endif // _CCCL_HAS_FLOAT128()
+  else
+  {
+    return __fp_format::__invalid;
+  }
+}
+
+template <class _Tp>
+inline constexpr __fp_format __fp_format_of_v = ::cuda::std::__fp_format_of_v_impl<_Tp>();
+
+template <class _Tp>
+inline constexpr __fp_format __fp_format_of_v<const _Tp> = __fp_format_of_v<_Tp>;
+
+template <class _Tp>
+inline constexpr __fp_format __fp_format_of_v<volatile _Tp> = __fp_format_of_v<_Tp>;
+
+template <class _Tp>
+inline constexpr __fp_format __fp_format_of_v<const volatile _Tp> = __fp_format_of_v<_Tp>;
+
+template <__fp_format _Fmt>
+inline constexpr __fp_format __fp_format_of_v<__cccl_fp<_Fmt>> = _Fmt;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FLOATING_POINT_FORMAT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/properties.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/properties.h
new file mode 100644
index 00000000..2e11a25d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/properties.h
@@ -0,0 +1,229 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FLOATING_POINT_PROPERTIES_H
+#define _CUDA_STD___FLOATING_POINT_PROPERTIES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__floating_point/format.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// __fp_is_signed_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_is_signed_v = true;
+
+template <>
+inline constexpr bool __fp_is_signed_v<__fp_format::__fp8_nv_e8m0> = false;
+
+// __fp_exp_nbits_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_exp_nbits_v = 0;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__binary16> = 5;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__binary32> = 8;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__binary64> = 11;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__binary128> = 15;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__bfloat16> = 8;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp80_x86> = 15;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp8_nv_e4m3> = 4;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp8_nv_e5m2> = 5;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp8_nv_e8m0> = 8;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp6_nv_e2m3> = 2;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp6_nv_e3m2> = 3;
+
+template <>
+inline constexpr int __fp_exp_nbits_v<__fp_format::__fp4_nv_e2m1> = 2;
+
+// __fp_exp_bias_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_exp_bias_v = (1 << (__fp_exp_nbits_v<_Fmt> - 1)) - 1;
+
+// __fp_exp_min_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_exp_min_v = 1 - __fp_exp_bias_v<_Fmt>;
+
+template <>
+inline constexpr int __fp_exp_min_v<__fp_format::__fp8_nv_e8m0> = -127;
+
+// __fp_exp_max_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_exp_max_v = (1 << __fp_exp_nbits_v<_Fmt>) -2 - __fp_exp_bias_v<_Fmt>;
+
+template <>
+inline constexpr int __fp_exp_max_v<__fp_format::__fp8_nv_e4m3> = 8;
+
+template <>
+inline constexpr int __fp_exp_max_v<__fp_format::__fp6_nv_e2m3> = 2;
+
+template <>
+inline constexpr int __fp_exp_max_v<__fp_format::__fp6_nv_e3m2> = 4;
+
+template <>
+inline constexpr int __fp_exp_max_v<__fp_format::__fp4_nv_e2m1> = 2;
+
+// __fp_mant_nbits_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_mant_nbits_v = 0;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__binary16> = 10;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__binary32> = 23;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__binary64> = 52;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__binary128> = 112;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__bfloat16> = 7;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp80_x86> = 64;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp8_nv_e4m3> = 3;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp8_nv_e5m2> = 2;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp8_nv_e8m0> = 0;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp6_nv_e2m3> = 3;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp6_nv_e3m2> = 2;
+
+template <>
+inline constexpr int __fp_mant_nbits_v<__fp_format::__fp4_nv_e2m1> = 1;
+
+// __fp_has_implicit_bit_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_has_implicit_bit_v = true;
+
+template <>
+inline constexpr bool __fp_has_implicit_bit_v<__fp_format::__fp80_x86> = false;
+
+// __fp_digits_v
+
+template <__fp_format _Fmt>
+inline constexpr int __fp_digits_v = __fp_mant_nbits_v<_Fmt> + static_cast<int>(__fp_has_implicit_bit_v<_Fmt>);
+
+// __fp_has_denorm_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_has_denorm_v = true;
+
+template <>
+inline constexpr bool __fp_has_denorm_v<__fp_format::__fp8_nv_e8m0> = false;
+
+// __fp_has_inf_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_has_inf_v = true;
+
+template <>
+inline constexpr bool __fp_has_inf_v<__fp_format::__fp8_nv_e4m3> = false;
+
+template <>
+inline constexpr bool __fp_has_inf_v<__fp_format::__fp8_nv_e8m0> = false;
+
+template <>
+inline constexpr bool __fp_has_inf_v<__fp_format::__fp6_nv_e2m3> = false;
+
+template <>
+inline constexpr bool __fp_has_inf_v<__fp_format::__fp6_nv_e3m2> = false;
+
+template <>
+inline constexpr bool __fp_has_inf_v<__fp_format::__fp4_nv_e2m1> = false;
+
+// __fp_has_nan_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_has_nan_v = true;
+
+template <>
+inline constexpr bool __fp_has_nan_v<__fp_format::__fp6_nv_e2m3> = false;
+
+template <>
+inline constexpr bool __fp_has_nan_v<__fp_format::__fp6_nv_e3m2> = false;
+
+template <>
+inline constexpr bool __fp_has_nan_v<__fp_format::__fp4_nv_e2m1> = false;
+
+// __fp_has_nans_v
+
+template <__fp_format _Fmt>
+inline constexpr bool __fp_has_nans_v = true;
+
+template <>
+inline constexpr bool __fp_has_nans_v<__fp_format::__fp8_nv_e4m3> = false;
+
+template <>
+inline constexpr bool __fp_has_nans_v<__fp_format::__fp8_nv_e8m0> = false;
+
+template <>
+inline constexpr bool __fp_has_nans_v<__fp_format::__fp6_nv_e2m3> = false;
+
+template <>
+inline constexpr bool __fp_has_nans_v<__fp_format::__fp6_nv_e3m2> = false;
+
+template <>
+inline constexpr bool __fp_has_nans_v<__fp_format::__fp4_nv_e2m1> = false;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FLOATING_POINT_PROPERTIES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/storage.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/storage.h
new file mode 100644
index 00000000..37af5ec4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/storage.h
@@ -0,0 +1,260 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FLOATING_POINT_STORAGE_H
+#define _CUDA_STD___FLOATING_POINT_STORAGE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__bit/bit_cast.h>
+#include <cuda/std/__floating_point/format.h>
+#include <cuda/std/__floating_point/traits.h>
+#include <cuda/std/__type_traits/always_false.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/cstdint>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <__fp_format _Fmt>
+[[nodiscard]] _CCCL_API constexpr auto __fp_storage_type_impl() noexcept
+{
+  if constexpr (_Fmt == __fp_format::__fp8_nv_e4m3 || _Fmt == __fp_format::__fp8_nv_e5m2
+                || _Fmt == __fp_format::__fp8_nv_e8m0 || _Fmt == __fp_format::__fp6_nv_e2m3
+                || _Fmt == __fp_format::__fp6_nv_e3m2 || _Fmt == __fp_format::__fp4_nv_e2m1)
+  {
+    return uint8_t{};
+  }
+  else if constexpr (_Fmt == __fp_format::__binary16 || _Fmt == __fp_format::__bfloat16)
+  {
+    return uint16_t{};
+  }
+  else if constexpr (_Fmt == __fp_format::__binary32)
+  {
+    return uint32_t{};
+  }
+  else if constexpr (_Fmt == __fp_format::__binary64)
+  {
+    return uint64_t{};
+  }
+#if _CCCL_HAS_INT128()
+  else if constexpr (_Fmt == __fp_format::__fp80_x86 || _Fmt == __fp_format::__binary128)
+  {
+    return __uint128_t{};
+  }
+#endif // _CCCL_HAS_INT128()
+  else
+  {
+    static_assert(__always_false_v<decltype(_Fmt)>, "Unsupported floating point format");
+  }
+}
+
+template <__fp_format _Fmt>
+using __fp_storage_t = decltype(__fp_storage_type_impl<_Fmt>());
+
+template <class _Tp>
+using __fp_storage_of_t = __fp_storage_t<__fp_format_of_v<_Tp>>;
+
+#if !_CCCL_TILE_COMPILATION()
+template <class _Tp>
+struct __cccl_nvfp_manip_helper : _Tp
+{
+  using _Tp::__x;
+};
+#endif // _CCCL_TILE_COMPILATION()
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp __fp_from_storage(__fp_storage_of_t<_Tp> __v) noexcept
+{
+  if constexpr (__is_std_fp_v<_Tp> || __is_ext_compiler_fp_v<_Tp>)
+  {
+    return ::cuda::std::bit_cast<_Tp>(__v);
+  }
+  else if constexpr (__is_ext_cccl_fp_v<_Tp>)
+  {
+    _Tp __ret{};
+    __ret.__storage_ = __v;
+    return __ret;
+  }
+#if _CCCL_HAS_NVFP16()
+  else if constexpr (is_same_v<_Tp, __half>)
+  {
+#  if _CCCL_TILE_COMPILATION()
+    return ::cuda::std::bit_cast<_Tp>(__v);
+#  else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION()
+    __cccl_nvfp_manip_helper<_Tp> __helper{};
+    __helper.__x = __v;
+    return __helper;
+#  endif // !_CCCL_TILE_COMPILATION()
+  }
+#endif // _CCCL_HAS_NVFP16()
+#if _CCCL_HAS_NVBF16()
+  else if constexpr (is_same_v<_Tp, __nv_bfloat16>)
+  {
+#  if _CCCL_TILE_COMPILATION()
+    return ::cuda::std::bit_cast<_Tp>(__v);
+#  else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION()
+    __cccl_nvfp_manip_helper<_Tp> __helper{};
+    __helper.__x = __v;
+    return __helper;
+#  endif // !_CCCL_TILE_COMPILATION()
+  }
+#endif // _CCCL_HAS_NVBF16()
+#if _CCCL_HAS_NVFP8_E4M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e4m3>)
+  {
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP8_E4M3()
+#if _CCCL_HAS_NVFP8_E5M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e5m2>)
+  {
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP8_E5M2()
+#if _CCCL_HAS_NVFP8_E8M0()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e8m0>)
+  {
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP8_E8M0()
+#if _CCCL_HAS_NVFP6_E2M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e2m3>)
+  {
+    _CCCL_ASSERT((__v & 0xc0u) == 0u, "Invalid __nv_fp6_e2m3 storage value");
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP6_E2M3()
+#if _CCCL_HAS_NVFP6_E3M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e3m2>)
+  {
+    _CCCL_ASSERT((__v & 0xc0u) == 0u, "Invalid __nv_fp6_e3m2 storage value");
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP6_E3M2()
+#if _CCCL_HAS_NVFP4_E2M1()
+  else if constexpr (is_same_v<_Tp, __nv_fp4_e2m1>)
+  {
+    _CCCL_ASSERT((__v & 0xf0u) == 0u, "Invalid __nv_fp4_e2m1 storage value");
+    _Tp __ret{};
+    __ret.__x = __v;
+    return __ret;
+  }
+#endif // _CCCL_HAS_NVFP4_E2M1()
+  else
+  {
+    static_assert(__always_false_v<_Tp>, "Unsupported floating point format");
+  }
+}
+
+_CCCL_TEMPLATE(class _Tp, class _Up)
+_CCCL_REQUIRES((!is_same_v<_Up, __fp_storage_of_t<_Tp>>) )
+_CCCL_API constexpr _Tp __fp_from_storage(const _Up& __v) noexcept = delete;
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr __fp_storage_of_t<_Tp> __fp_get_storage(_Tp __v) noexcept
+{
+  if constexpr (__is_std_fp_v<_Tp> || __is_ext_compiler_fp_v<_Tp>)
+  {
+    return ::cuda::std::bit_cast<__fp_storage_of_t<_Tp>>(__v);
+  }
+  else if constexpr (__is_ext_cccl_fp_v<_Tp>)
+  {
+    return __v.__storage_;
+  }
+#if _CCCL_HAS_NVFP16()
+  else if constexpr (is_same_v<_Tp, __half>)
+  {
+#  if _CCCL_TILE_COMPILATION()
+    return ::cuda::std::bit_cast<__fp_storage_of_t<_Tp>>(__v);
+#  else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION() vvv
+    return __cccl_nvfp_manip_helper<_Tp>{__v}.__x;
+#  endif // !_CCCL_TILE_COMPILATION()
+  }
+#endif // _CCCL_HAS_NVFP16()
+#if _CCCL_HAS_NVBF16()
+  else if constexpr (is_same_v<_Tp, __nv_bfloat16>)
+  {
+#  if _CCCL_TILE_COMPILATION()
+    return ::cuda::std::bit_cast<__fp_storage_of_t<_Tp>>(__v);
+#  else // ^^^ _CCCL_TILE_COMPILATION() ^^^ / vvv !_CCCL_TILE_COMPILATION() vvv
+    return __cccl_nvfp_manip_helper<_Tp>{__v}.__x;
+#  endif // !_CCCL_TILE_COMPILATION()
+  }
+#endif // _CCCL_HAS_NVBF16()
+  // Distinct extended floating-point types expose the same storage member.
+  // NOLINTBEGIN(bugprone-branch-clone)
+#if _CCCL_HAS_NVFP8_E4M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e4m3>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP8_E4M3()
+#if _CCCL_HAS_NVFP8_E5M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e5m2>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP8_E5M2()
+#if _CCCL_HAS_NVFP8_E8M0()
+  else if constexpr (is_same_v<_Tp, __nv_fp8_e8m0>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP8_E8M0()
+#if _CCCL_HAS_NVFP6_E2M3()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e2m3>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP6_E2M3()
+#if _CCCL_HAS_NVFP6_E3M2()
+  else if constexpr (is_same_v<_Tp, __nv_fp6_e3m2>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP6_E3M2()
+#if _CCCL_HAS_NVFP4_E2M1()
+  else if constexpr (is_same_v<_Tp, __nv_fp4_e2m1>)
+  {
+    return __v.__x;
+  }
+#endif // _CCCL_HAS_NVFP4_E2M1()
+  // NOLINTEND(bugprone-branch-clone)
+  else
+  {
+    static_assert(__always_false_v<_Tp>, "Unsupported floating point format");
+  }
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FLOATING_POINT_STORAGE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/traits.h
new file mode 100644
index 00000000..85ccd673
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__floating_point/traits.h
@@ -0,0 +1,171 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FLOATING_POINT_TRAITS_H
+#define _CUDA_STD___FLOATING_POINT_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__floating_point/properties.h>
+#include <cuda/std/__fwd/fp.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// __is_std_fp_v
+
+template <class _Tp>
+inline constexpr bool __is_std_fp_v = false;
+
+template <class _Tp>
+inline constexpr bool __is_std_fp_v<const _Tp> = __is_std_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_std_fp_v<volatile _Tp> = __is_std_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_std_fp_v<const volatile _Tp> = __is_std_fp_v<_Tp>;
+
+template <>
+inline constexpr bool __is_std_fp_v<float> = true;
+
+template <>
+inline constexpr bool __is_std_fp_v<double> = true;
+
+template <>
+inline constexpr bool __is_std_fp_v<long double> = true;
+
+// __is_ext_nv_fp_v
+
+template <class _Tp>
+inline constexpr bool __is_ext_nv_fp_v = false;
+
+template <class _Tp>
+inline constexpr bool __is_ext_nv_fp_v<const _Tp> = __is_ext_nv_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_nv_fp_v<volatile _Tp> = __is_ext_nv_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_nv_fp_v<const volatile _Tp> = __is_ext_nv_fp_v<_Tp>;
+
+#if _CCCL_HAS_NVFP16()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__half> = true;
+#endif // _CCCL_HAS_NVFP16()
+
+#if _CCCL_HAS_NVBF16()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_bfloat16> = true;
+#endif // _CCCL_HAS_NVBF16()
+
+#if _CCCL_HAS_NVFP8_E4M3()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp8_e4m3> = true;
+#endif // _CCCL_HAS_NVFP8_E4M3()
+
+#if _CCCL_HAS_NVFP8_E5M2()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp8_e5m2> = true;
+#endif // _CCCL_HAS_NVFP8_E5M2()
+
+#if _CCCL_HAS_NVFP8_E8M0()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp8_e8m0> = true;
+#endif // _CCCL_HAS_NVFP8_E8M0()
+
+#if _CCCL_HAS_NVFP6_E2M3()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp6_e2m3> = true;
+#endif // _CCCL_HAS_NVFP6_E2M3()
+
+#if _CCCL_HAS_NVFP6_E3M2()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp6_e3m2> = true;
+#endif // _CCCL_HAS_NVFP6_E3M2()
+
+#if _CCCL_HAS_NVFP4_E2M1()
+template <>
+inline constexpr bool __is_ext_nv_fp_v<__nv_fp4_e2m1> = true;
+#endif // _CCCL_HAS_NVFP4_E2M1()
+
+// __is_ext_compiler_fp_v
+
+template <class _Tp>
+inline constexpr bool __is_ext_compiler_fp_v = false;
+
+template <class _Tp>
+inline constexpr bool __is_ext_compiler_fp_v<const _Tp> = __is_ext_compiler_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_compiler_fp_v<volatile _Tp> = __is_ext_compiler_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_compiler_fp_v<const volatile _Tp> = __is_ext_compiler_fp_v<_Tp>;
+
+#if _CCCL_HAS_FLOAT128()
+template <>
+inline constexpr bool __is_ext_compiler_fp_v<__float128> = true;
+#endif // _CCCL_HAS_FLOAT128()
+
+// __is_ext_cccl_fp_v
+
+template <class _Tp>
+inline constexpr bool __is_ext_cccl_fp_v = false;
+
+template <class _Tp>
+inline constexpr bool __is_ext_cccl_fp_v<const _Tp> = __is_ext_cccl_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_cccl_fp_v<volatile _Tp> = __is_ext_cccl_fp_v<_Tp>;
+
+template <class _Tp>
+inline constexpr bool __is_ext_cccl_fp_v<const volatile _Tp> = __is_ext_cccl_fp_v<_Tp>;
+
+template <__fp_format _Fmt>
+inline constexpr bool __is_ext_cccl_fp_v<__cccl_fp<_Fmt>> = true;
+
+// __is_ext_fp_v
+
+template <class _Tp>
+inline constexpr bool __is_ext_fp_v = __is_ext_nv_fp_v<_Tp> || __is_ext_compiler_fp_v<_Tp> || __is_ext_cccl_fp_v<_Tp>;
+
+// __is_fp_v (todo: use cuda::std::is_floating_point_v instead in the future)
+
+template <class _Tp>
+inline constexpr bool __is_fp_v = __is_std_fp_v<_Tp> || __is_ext_fp_v<_Tp>;
+
+// __fp_is_subset_v
+
+template <__fp_format _LhsFmt, __fp_format _RhsFmt>
+inline constexpr bool __fp_is_subset_v =
+  (!__fp_is_signed_v<_LhsFmt> || __fp_is_signed_v<_RhsFmt>)
+  && __fp_exp_min_v<_LhsFmt> >= __fp_exp_min_v<_RhsFmt> && __fp_exp_max_v<_LhsFmt> <= __fp_exp_max_v<_RhsFmt>
+  && __fp_digits_v<_LhsFmt> <= __fp_digits_v<_RhsFmt> && (!__fp_has_denorm_v<_LhsFmt> || __fp_has_denorm_v<_RhsFmt>);
+
+// __fp_is_subset_of_v
+
+template <class _Lhs, class _Rhs>
+inline constexpr bool __fp_is_subset_of_v = __fp_is_subset_v<__fp_format_of_v<_Lhs>, __fp_format_of_v<_Rhs>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FLOATING_POINT_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/binary_function.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/binary_function.h
new file mode 100644
index 00000000..b81dfd42
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/binary_function.h
@@ -0,0 +1,64 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_BINARY_FUNCTION_H
+#define _CUDA_STD___FUNCTIONAL_BINARY_FUNCTION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION)
+
+template <class _Arg1, class _Arg2, class _Result>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT CCCL_DEPRECATED binary_function
+{
+  using first_argument_type  = _Arg1;
+  using second_argument_type = _Arg2;
+  using result_type          = _Result;
+};
+
+#endif // defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION)
+
+template <class _Arg1, class _Arg2, class _Result>
+struct __binary_function_keep_layout_base
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using first_argument_type CCCL_DEPRECATED  = _Arg1;
+  using second_argument_type CCCL_DEPRECATED = _Arg2;
+  using result_type CCCL_DEPRECATED          = _Result;
+#endif // _LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS
+};
+
+#if defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION)
+_CCCL_SUPPRESS_DEPRECATED_PUSH
+_CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+template <class _Arg1, class _Arg2, class _Result>
+using __binary_function = binary_function<_Arg1, _Arg2, _Result>;
+_CCCL_SUPPRESS_DEPRECATED_POP
+#else
+template <class _Arg1, class _Arg2, class _Result>
+using __binary_function = __binary_function_keep_layout_base<_Arg1, _Arg2, _Result>;
+#endif // !_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_BINARY_FUNCTION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/identity.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/identity.h
new file mode 100644
index 00000000..bc8e801f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/identity.h
@@ -0,0 +1,57 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_IDENTITY_H
+#define _CUDA_STD___FUNCTIONAL_IDENTITY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__functional/reference_wrapper.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __is_identity_v = false;
+
+struct identity
+{
+  template <class _Tp>
+  [[nodiscard]] _CCCL_API constexpr _Tp&& operator()(_Tp&& __t) const noexcept
+  {
+    return ::cuda::std::forward<_Tp>(__t);
+  }
+
+  using is_transparent = void;
+};
+
+template <>
+inline constexpr bool __is_identity_v<identity> = true;
+template <>
+inline constexpr bool __is_identity_v<reference_wrapper<identity>> = true;
+template <>
+inline constexpr bool __is_identity_v<reference_wrapper<const identity>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_IDENTITY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/invoke.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/invoke.h
new file mode 100644
index 00000000..0fc242a2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/invoke.h
@@ -0,0 +1,298 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_INVOKE_H
+#define _CUDA_STD___FUNCTIONAL_INVOKE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__fwd/reference_wrapper.h>
+#include <cuda/std/__type_traits/decay.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_core_convertible.h>
+#include <cuda/std/__type_traits/is_member_function_pointer.h>
+#include <cuda/std/__type_traits/is_member_object_pointer.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/nat.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct __any
+{
+  _CCCL_API inline __any(...);
+};
+
+template <class _DecayedFp>
+struct __member_pointer_class_type
+{};
+
+template <class _Ret, class _ClassType>
+struct __member_pointer_class_type<_Ret _ClassType::*>
+{
+  using type = _ClassType;
+};
+
+template <class _DecayedFp>
+using __member_pointer_class_type_t = typename __member_pointer_class_type<_DecayedFp>::type;
+
+template <class _Fp,
+          class _A0,
+          class _DecayFp = decay_t<_Fp>,
+          class _DecayA0 = decay_t<_A0>,
+          class _ClassT  = __member_pointer_class_type_t<_DecayFp>>
+using __enable_if_bullet1 = enable_if_t<is_member_function_pointer_v<_DecayFp> && is_base_of_v<_ClassT, _DecayA0>>;
+
+template <class _Fp, class _A0, class _DecayFp = decay_t<_Fp>, class _DecayA0 = decay_t<_A0>>
+using __enable_if_bullet2 =
+  enable_if_t<is_member_function_pointer_v<_DecayFp> && __is_cuda_std_reference_wrapper_v<_DecayA0>>;
+
+template <class _Fp,
+          class _A0,
+          class _DecayFp = decay_t<_Fp>,
+          class _DecayA0 = decay_t<_A0>,
+          class _ClassT  = __member_pointer_class_type_t<_DecayFp>>
+using __enable_if_bullet3 = enable_if_t<is_member_function_pointer_v<_DecayFp> && !is_base_of_v<_ClassT, _DecayA0>
+                                        && !__is_cuda_std_reference_wrapper_v<_DecayA0>>;
+
+template <class _Fp,
+          class _A0,
+          class _DecayFp = decay_t<_Fp>,
+          class _DecayA0 = decay_t<_A0>,
+          class _ClassT  = __member_pointer_class_type_t<_DecayFp>>
+using __enable_if_bullet4 = enable_if_t<is_member_object_pointer_v<_DecayFp> && is_base_of_v<_ClassT, _DecayA0>>;
+
+template <class _Fp, class _A0, class _DecayFp = decay_t<_Fp>, class _DecayA0 = decay_t<_A0>>
+using __enable_if_bullet5 =
+  enable_if_t<is_member_object_pointer_v<_DecayFp> && __is_cuda_std_reference_wrapper_v<_DecayA0>>;
+
+template <class _Fp,
+          class _A0,
+          class _DecayFp = decay_t<_Fp>,
+          class _DecayA0 = decay_t<_A0>,
+          class _ClassT  = __member_pointer_class_type_t<_DecayFp>>
+using __enable_if_bullet6 = enable_if_t<is_member_object_pointer_v<_DecayFp> && !is_base_of_v<_ClassT, _DecayA0>
+                                        && !__is_cuda_std_reference_wrapper_v<_DecayA0>>;
+
+// __invoke forward declarations
+
+// fall back - none of the bullets
+
+template <class... _Args>
+_CCCL_API inline __nat __invoke(__any, _Args&&... __args);
+
+// bullets 1, 2 and 3
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class... _Args, class = __enable_if_bullet1<_Fp, _A0>>
+_CCCL_API constexpr decltype((::cuda::std::declval<_A0>()
+                              .*::cuda::std::declval<_Fp>())(::cuda::std::declval<_Args>()...))
+__invoke(_Fp&& __f,
+         _A0&& __a0,
+         _Args&&... __args) noexcept(noexcept((static_cast<_A0&&>(__a0).*__f)(static_cast<_Args&&>(__args)...)))
+{
+  return (static_cast<_A0&&>(__a0).*__f)(static_cast<_Args&&>(__args)...);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class... _Args, class = __enable_if_bullet2<_Fp, _A0>>
+_CCCL_API constexpr decltype((::cuda::std::declval<_A0>().get()
+                              .*::cuda::std::declval<_Fp>())(::cuda::std::declval<_Args>()...))
+__invoke(_Fp&& __f, _A0&& __a0, _Args&&... __args) noexcept(noexcept((__a0.get().*__f)(static_cast<_Args&&>(__args)...)))
+{
+  return (__a0.get().*__f)(static_cast<_Args&&>(__args)...);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class... _Args, class = __enable_if_bullet3<_Fp, _A0>>
+_CCCL_API constexpr decltype(((*::cuda::std::declval<_A0>())
+                              .*::cuda::std::declval<_Fp>())(::cuda::std::declval<_Args>()...))
+__invoke(_Fp&& __f,
+         _A0&& __a0,
+         _Args&&... __args) noexcept(noexcept(((*static_cast<_A0&&>(__a0)).*__f)(static_cast<_Args&&>(__args)...)))
+{
+  return ((*static_cast<_A0&&>(__a0)).*__f)(static_cast<_Args&&>(__args)...);
+}
+
+// bullets 4, 5 and 6
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class = __enable_if_bullet4<_Fp, _A0>>
+_CCCL_API constexpr decltype(::cuda::std::declval<_A0>().*::cuda::std::declval<_Fp>())
+__invoke(_Fp&& __f, _A0&& __a0) noexcept(noexcept(static_cast<_A0&&>(__a0).*__f))
+{
+  return static_cast<_A0&&>(__a0).*__f;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class = __enable_if_bullet5<_Fp, _A0>>
+_CCCL_API constexpr decltype(::cuda::std::declval<_A0>().get().*::cuda::std::declval<_Fp>())
+__invoke(_Fp&& __f, _A0&& __a0) noexcept(noexcept(__a0.get().*__f))
+{
+  return __a0.get().*__f;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class _A0, class = __enable_if_bullet6<_Fp, _A0>>
+_CCCL_API constexpr decltype((*::cuda::std::declval<_A0>()).*::cuda::std::declval<_Fp>())
+__invoke(_Fp&& __f, _A0&& __a0) noexcept(noexcept((*static_cast<_A0&&>(__a0)).*__f))
+{
+  return (*static_cast<_A0&&>(__a0)).*__f;
+}
+
+// bullet 7
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Fp, class... _Args>
+_CCCL_API constexpr decltype(::cuda::std::declval<_Fp>()(::cuda::std::declval<_Args>()...))
+__invoke(_Fp&& __f, _Args&&... __args) noexcept(noexcept(static_cast<_Fp&&>(__f)(static_cast<_Args&&>(__args)...)))
+{
+  return static_cast<_Fp&&>(__f)(static_cast<_Args&&>(__args)...);
+}
+
+// __is_invocable
+template <class _Fp, class... _Args>
+using __invoke_result_t =
+  decltype(::cuda::std::__invoke(::cuda::std::declval<_Fp>(), ::cuda::std::declval<_Args>()...));
+
+template <class _Fp, class... _Args>
+_CCCL_CONCEPT __is_invocable =
+  _CCCL_REQUIRES_EXPR((_Fp, variadic _Args))(requires(!is_same_v<__nat, __invoke_result_t<_Fp, _Args...>>));
+
+template <class _Ret, class _Fp, class... _Args>
+_CCCL_CONCEPT __is_invocable_r = _CCCL_REQUIRES_EXPR((_Ret, _Fp, variadic _Args))(
+  requires(__is_invocable<_Fp, _Args...>),
+  requires((is_void_v<_Ret> || __is_core_convertible<__invoke_result_t<_Fp, _Args...>, _Ret>::value)));
+
+template <class _Fp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT invoke_result //
+    : public enable_if<__is_invocable<_Fp, _Args...>, __invoke_result_t<_Fp, _Args...>>
+{
+#if _CCCL_CUDA_COMPILER(NVCC) && defined(__CUDACC_EXTENDED_LAMBDA__) && !_CCCL_DEVICE_COMPILATION()
+#  if _CCCL_CUDACC_BELOW(12, 3)
+  static_assert(!__nv_is_extended_device_lambda_closure_type(remove_cvref_t<_Fp>),
+                "Attempt to use an extended __device__ lambda in a context "
+                "that requires querying its return type in host code. Use a "
+                "named function object, an extended __host__ __device__ lambda, or "
+                "cuda::proclaim_return_type instead.");
+#  else // ^^^ _CCCL_CUDACC_BELOW(12, 3) ^^^ / vvv _CCCL_CUDACC_AT_LEAST(12, 3) vvv
+  static_assert(
+    !__nv_is_extended_device_lambda_closure_type(remove_cvref_t<_Fp>)
+      || __nv_is_extended_host_device_lambda_closure_type(remove_cvref_t<_Fp>)
+      || __nv_is_extended_device_lambda_with_preserved_return_type(remove_cvref_t<_Fp>),
+    "Attempt to use an extended __device__ lambda in a context "
+    "that requires querying its return type in host code. Use a "
+    "named function object, an extended __host__ __device__ lambda, "
+    "cuda::proclaim_return_type, or an extended __device__ lambda "
+    "with a trailing return type instead ([] __device__ (...) -> RETURN_TYPE {...}).");
+#  endif // _CCCL_CUDACC_AT_LEAST(12, 3)
+#endif
+};
+
+// is_invocable
+
+template <class _Fn, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_invocable : bool_constant<__is_invocable<_Fn, _Args...>>
+{};
+
+template <class _Ret, class _Fn, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_invocable_r : bool_constant<__is_invocable_r<_Ret, _Fn, _Args...>>
+{};
+
+template <class _Fn, class... _Args>
+inline constexpr bool is_invocable_v = __is_invocable<_Fn, _Args...>;
+
+template <class _Ret, class _Fn, class... _Args>
+inline constexpr bool is_invocable_r_v = __is_invocable_r<_Ret, _Fn, _Args...>;
+
+// is_nothrow_invocable
+
+template <class _Tp>
+_CCCL_API constexpr void __cccl_test_noexcept_conversion(_Tp) noexcept;
+
+template <bool _IsInvocable, bool _IsCVVoid, class _Ret, class _Fp, class... _Args>
+inline constexpr bool __nothrow_invocable_r_imp = false;
+
+template <class _Ret, class _Fp, class... _Args>
+inline constexpr bool __nothrow_invocable_r_imp<true, false, _Ret, _Fp, _Args...> =
+  noexcept(::cuda::std::__cccl_test_noexcept_conversion<_Ret>(
+    ::cuda::std::__invoke(declval<_Fp>(), ::cuda::std::declval<_Args>()...)));
+
+template <class _Ret, class _Fp, class... _Args>
+inline constexpr bool __nothrow_invocable_r_imp<true, true, _Ret, _Fp, _Args...> =
+  noexcept(::cuda::std::__invoke(::cuda::std::declval<_Fp>(), ::cuda::std::declval<_Args>()...));
+
+template <class _Fp, class... _Args>
+inline constexpr bool is_nothrow_invocable_v =
+  __nothrow_invocable_r_imp<__is_invocable<_Fp, _Args...>, true, void, _Fp, _Args...>;
+
+template <class _Ret, class _Fp, class... _Args>
+inline constexpr bool is_nothrow_invocable_r_v =
+  __nothrow_invocable_r_imp<__is_invocable_r<_Ret, _Fp, _Args...>, is_void_v<_Ret>, _Ret, _Fp, _Args...>;
+
+template <class _Fn, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_invocable : bool_constant<is_nothrow_invocable_v<_Fn, _Args...>>
+{};
+
+template <class _Ret, class _Fn, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_invocable_r : bool_constant<is_nothrow_invocable_r_v<_Ret, _Fn, _Args...>>
+{};
+
+// Not going directly through __invoke_result_t because we want the additional device lambda checks in invoke_result
+template <class _Fn, class... _Args>
+using invoke_result_t = typename invoke_result<_Fn, _Args...>::type;
+
+template <class _Fn, class... _Args>
+_CCCL_API constexpr invoke_result_t<_Fn, _Args...>
+invoke(_Fn&& __f, _Args&&... __args) noexcept(is_nothrow_invocable_v<_Fn, _Args...>)
+{
+  return ::cuda::std::__invoke(::cuda::std::forward<_Fn>(__f), ::cuda::std::forward<_Args>(__args)...);
+}
+
+_CCCL_TEMPLATE(class _Ret, class _Fn, class... _Args)
+_CCCL_REQUIRES(is_invocable_r_v<_Ret, _Fn, _Args...>)
+_CCCL_API constexpr _Ret invoke_r(_Fn&& __f, _Args&&... __args) noexcept(is_nothrow_invocable_r_v<_Ret, _Fn, _Args...>)
+{
+  if constexpr (is_void_v<_Ret>)
+  {
+    ::cuda::std::__invoke(::cuda::std::forward<_Fn>(__f), ::cuda::std::forward<_Args>(__args)...);
+  }
+  else
+  {
+    return ::cuda::std::__invoke(::cuda::std::forward<_Fn>(__f), ::cuda::std::forward<_Args>(__args)...);
+  }
+}
+
+/// The type of intermediate accumulator (according to P2322R6)
+template <typename Invocable, typename InputT, typename InitT = InputT>
+using __accumulator_t = decay_t<invoke_result_t<Invocable, InitT, InputT>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_INVOKE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/reference_wrapper.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/reference_wrapper.h
new file mode 100644
index 00000000..ed56284d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/reference_wrapper.h
@@ -0,0 +1,116 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_REFERENCE_WRAPPER_H
+#define _CUDA_STD___FUNCTIONAL_REFERENCE_WRAPPER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__functional/weak_result_type.h>
+#include <cuda/std/__fwd/reference_wrapper.h>
+#include <cuda/std/__memory/addressof.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+class _CCCL_TYPE_VISIBILITY_DEFAULT reference_wrapper : public __weak_result_type<_Tp>
+{
+public:
+  // types
+  using type = _Tp;
+
+private:
+  type* __f_{};
+
+  static _CCCL_API void __fun(_Tp&) noexcept;
+  static void __fun(_Tp&&) = delete; // NOLINT(modernize-use-equals-delete)
+
+public:
+  // NOLINTBEGIN(bugprone-forwarding-reference-overload)
+  template <
+    class _Up,
+    class = enable_if_t<!__is_same_uncvref<_Up, reference_wrapper>::value, decltype(__fun(::cuda::std::declval<_Up>()))>>
+  _CCCL_API constexpr reference_wrapper(_Up&& __u) noexcept(noexcept(__fun(::cuda::std::declval<_Up>())))
+  {
+    type& __f = static_cast<_Up&&>(__u);
+    __f_      = ::cuda::std::addressof(__f);
+  }
+  // NOLINTEND(bugprone-forwarding-reference-overload)
+
+  // access
+  _CCCL_API constexpr operator type&() const noexcept
+  {
+    return *__f_;
+  }
+  [[nodiscard]] _CCCL_API constexpr type& get() const noexcept
+  {
+    return *__f_;
+  }
+
+  // invoke
+  template <class... _ArgTypes>
+  _CCCL_API constexpr invoke_result_t<type&, _ArgTypes...> operator()(_ArgTypes&&... __args) const
+    noexcept(is_nothrow_invocable_v<_Tp&, _ArgTypes...>)
+  {
+    return ::cuda::std::invoke(get(), ::cuda::std::forward<_ArgTypes>(__args)...);
+  }
+};
+
+template <class _Tp>
+_CCCL_DEDUCTION_GUIDE_ATTRIBUTES reference_wrapper(_Tp&) -> reference_wrapper<_Tp>;
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr reference_wrapper<_Tp> ref(_Tp& __t) noexcept
+{
+  return reference_wrapper<_Tp>(__t);
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr reference_wrapper<_Tp> ref(reference_wrapper<_Tp> __t) noexcept
+{
+  return __t;
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr reference_wrapper<const _Tp> cref(const _Tp& __t) noexcept
+{
+  return reference_wrapper<const _Tp>(__t);
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr reference_wrapper<const _Tp> cref(reference_wrapper<_Tp> __t) noexcept
+{
+  return __t;
+}
+
+template <class _Tp>
+void ref(const _Tp&&) = delete;
+template <class _Tp>
+void cref(const _Tp&&) = delete;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_REFERENCE_WRAPPER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unary_function.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unary_function.h
new file mode 100644
index 00000000..e405b62c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unary_function.h
@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_UNARY_FUNCTION_H
+#define _CUDA_STD___FUNCTIONAL_UNARY_FUNCTION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION)
+
+template <class _Arg, class _Result>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT CCCL_DEPRECATED unary_function
+{
+  using argument_type = _Arg;
+  using result_type   = _Result;
+};
+
+#endif // _LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION
+
+template <class _Arg, class _Result>
+struct __unary_function_keep_layout_base
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using argument_type CCCL_DEPRECATED = _Arg;
+  using result_type CCCL_DEPRECATED   = _Result;
+#endif
+};
+
+#if defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION)
+
+_CCCL_SUPPRESS_DEPRECATED_PUSH
+_CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+template <class _Arg, class _Result>
+using __unary_function = unary_function<_Arg, _Result>;
+_CCCL_SUPPRESS_DEPRECATED_POP
+
+#else
+template <class _Arg, class _Result>
+using __unary_function = __unary_function_keep_layout_base<_Arg, _Result>;
+#endif // !_LIBCUDACXX_ENABLE_CXX17_REMOVED_UNARY_BINARY_FUNCTION
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_UNARY_FUNCTION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unwrap_ref.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unwrap_ref.h
new file mode 100644
index 00000000..d9b24ebb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/unwrap_ref.h
@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_UNWRAP_REF_H
+#define _CUDA_STD___FUNCTIONAL_UNWRAP_REF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/reference_wrapper.h>
+#include <cuda/std/__type_traits/decay.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct unwrap_reference
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+template <class _Tp>
+struct unwrap_reference<reference_wrapper<_Tp>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp&;
+};
+
+template <class _Tp>
+using unwrap_reference_t = typename unwrap_reference<_Tp>::type;
+
+template <class _Tp>
+struct unwrap_ref_decay : unwrap_reference<decay_t<_Tp>>
+{};
+
+template <class _Tp>
+using unwrap_ref_decay_t = typename unwrap_ref_decay<_Tp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_UNWRAP_REF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/weak_result_type.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/weak_result_type.h
new file mode 100644
index 00000000..40ce6801
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__functional/weak_result_type.h
@@ -0,0 +1,262 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_WEAK_RESULT_TYPE_H
+#define _CUDA_STD___FUNCTIONAL_WEAK_RESULT_TYPE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__functional/binary_function.h>
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__functional/unary_function.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_result_type = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::result_type));
+
+// __weak_result_type
+
+template <class _Tp>
+struct __derives_from_unary_function
+{
+private:
+  struct __two
+  {
+    char __lx;
+    char __lxx;
+  };
+  static _CCCL_API inline __two __test(...);
+  template <class _Ap, class _Rp>
+  static _CCCL_API inline __unary_function<_Ap, _Rp> __test(const volatile __unary_function<_Ap, _Rp>*);
+
+public:
+  static const bool value = !is_same_v<decltype(__test((_Tp*) nullptr)), __two>;
+  using type              = decltype(__test((_Tp*) nullptr));
+};
+
+template <class _Tp>
+struct __derives_from_binary_function
+{
+private:
+  struct __two
+  {
+    char __lx;
+    char __lxx;
+  };
+  static __two _CCCL_API inline __test(...);
+  template <class _A1, class _A2, class _Rp>
+  static _CCCL_API inline __binary_function<_A1, _A2, _Rp> __test(const volatile __binary_function<_A1, _A2, _Rp>*);
+
+public:
+  static const bool value = !is_same_v<decltype(__test((_Tp*) nullptr)), __two>;
+  using type              = decltype(__test((_Tp*) nullptr));
+};
+
+template <class _Tp, bool = __derives_from_unary_function<_Tp>::value>
+struct __maybe_derive_from_unary_function // bool is true
+    : public __derives_from_unary_function<_Tp>::type
+{};
+
+template <class _Tp>
+struct __maybe_derive_from_unary_function<_Tp, false>
+{};
+
+template <class _Tp, bool = __derives_from_binary_function<_Tp>::value>
+struct __maybe_derive_from_binary_function // bool is true
+    : public __derives_from_binary_function<_Tp>::type
+{};
+
+template <class _Tp>
+struct __maybe_derive_from_binary_function<_Tp, false>
+{};
+
+template <class _Tp, bool = __has_member_result_type<_Tp>>
+struct __weak_result_type_imp // bool is true
+    : public __maybe_derive_from_unary_function<_Tp>
+    , public __maybe_derive_from_binary_function<_Tp>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = typename _Tp::result_type;
+#endif
+};
+
+template <class _Tp>
+struct __weak_result_type_imp<_Tp, false>
+    : public __maybe_derive_from_unary_function<_Tp>
+    , public __maybe_derive_from_binary_function<_Tp>
+{};
+
+template <class _Tp>
+struct __weak_result_type : public __weak_result_type_imp<_Tp>
+{};
+
+// 0 argument case
+
+template <class _Rp>
+struct __weak_result_type<_Rp()>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp>
+struct __weak_result_type<_Rp (&)()>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp>
+struct __weak_result_type<_Rp (*)()>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+// 1 argument case
+
+template <class _Rp, class _A1>
+struct __weak_result_type<_Rp(_A1)> : public __unary_function<_A1, _Rp>
+{};
+
+template <class _Rp, class _A1>
+struct __weak_result_type<_Rp (&)(_A1)> : public __unary_function<_A1, _Rp>
+{};
+
+template <class _Rp, class _A1>
+struct __weak_result_type<_Rp (*)(_A1)> : public __unary_function<_A1, _Rp>
+{};
+
+template <class _Rp, class _Cp>
+struct __weak_result_type<_Rp (_Cp::*)()> : public __unary_function<_Cp*, _Rp>
+{};
+
+template <class _Rp, class _Cp>
+struct __weak_result_type<_Rp (_Cp::*)() const> : public __unary_function<const _Cp*, _Rp>
+{};
+
+template <class _Rp, class _Cp>
+struct __weak_result_type<_Rp (_Cp::*)() volatile> : public __unary_function<volatile _Cp*, _Rp>
+{};
+
+template <class _Rp, class _Cp>
+struct __weak_result_type<_Rp (_Cp::*)() const volatile> : public __unary_function<const volatile _Cp*, _Rp>
+{};
+
+// 2 argument case
+
+template <class _Rp, class _A1, class _A2>
+struct __weak_result_type<_Rp(_A1, _A2)> : public __binary_function<_A1, _A2, _Rp>
+{};
+
+template <class _Rp, class _A1, class _A2>
+struct __weak_result_type<_Rp (*)(_A1, _A2)> : public __binary_function<_A1, _A2, _Rp>
+{};
+
+template <class _Rp, class _A1, class _A2>
+struct __weak_result_type<_Rp (&)(_A1, _A2)> : public __binary_function<_A1, _A2, _Rp>
+{};
+
+template <class _Rp, class _Cp, class _A1>
+struct __weak_result_type<_Rp (_Cp::*)(_A1)> : public __binary_function<_Cp*, _A1, _Rp>
+{};
+
+template <class _Rp, class _Cp, class _A1>
+struct __weak_result_type<_Rp (_Cp::*)(_A1) const> : public __binary_function<const _Cp*, _A1, _Rp>
+{};
+
+template <class _Rp, class _Cp, class _A1>
+struct __weak_result_type<_Rp (_Cp::*)(_A1) volatile> : public __binary_function<volatile _Cp*, _A1, _Rp>
+{};
+
+template <class _Rp, class _Cp, class _A1>
+struct __weak_result_type<_Rp (_Cp::*)(_A1) const volatile> : public __binary_function<const volatile _Cp*, _A1, _Rp>
+{};
+
+// 3 or more arguments
+
+template <class _Rp, class _A1, class _A2, class _A3, class... _A4>
+struct __weak_result_type<_Rp(_A1, _A2, _A3, _A4...)>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _A1, class _A2, class _A3, class... _A4>
+struct __weak_result_type<_Rp (&)(_A1, _A2, _A3, _A4...)>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _A1, class _A2, class _A3, class... _A4>
+struct __weak_result_type<_Rp (*)(_A1, _A2, _A3, _A4...)>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _Cp, class _A1, class _A2, class... _A3>
+struct __weak_result_type<_Rp (_Cp::*)(_A1, _A2, _A3...)>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _Cp, class _A1, class _A2, class... _A3>
+struct __weak_result_type<_Rp (_Cp::*)(_A1, _A2, _A3...) const>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _Cp, class _A1, class _A2, class... _A3>
+struct __weak_result_type<_Rp (_Cp::*)(_A1, _A2, _A3...) volatile>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+template <class _Rp, class _Cp, class _A1, class _A2, class... _A3>
+struct __weak_result_type<_Rp (_Cp::*)(_A1, _A2, _A3...) const volatile>
+{
+#if _CCCL_STD_VER <= 2017 || defined(_LIBCUDACXX_ENABLE_CXX20_REMOVED_BINDER_TYPEDEFS)
+  using result_type _CCCL_NODEBUG_ALIAS CCCL_DEPRECATED = _Rp;
+#endif
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_WEAK_RESULT_TYPE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/array.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/array.h
new file mode 100644
index 00000000..887c9d35
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/array.h
@@ -0,0 +1,68 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_ARRAY_H
+#define _CUDA_STD___FWD_ARRAY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+#  if _CCCL_HOST_STD_LIB(STL)
+template <class _Tp, size_t _Size>
+class array;
+#  else // ^^^ _CCCL_HOST_STD_LIB(STL) ^^^ / vvv !_CCCL_HOST_STD_LIB(STL) vvv
+template <class _Tp, size_t _Size>
+struct array;
+#  endif // !_CCCL_HOST_STD_LIB(STL)
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// ::cuda::std:: forward declaration
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, size_t _Size>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT array;
+
+template <class _Tp>
+inline constexpr bool __is_cuda_std_array_v = false;
+
+template <class _Tp, size_t _Sz>
+inline constexpr bool __is_cuda_std_array_v<array<_Tp, _Sz>> = true;
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Tp>
+inline constexpr bool __is_std_array_v = false;
+
+template <class _Tp, size_t _Sz>
+inline constexpr bool __is_std_array_v<::std::array<_Tp, _Sz>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_ARRAY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/complex.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/complex.h
new file mode 100644
index 00000000..363afb90
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/complex.h
@@ -0,0 +1,75 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_COMPLEX_H
+#define _CUDA_STD___FWD_COMPLEX_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class>
+class complex;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+class _CCCL_TYPE_VISIBILITY_DEFAULT complex;
+
+// __is_std_complex_v
+
+template <class _Tp>
+inline constexpr bool __is_std_complex_v = false;
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Tp>
+inline constexpr bool __is_std_complex_v<const _Tp> = __is_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_std_complex_v<volatile _Tp> = __is_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_std_complex_v<const volatile _Tp> = __is_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_std_complex_v<::std::complex<_Tp>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// __is_cuda_std_complex_v
+
+template <class _Tp>
+inline constexpr bool __is_cuda_std_complex_v = false;
+template <class _Tp>
+inline constexpr bool __is_cuda_std_complex_v<const _Tp> = __is_cuda_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_std_complex_v<volatile _Tp> = __is_cuda_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_std_complex_v<const volatile _Tp> = __is_cuda_std_complex_v<_Tp>;
+template <class _Tp>
+inline constexpr bool __is_cuda_std_complex_v<complex<_Tp>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_COMPLEX_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/format.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/format.h
new file mode 100644
index 00000000..e96561b1
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/format.h
@@ -0,0 +1,120 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_FORMAT_H
+#define _CUDA_STD___FWD_FORMAT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/iterator.h>
+#include <cuda/std/__type_traits/type_identity.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if __cpp_lib_format >= 201907L
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class, class>
+struct formatter;
+
+_CCCL_END_NAMESPACE_STD
+
+#endif // __cpp_lib_format >= 201907L
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _CharT>
+class _CCCL_TYPE_VISIBILITY_DEFAULT basic_format_parse_context;
+
+using format_parse_context = basic_format_parse_context<char>;
+#if _CCCL_HAS_WCHAR_T()
+using wformat_parse_context = basic_format_parse_context<wchar_t>;
+#endif // _CCCL_HAS_WCHAR_T()
+
+template <class _CharT>
+class _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_PREFERRED_NAME(format_parse_context)
+#if _CCCL_HAS_WCHAR_T()
+  _CCCL_PREFERRED_NAME(wformat_parse_context)
+#endif // _CCCL_HAS_WCHAR_T()
+    basic_format_parse_context;
+
+template <class _CharT>
+class __fmt_output_buffer;
+
+template <class _Context>
+class _CCCL_TYPE_VISIBILITY_DEFAULT basic_format_arg;
+
+template <class _OutIt, class _CharT>
+class _CCCL_TYPE_VISIBILITY_DEFAULT basic_format_context;
+
+using format_context = basic_format_context<__back_insert_iterator<__fmt_output_buffer<char>>, char>;
+#if _CCCL_HAS_WCHAR_T()
+using wformat_context = basic_format_context<__back_insert_iterator<__fmt_output_buffer<wchar_t>>, wchar_t>;
+#endif // _CCCL_HAS_WCHAR_T()
+
+template <class _OutIt, class _CharT>
+class _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_PREFERRED_NAME(format_context)
+#if _CCCL_HAS_WCHAR_T()
+  _CCCL_PREFERRED_NAME(wformat_context)
+#endif // _CCCL_HAS_WCHAR_T()
+    basic_format_context;
+
+template <class _Context>
+class _CCCL_TYPE_VISIBILITY_DEFAULT basic_format_args;
+
+using format_args = basic_format_args<format_context>;
+#if _CCCL_HAS_WCHAR_T()
+using wformat_args = basic_format_args<wformat_context>;
+#endif // _CCCL_HAS_WCHAR_T()
+
+template <class _Context>
+class _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_PREFERRED_NAME(format_args)
+#if _CCCL_HAS_WCHAR_T()
+  _CCCL_PREFERRED_NAME(wformat_args)
+#endif // _CCCL_HAS_WCHAR_T()
+    basic_format_args;
+
+template <class _CharT, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT basic_format_string;
+
+template <class... _Args>
+using format_string = basic_format_string<char, type_identity_t<_Args>...>;
+
+#if _CCCL_HAS_WCHAR_T()
+template <class... _Args>
+using wformat_string = basic_format_string<wchar_t, type_identity_t<_Args>...>;
+#endif // _CCCL_HAS_WCHAR_T()
+
+template <class _OutIt>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT format_to_n_result;
+
+enum class range_format
+{
+  disabled,
+  map,
+  set,
+  sequence,
+  string,
+  debug_string,
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_FORMAT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/fp.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/fp.h
new file mode 100644
index 00000000..7aee6d06
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/fp.h
@@ -0,0 +1,37 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_FP_H
+#define _CUDA_STD___FWD_FP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+enum class __fp_format;
+
+template <__fp_format _Fmt>
+class __cccl_fp;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_FP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/get.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/get.h
new file mode 100644
index 00000000..26f27725
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/get.h
@@ -0,0 +1,132 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_GET_H
+#define _CUDA_STD___FWD_GET_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/copyable.h>
+#include <cuda/std/__fwd/array.h>
+#include <cuda/std/__fwd/complex.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__fwd/subrange.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <size_t _Ip, class... _Tp>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, tuple<_Tp...>>& get(tuple<_Tp...>&) noexcept;
+
+template <size_t _Ip, class... _Tp>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, tuple<_Tp...>>& get(const tuple<_Tp...>&) noexcept;
+
+template <size_t _Ip, class... _Tp>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, tuple<_Tp...>>&& get(tuple<_Tp...>&&) noexcept;
+
+template <size_t _Ip, class... _Tp>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, tuple<_Tp...>>&& get(const tuple<_Tp...>&&) noexcept;
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, pair<_T1, _T2>>& get(pair<_T1, _T2>&) noexcept;
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, pair<_T1, _T2>>& get(const pair<_T1, _T2>&) noexcept;
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, pair<_T1, _T2>>&& get(pair<_T1, _T2>&&) noexcept;
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, pair<_T1, _T2>>&& get(const pair<_T1, _T2>&&) noexcept;
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr _Tp& get(array<_Tp, _Size>&) noexcept;
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr const _Tp& get(const array<_Tp, _Size>&) noexcept;
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr _Tp&& get(array<_Tp, _Size>&&) noexcept;
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr const _Tp&& get(const array<_Tp, _Size>&&) noexcept;
+
+template <size_t _Ip, class _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp& get(complex<_Tp>&) noexcept;
+
+template <size_t _Ip, class _Tp>
+[[nodiscard]] _CCCL_API constexpr _Tp&& get(complex<_Tp>&&) noexcept;
+
+template <size_t _Ip, class _Tp>
+[[nodiscard]] _CCCL_API constexpr const _Tp& get(const complex<_Tp>&) noexcept;
+
+template <size_t _Ip, class _Tp>
+[[nodiscard]] _CCCL_API constexpr const _Tp&& get(const complex<_Tp>&&) noexcept;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+#if _CCCL_HAS_CONCEPTS()
+template <size_t _Index, class _Iter, class _Sent, subrange_kind _Kind>
+  requires((_Index == 0) && copyable<_Iter>) || (_Index == 1)
+#else // ^^^ C++20 ^^^ / vvv C++17 vvv
+template <size_t _Index,
+          class _Iter,
+          class _Sent,
+          subrange_kind _Kind,
+          enable_if_t<((_Index == 0) && copyable<_Iter>) || (_Index == 1), int> = 0>
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+_CCCL_API constexpr auto get(const subrange<_Iter, _Sent, _Kind>& __subrange);
+
+#if _CCCL_HAS_CONCEPTS()
+template <size_t _Index, class _Iter, class _Sent, subrange_kind _Kind>
+  requires(_Index < 2)
+#else // ^^^ C++20 ^^^ / vvv C++17 vvv
+template <size_t _Index,
+          class _Iter,
+          class _Sent,
+          subrange_kind _Kind,
+          enable_if_t<_Index<2, int> = 0>
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+_CCCL_API constexpr auto get(subrange<_Iter, _Sent, _Kind>&& __subrange);
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+using ::cuda::std::ranges::get;
+
+// Explicitly rely on ADL, mostly for constructors of host STL types, where we cannot squeze using ::cuda::std::get; in
+_CCCL_EXEC_CHECK_DISABLE
+template <size_t _Ip, class _TupleLike>
+[[nodiscard]] _CCCL_API constexpr decltype(auto) __adl_get(_TupleLike&& __t) noexcept
+{
+  using ::cuda::std::get;
+  return get<_Ip>(::cuda::std::forward<_TupleLike>(__t));
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_GET_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/iterator.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/iterator.h
new file mode 100644
index 00000000..7aeb940a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/iterator.h
@@ -0,0 +1,45 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_ITERATOR_H
+#define _CUDA_STD___FWD_ITERATOR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Container>
+class _CCCL_TYPE_VISIBILITY_DEFAULT __back_insert_iterator;
+
+template <class, class = void>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT iterator_traits;
+
+_LIBCUDACXX_BEGIN_HIDDEN_FRIEND_NAMESPACE
+
+template <class _Iter>
+class _CCCL_TYPE_VISIBILITY_DEFAULT reverse_iterator;
+
+_LIBCUDACXX_END_HIDDEN_FRIEND_NAMESPACE(reverse_iterator)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_ITERATOR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/pair.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/pair.h
new file mode 100644
index 00000000..e526d6b1
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/pair.h
@@ -0,0 +1,53 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_PAIR_H
+#define _CUDA_STD___FWD_PAIR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class, class>
+struct pair;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class, class>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT pair;
+
+template <class>
+inline constexpr bool __is_cuda_std_pair = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __is_cuda_std_pair<pair<_Tp, _Up>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_PAIR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/reference_wrapper.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/reference_wrapper.h
new file mode 100644
index 00000000..e95bf5ab
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/reference_wrapper.h
@@ -0,0 +1,52 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_REFERENCE_WRAPPER_H
+#define _CUDA_STD___FWD_REFERENCE_WRAPPER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+// std:: forward declarations
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class>
+class reference_wrapper;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+// cuda::std:: forward declarations
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+class _CCCL_TYPE_VISIBILITY_DEFAULT reference_wrapper;
+
+template <class _Tp>
+inline constexpr bool __is_cuda_std_reference_wrapper_v = false;
+template <class _Tp>
+inline constexpr bool __is_cuda_std_reference_wrapper_v<reference_wrapper<_Tp>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_REFERENCE_WRAPPER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/span.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/span.h
new file mode 100644
index 00000000..a7d28727
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/span.h
@@ -0,0 +1,45 @@
+// -*- C++ -*-
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_SPAN_H
+#define _CUDA_STD___FWD_SPAN_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+inline constexpr size_t dynamic_extent = static_cast<size_t>(-1);
+
+template <typename _Tp, size_t _Extent = dynamic_extent>
+class span;
+
+template <class _Tp>
+inline constexpr bool __is_cuda_std_span_v = false;
+
+template <class _Tp, size_t _Extent>
+inline constexpr bool __is_cuda_std_span_v<span<_Tp, _Extent>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_SPAN_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/subrange.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/subrange.h
new file mode 100644
index 00000000..ccd8aadc
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/subrange.h
@@ -0,0 +1,65 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_SUBRANGE_H
+#define _CUDA_STD___FWD_SUBRANGE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/concepts.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+enum class _CCCL_TYPE_VISIBILITY_DEFAULT subrange_kind : bool
+{
+  unsized,
+  sized
+};
+
+#if _CCCL_HAS_CONCEPTS()
+template <input_or_output_iterator _Iter,
+          sentinel_for<_Iter> _Sent = _Iter,
+          subrange_kind _Kind       = sized_sentinel_for<_Sent, _Iter> ? subrange_kind::sized : subrange_kind::unsized>
+  requires(_Kind == subrange_kind::sized || !sized_sentinel_for<_Sent, _Iter>)
+class _CCCL_TYPE_VISIBILITY_DEFAULT subrange;
+#else // ^^^ C++20 ^^^ / vvv C++17 vvv
+template <class _Iter,
+          class _Sent         = _Iter,
+          subrange_kind _Kind = sized_sentinel_for<_Sent, _Iter> ? subrange_kind::sized : subrange_kind::unsized,
+          enable_if_t<input_or_output_iterator<_Iter>, int>                                      = 0,
+          enable_if_t<sentinel_for<_Sent, _Iter>, int>                                           = 0,
+          enable_if_t<(_Kind == subrange_kind::sized || !sized_sentinel_for<_Sent, _Iter>), int> = 0>
+class _CCCL_TYPE_VISIBILITY_DEFAULT subrange;
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class>
+inline constexpr bool __is_cuda_std_ranges_subrange_v = false;
+
+template <class _Iter, class _Sent, ::cuda::std::ranges::subrange_kind _Kind>
+inline constexpr bool __is_cuda_std_ranges_subrange_v<::cuda::std::ranges::subrange<_Iter, _Sent, _Kind>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_SUBRANGE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/tuple.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/tuple.h
new file mode 100644
index 00000000..4720dc11
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__fwd/tuple.h
@@ -0,0 +1,52 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FWD_TUPLE_H
+#define _CUDA_STD___FWD_TUPLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_HAS_HOST_STD_LIB()
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class...>
+class tuple;
+
+_CCCL_END_NAMESPACE_STD
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+class _CCCL_TYPE_VISIBILITY_DEFAULT tuple;
+
+template <class>
+inline constexpr bool __is_tuple_of_iterator_references_v = false;
+
+template <class>
+inline constexpr bool __is_cuda_std_tuple = false;
+
+template <class... _Types>
+inline constexpr bool __is_cuda_std_tuple<tuple<_Types...>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FWD_TUPLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/cstdio b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/cstdio
new file mode 100644
index 00000000..87950e63
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/cstdio
@@ -0,0 +1,28 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___HOST_STDLIB_CSTDIO
+#define _CUDA_STD___HOST_STDLIB_CSTDIO
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <cstdio>
+#endif // _CCCL_HOSTED()
+
+#endif // _CUDA_STD___HOST_STDLIB_CSTDIO
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/math.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/math.h
new file mode 100644
index 00000000..d6396304
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/math.h
@@ -0,0 +1,49 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___HOST_STDLIB_MATH_H
+#define _CUDA_STD___HOST_STDLIB_MATH_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <math.h>
+
+// Standard C++ library comes with it's own <math.h> C++ compatible header. However, if the include paths are jumbled,
+// it might happen that the original C <math.h> is found first. This is a problem because C headers define many of the
+// math functions as macros which would change our definitions. So, we check whether any of the functions are defined
+// as a macro to distinguish the C++ copatibility header from the C header.
+#  if defined(fabs) || defined(fmod) || defined(remainder) || defined(remquo) || defined(fma) || defined(fmax)      \
+    || defined(fmin) || defined(fdim) || defined(exp) || defined(exp2) || defined(expm1) || defined(log)            \
+    || defined(log10) || defined(log2) || defined(log1p) || defined(pow) || defined(sqrt) || defined(cbrt)          \
+    || defined(hypot) || defined(sin) || defined(cos) || defined(tan) || defined(asin) || defined(acos)             \
+    || defined(atan) || defined(atan2) || defined(sinh) || defined(cosh) || defined(tanh) || defined(asinh)         \
+    || defined(acosh) || defined(atanh) || defined(erf) || defined(erfc) || defined(tgamma) || defined(lgamma)      \
+    || defined(ceil) || defined(floor) || defined(trunc) || defined(round) || defined(lround) || defined(llround)   \
+    || defined(nearbyint) || defined(rint) || defined(lrint) || defined(llrint) || defined(frexp) || defined(ldexp) \
+    || defined(scalbn) || defined(scalbln) || defined(ilogb) || defined(logb) || defined(nextafter)                 \
+    || defined(nexttoward) || defined(copysign) || defined(fpclassify) || defined(isfinite) || defined(isinf)       \
+    || defined(isnan) || defined(isnormal) || defined(signbit) || defined(isgreater) || defined(isgreaterequal)     \
+    || defined(isless) || defined(islessequal) || defined(islessgreater) || defined(isunordered)
+#    error \
+      "libcu++ requires the C++ compatibility <math.h> header, not the C <math.h> header. Please, check your include paths."
+#  endif // math functions defined as macros
+
+#endif // _CCCL_HOSTED()
+
+#endif // _CUDA_STD___HOST_STDLIB_MATH_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/memory b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/memory
new file mode 100644
index 00000000..119de005
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/memory
@@ -0,0 +1,36 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___HOST_STDLIB_MEMORY
+#define _CUDA_STD___HOST_STDLIB_MEMORY
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// When nvc++ uses CCCL components as part of its implementation of
+// Standard C++ algorithms, a cycle of included files may result when CCCL code
+// tries to use a standard algorithm. The THRUST_INCLUDING_ALGORITHMS_HEADER macro
+// is defined only when CCCL is including an algorithms-related header, giving
+// the compiler a chance to detect and break the cycle of includes.
+
+#if _CCCL_HOSTED()
+#  define THRUST_INCLUDING_ALGORITHMS_HEADER
+#  include <memory>
+#  undef THRUST_INCLUDING_ALGORITHMS_HEADER
+#endif // _CCCL_HOSTED()
+
+#endif // _CUDA_STD___HOST_STDLIB_MEMORY
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/new b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/new
new file mode 100644
index 00000000..cc2db623
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/new
@@ -0,0 +1,29 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___HOST_STDLIB_NEW
+#define _CUDA_STD___HOST_STDLIB_NEW
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// HostJiT also needs <new>
+#if !_CCCL_COMPILER(NVRTC)
+#  include <new>
+#endif // !_CCCL_COMPILER(NVRTC)
+
+#endif // _CUDA_STD___HOST_STDLIB_NEW
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/stdexcept b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/stdexcept
new file mode 100644
index 00000000..053f822f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__host_stdlib/stdexcept
@@ -0,0 +1,28 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___HOST_STDLIB_STDEXCEPT
+#define _CUDA_STD___HOST_STDLIB_STDEXCEPT
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <stdexcept>
+#endif // _CCCL_HOSTED()
+
+#endif // _CUDA_STD___HOST_STDLIB_STDEXCEPT
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/atomic.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/atomic.h
new file mode 100644
index 00000000..94d87d23
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/atomic.h
@@ -0,0 +1,55 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_ATOMIC_H
+#define _CUDA_STD___INTERNAL_ATOMIC_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__internal/features.h>
+
+#if _CCCL_CUDA_COMPILATION()
+#  define _CCCL_ATOMIC_ALWAYS_LOCK_FREE(size, ptr) (size <= 8)
+#elif _CCCL_COMPILER(CLANG) || _CCCL_COMPILER(GCC)
+#  define _CCCL_ATOMIC_ALWAYS_LOCK_FREE(...) __atomic_always_lock_free(__VA_ARGS__)
+#endif // _CCCL_CUDA_COMPILER
+
+// Enable bypassing automatic storage checks in atomics when using CTK 12.2 and below and if NDEBUG is defined.
+// A compiler bug prevents the safe use of `__is_local` and PTX spacep until after 13.0.
+#ifndef _CCCL_ATOMIC_UNSAFE_AUTOMATIC_STORAGE
+#  if _CCCL_CUDACC_BELOW(13, 1) && !defined(NDEBUG)
+#    define _CCCL_ATOMIC_UNSAFE_AUTOMATIC_STORAGE
+#  endif // _CCCL_CUDACC_BELOW(13, 1)
+#endif // _CCCL_ATOMIC_UNSAFE_AUTOMATIC_STORAGE
+
+#define _CCCL_ATOMIC_FLAG_TYPE int
+
+// Clang provides 128b atomics as a builtin
+#if defined(CCCL_ENABLE_EXPERIMENTAL_HOST_ATOMICS_128B)
+#  define _CCCL_HOST_128_ATOMICS_ENABLED() 1
+#  define _CCCL_HOST_128_ATOMICS_MAYBE()   0
+// GCC does not provide 128b atomics, but they may be available as a library, this requires opt-in usage.
+// See: https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html "-mcx16" for more
+#elif _CCCL_COMPILER(CLANG) || _CCCL_COMPILER(GCC)
+#  define _CCCL_HOST_128_ATOMICS_ENABLED() 0
+#  define _CCCL_HOST_128_ATOMICS_MAYBE()   1
+#else
+#  define _CCCL_HOST_128_ATOMICS_ENABLED() 0
+#  define _CCCL_HOST_128_ATOMICS_MAYBE()   0
+#endif
+
+#endif // _CUDA_STD___INTERNAL_ATOMIC_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/cpp_dialect.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/cpp_dialect.h
new file mode 100644
index 00000000..e5d9f533
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/cpp_dialect.h
@@ -0,0 +1,44 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_CPP_DIALECT_H
+#define _CUDA_STD___INTERNAL_CPP_DIALECT_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// Define LIBCUDACXX_COMPILER_DEPRECATION macro:
+#if _CCCL_COMPILER(MSVC) || _CCCL_COMPILER(NVRTC)
+#  define LIBCUDACXX_COMP_DEPR_IMPL(msg) \
+    _CCCL_PRAGMA(message(__FILE__ ":" _CCCL_TO_STRING(__LINE__) ": warning: " #msg))
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#  define LIBCUDACXX_COMP_DEPR_IMPL(msg) _CCCL_PRAGMA(GCC warning #msg)
+#endif // !_CCCL_COMPILER(MSVC)
+
+// clang-format off
+#define LIBCUDACXX_DIALECT_DEPRECATION(REQ, CUR)                                                           \
+  LIBCUDACXX_COMP_DEPR_IMPL(                                                                               \
+    libcu++ requires at least REQ. CUR is deprecated but still supported. CUR support will be removed in a \
+      future release. Define CCCL_IGNORE_DEPRECATED_CPP_DIALECT to suppress this message.)
+// clang-format on
+
+#ifndef CCCL_IGNORE_DEPRECATED_CPP_DIALECT
+#  if _CCCL_STD_VER < 2017
+#    error libcu++ requires at least C++ 17. Define CCCL_IGNORE_DEPRECATED_CPP_DIALECT to suppress this message.
+#  endif // _CCCL_STD_VER < 2017
+#endif // CCCL_IGNORE_DEPRECATED_CPP_DIALECT
+
+#endif // _CUDA_STD___INTERNAL_CPP_DIALECT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/features.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/features.h
new file mode 100644
index 00000000..c4e4662f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/features.h
@@ -0,0 +1,127 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_FEATURES_H
+#define _CUDA_STD___INTERNAL_FEATURES_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#define _LIBCUDACXX_HAS_CXX20_CHRONO_LITERALS() (!_CCCL_COMPILER(CLANG) || _CCCL_STD_VER >= 2020)
+#define _LIBCUDACXX_HAS_MONOTONIC_CLOCK()       0
+#define _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()    0
+
+#if _CCCL_CUDA_COMPILATION() || __cpp_aligned_new < 201606
+#  define _LIBCUDACXX_HAS_ALIGNED_ALLOCATION() 0
+#else
+#  define _LIBCUDACXX_HAS_ALIGNED_ALLOCATION() 1
+#endif // !_CCCL_CUDA_COMPILATION() && __cpp_aligned_new >= 201606
+
+// We need `is_constant_evaluated` for clang and gcc. MSVC also needs extensive rework
+#if !defined(_CCCL_BUILTIN_IS_CONSTANT_EVALUATED)
+#  define _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS() 0
+#elif _CCCL_COMPILER(NVRTC)
+#  define _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS() 0
+#elif _CCCL_COMPILER(MSVC)
+#  define _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS() 0
+#elif _CCCL_CUDA_COMPILER(CLANG)
+#  define _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS() 0
+#else
+#  define _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS() 1
+#endif
+
+#if _LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS()
+#  define _CCCL_CONSTEXPR_COMPLEX constexpr
+#else
+#  define _CCCL_CONSTEXPR_COMPLEX
+#endif // !_LIBCUDACXX_HAS_CONSTEXPR_COMPLEX_OPERATIONS()
+
+#ifndef _LIBCUDACXX_HAS_NO_INCOMPLETE_RANGES
+#  define _LIBCUDACXX_HAS_NO_INCOMPLETE_RANGES
+#endif // _LIBCUDACXX_HAS_NO_INCOMPLETE_RANGES
+
+// libcu++ requires host device support for its tests. Until then restrict usage to at least 12.2
+#if _CCCL_HAS_NVFP16() && _CCCL_CTK_AT_LEAST(12, 2)
+#  define _LIBCUDACXX_HAS_NVFP16() 1
+#else
+#  define _LIBCUDACXX_HAS_NVFP16() 0
+#endif // _CCCL_HAS_NVFP16() && _CCCL_CTK_AT_LEAST(12, 2)
+
+// libcu++ requires host device support for its tests. Until then restrict usage to at least 12.2
+#if _CCCL_HAS_NVBF16() && _CCCL_CTK_AT_LEAST(12, 2)
+#  define _LIBCUDACXX_HAS_NVBF16() 1
+#else
+#  define _LIBCUDACXX_HAS_NVBF16() 0
+#endif // _CCCL_HAS_NVBF16() && _CCCL_CTK_AT_LEAST(12, 2)
+
+#if _CCCL_COMPILER(MSVC)
+#  define _CCCL_ALIGNAS_TYPE(x) alignas(x)
+#  define _CCCL_ALIGNAS(x)      __declspec(align(x))
+#elif _CCCL_HAS_FEATURE(cxx_alignas)
+#  define _CCCL_ALIGNAS_TYPE(x) alignas(x)
+#  define _CCCL_ALIGNAS(x)      alignas(x)
+#else
+#  define _CCCL_ALIGNAS_TYPE(x) __attribute__((__aligned__(alignof(x))))
+#  define _CCCL_ALIGNAS(x)      __attribute__((__aligned__(x)))
+#endif // !_CCCL_COMPILER(MSVC) && !_CCCL_HAS_FEATURE(cxx_alignas)
+
+// We can only expose constexpr allocations if the compiler supports it
+// For now disable constexpr allocation support until we can actually use
+#if 0 && __cpp_constexpr_dynamic_alloc >= 201907L && __cpp_lib_constexpr_dynamic_alloc >= 201907L \
+  && _CCCL_STD_VER >= 2020 && !_CCCL_COMPILER(NVRTC)
+#  define _CCCL_HAS_CONSTEXPR_ALLOCATION
+#  define _CCCL_CONSTEXPR_CXX20_ALLOCATION constexpr
+#else // ^^^ has constexpr allocations ^^^ / vvv no constexpr allocations vvv
+#  define _CCCL_CONSTEXPR_CXX20_ALLOCATION
+#endif // ^^^ no constexpr allocations ^^^
+
+// Enable removed C++17 features
+#if defined(_LIBCUDACXX_ENABLE_CXX17_REMOVED_FEATURES)
+#  define _LIBCUDACXX_ENABLE_CXX17_REMOVED_BINDERS
+#endif // _LIBCUDACXX_ENABLE_CXX17_REMOVED_FEATURES
+
+#ifndef _CCCL_DISABLE_ADDITIONAL_DIAGNOSTICS
+#  define _CCCL_DIAGNOSE_WARNING(_COND, _MSG) _CCCL_DIAGNOSE_IF(_COND, _MSG, "warning")
+#  define _CCCL_DIAGNOSE_ERROR(_COND, _MSG)   _CCCL_DIAGNOSE_IF(_COND, _MSG, "error")
+#else
+#  define _CCCL_DIAGNOSE_WARNING(_COND, _MSG)
+#  define _CCCL_DIAGNOSE_ERROR(_COND, _MSG)
+#endif
+
+#define _CCCL_HAS_SIMD_F32X2_INTRINSICS() \
+  (_CCCL_CUDACC_AT_LEAST(12, 8) && _CCCL_HAS_CTK() && !_CCCL_CUDA_COMPILER(CLANG))
+#define _CCCL_HAS_SIMD_F32X2_PTX() (__cccl_ptx_isa >= 860ULL)
+#define _CCCL_HAS_SIMD_F32X2() \
+  ((_CCCL_HAS_SIMD_F32X2_INTRINSICS() || _CCCL_HAS_SIMD_F32X2_PTX()) && !_CCCL_TILE_COMPILATION())
+
+// nvcc >= 12.8 already optimizes 16-bit X2 min/max operations to SIMD instructions
+#define _CCCL_HAS_SIMD_16BIT_MIN_MAX_COMPILER_OPTIMIZATION() _CCCL_CUDA_COMPILER(NVCC, >=, 12, 8)
+
+#define _CCCL_HAS_SIMD_8BIT_INTRINSICS() 0 // TODO(fbusato): CTK 13.2 produces non-optimal code for 8-bit SIMD instrs.
+#define _CCCL_HAS_SIMD_8BIT_PTX()        (__cccl_ptx_isa >= 920ULL)
+#define _CCCL_HAS_SIMD_8BIT() \
+  ((_CCCL_HAS_SIMD_8BIT_PTX() || _CCCL_HAS_SIMD_8BIT_INTRINSICS()) && !_CCCL_TILE_COMPILATION())
+
+// Third party libraries
+
+#if (__has_include(<dlpack/dlpack.h>) || __has_include(<dlpack.h>)) && \
+     !_CCCL_COMPILER(NVRTC) && !defined(CCCL_DISABLE_DLPACK)
+#  define _CCCL_HAS_DLPACK() 1
+#else // ^^^ has dlpack ^^^ / vvv no dlpack vvv
+#  define _CCCL_HAS_DLPACK() 0
+#endif // ^^^ no dlpack ^^^
+
+#endif // _CUDA_STD___INTERNAL_FEATURES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/namespaces.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/namespaces.h
new file mode 100644
index 00000000..4876dd6f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/namespaces.h
@@ -0,0 +1,188 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_NAMESPACES_H
+#define _CUDA_STD___INTERNAL_NAMESPACES_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__internal/version.h>
+
+// During the header testing, we want to check if the code is wrapped by the prologue/epilogue
+#if defined(_CCCL_HEADER_TEST)
+#  define _CCCL_PROLOGUE_INCLUDE_CHECK() \
+    static_assert(_CCCL_PROLOGUE_INCLUDED(), "missing #include <cuda/std/__cccl/prologue.h>");
+#else // ^^^ defined(_CCCL_HEADER_TEST) ^^^ / vvv !defined(_CCCL_HEADER_TEST) vvv
+#  define _CCCL_PROLOGUE_INCLUDE_CHECK()
+#endif // ^^^ !defined(_CCCL_HEADER_TEST) ^^^
+
+#ifndef _LIBCUDACXX_ABI_NAMESPACE
+#  define _LIBCUDACXX_ABI_NAMESPACE _CCCL_PP_CAT(__, _LIBCUDACXX_CUDA_ABI_VERSION)
+#endif // _LIBCUDACXX_ABI_NAMESPACE
+
+#define _CCCL_BEGIN_NAMESPACE_NOVERSION(_NS)   \
+  _CCCL_PROLOGUE_INCLUDE_CHECK() namespace _NS \
+  {
+#define _CCCL_END_NAMESPACE_NOVERSION(_NS) \
+  }                                        \
+  _CCCL_PROLOGUE_INCLUDE_CHECK()
+#define _CCCL_BEGIN_NAMESPACE(_NS)                                                \
+  _CCCL_BEGIN_NAMESPACE_NOVERSION(_NS) inline namespace _LIBCUDACXX_ABI_NAMESPACE \
+  {
+#define _CCCL_END_NAMESPACE(_NS) \
+  }                              \
+  _CCCL_END_NAMESPACE_NOVERSION(_NS)
+
+// Open a namespace for APIs that were version bumped in a minor release
+// Version bump namespace should be removed from the APIs at the next major release
+#define _CCCL_BEGIN_NAMESPACE_ABI_VER4_BUMP                                           \
+  static_assert(_LIBCUDACXX_CUDA_ABI_VERSION == 4, "Version bump should be removed"); \
+  inline namespace __version_bump_ver4_                                               \
+  {
+#define _CCCL_END_NAMESPACE_ABI_VER4_BUMP                                             \
+  static_assert(_LIBCUDACXX_CUDA_ABI_VERSION == 4, "Version bump should be removed"); \
+  }
+
+// Standard namespaces with or without versioning
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD_NOVERSION _CCCL_BEGIN_NAMESPACE_NOVERSION(cuda::std)
+#define _CCCL_END_NAMESPACE_CUDA_STD_NOVERSION   _CCCL_END_NAMESPACE_NOVERSION(cuda::std)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD           _CCCL_BEGIN_NAMESPACE(cuda::std)
+#define _CCCL_END_NAMESPACE_CUDA_STD             _CCCL_END_NAMESPACE(cuda::std)
+
+// cuda specific namespaces
+#define _CCCL_BEGIN_NAMESPACE_CUDA                     _CCCL_BEGIN_NAMESPACE(cuda)
+#define _CCCL_END_NAMESPACE_CUDA                       _CCCL_END_NAMESPACE(cuda)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_MR                  _CCCL_BEGIN_NAMESPACE(cuda::mr)
+#define _CCCL_END_NAMESPACE_CUDA_MR                    _CCCL_END_NAMESPACE(cuda::mr)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_DEVICE              _CCCL_BEGIN_NAMESPACE(cuda::device)
+#define _CCCL_END_NAMESPACE_CUDA_DEVICE                _CCCL_END_NAMESPACE(cuda::device)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_PTX                 _CCCL_BEGIN_NAMESPACE(cuda::ptx)
+#define _CCCL_END_NAMESPACE_CUDA_PTX                   _CCCL_END_NAMESPACE(cuda::ptx)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_DEVICE_EXPERIMENTAL _CCCL_BEGIN_NAMESPACE(cuda::device::experimental)
+#define _CCCL_END_NAMESPACE_CUDA_DEVICE_EXPERIMENTAL   _CCCL_END_NAMESPACE(cuda::device::experimental)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_DRIVER              _CCCL_BEGIN_NAMESPACE(cuda::__driver)
+#define _CCCL_END_NAMESPACE_CUDA_DRIVER                _CCCL_END_NAMESPACE(cuda::__driver)
+
+// Namespaces related to <simd>
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD_SIMD _CCCL_BEGIN_NAMESPACE(cuda::std::simd)
+#define _CCCL_END_NAMESPACE_CUDA_STD_SIMD   _CCCL_END_NAMESPACE(cuda::std::simd)
+
+// Namespaces related to <ranges>
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES _CCCL_BEGIN_NAMESPACE(cuda::std::ranges)
+#define _CCCL_END_NAMESPACE_CUDA_STD_RANGES   _CCCL_END_NAMESPACE(cuda::std::ranges)
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD_VIEWS  _CCCL_BEGIN_NAMESPACE(cuda::std::ranges::views)
+#define _CCCL_END_NAMESPACE_CUDA_STD_VIEWS    _CCCL_END_NAMESPACE(cuda::std::ranges::views)
+
+#define _CCCL_BEGIN_NAMESPACE_CPO(_CPO) \
+  namespace _CPO                        \
+  {
+#define _CCCL_END_NAMESPACE_CPO }
+
+// Namespaces related to chrono / filesystem
+#define _CCCL_BEGIN_NAMESPACE_FILESYSTEM     \
+  _CCCL_BEGIN_NAMESPACE_CUDA_STD_NOVERSION   \
+  inline namespace __fs                      \
+  {                                          \
+  namespace filesystem                       \
+  {                                          \
+  inline namespace _LIBCUDACXX_ABI_NAMESPACE \
+  {
+#define _CCCL_END_NAMESPACE_FILESYSTEM \
+  }                                    \
+  }                                    \
+  }                                    \
+  _CCCL_END_NAMESPACE_CUDA_STD_NOVERSION
+
+// Shorthands for different qualifiers
+// Namespaces related to execution
+#define _CCCL_BEGIN_NAMESPACE_CUDA_STD_EXECUTION _CCCL_BEGIN_NAMESPACE(cuda::std::execution)
+#define _CCCL_END_NAMESPACE_CUDA_STD_EXECUTION   _CCCL_END_NAMESPACE(cuda::std::execution)
+
+#define _CCCL_BEGIN_NAMESPACE_CUDA_EXECUTION _CCCL_BEGIN_NAMESPACE(cuda::execution)
+#define _CCCL_END_NAMESPACE_CUDA_EXECUTION   _CCCL_END_NAMESPACE(cuda::execution)
+
+#define _CCCL_BEGIN_NAMESPACE_CUDA_ARGUMENT _CCCL_BEGIN_NAMESPACE(cuda::args)
+#define _CCCL_END_NAMESPACE_CUDA_ARGUMENT   _CCCL_END_NAMESPACE(cuda::args)
+
+// Namespace to avoid name collisions with CPOs on clang-16 (see
+// https://godbolt.org/z/9TadonrdM for example). MSVC's ancient parser also gets confused with
+// __cccl_true in the main iter_move template.
+#if _CCCL_COMPILER(CLANG, <=, 16) || _CCCL_COMPILER(MSVC)
+#  define _LIBCUDACXX_BEGIN_HIDDEN_FRIEND_NAMESPACE \
+    namespace __hidden                              \
+    {
+#  define _LIBCUDACXX_END_HIDDEN_FRIEND_NAMESPACE(_CLASS) \
+    }                                                     \
+    using __hidden::_CLASS;
+#else // ^^^ _CCCL_COMPILER(CLANG, <=, 16) ^^^ / vvv _CCCL_COMPILER(CLANG, >, 16) vvv
+#  define _LIBCUDACXX_BEGIN_HIDDEN_FRIEND_NAMESPACE
+#  define _LIBCUDACXX_END_HIDDEN_FRIEND_NAMESPACE(_CLASS)
+#endif // !_CCCL_COMPILER(CLANG, >, 16)
+
+#if defined(CCCL_DISABLE_ARCH_DEPENDENT_NAMESPACE)
+#  define _CCCL_BEGIN_NAMESPACE_ARCH_DEPENDENT
+#  define _CCCL_END_NAMESPACE_ARCH_DEPENDENT
+#else // not defined(CCCL_DISABLE_ARCH_DEPENDENT_NAMESPACE)
+#  if _CCCL_CUDA_COMPILER(NVHPC)
+#    define _CCCL_BEGIN_NAMESPACE_ARCH_DEPENDENT                                                     \
+      inline namespace _CCCL_PP_CAT(_CCCL_PP_SPLICE_WITH(_, _SM, NV_TARGET_SM_INTEGER_LIST), _NVHPC) \
+      {
+#    define _CCCL_END_NAMESPACE_ARCH_DEPENDENT }
+#  else // ^^^ _CCCL_CUDA_COMPILER(NVHPC) ^^^ / vvv !_CCCL_CUDA_COMPILER(NVHPC) vvv
+#    define _CCCL_BEGIN_NAMESPACE_ARCH_DEPENDENT                        \
+      inline namespace _CCCL_PP_SPLICE_WITH(_, _SM, __CUDA_ARCH_LIST__) \
+      {
+#    define _CCCL_END_NAMESPACE_ARCH_DEPENDENT }
+#  endif // ^^^ !_CCCL_CUDA_COMPILER(NVHPC) ^^^
+#endif // not defined(CCCL_DISABLE_ARCH_DEPENDENT_NAMESPACE)
+
+// Host standard library namespaces
+#if _CCCL_HOST_STD_LIB(LIBSTDCXX)
+// We don't appy attributes on forward declarations, so we omit the _GLIBCXX_VISIBILITY(default)
+#  if _GLIBCXX_INLINE_VERSION
+#    define _CCCL_BEGIN_NAMESPACE_STD              \
+      _CCCL_PROLOGUE_INCLUDE_CHECK() namespace std \
+      {                                            \
+        inline _GLIBCXX_BEGIN_NAMESPACE_VERSION
+#    define _CCCL_END_NAMESPACE_STD  \
+      _GLIBCXX_END_NAMESPACE_VERSION \
+      }                              \
+      _CCCL_PROLOGUE_INCLUDE_CHECK()
+#  else // ^^^ _GLIBCXX_INLINE_VERSION ^^^ / vvv !_GLIBCXX_INLINE_VERSION vvv
+#    define _CCCL_BEGIN_NAMESPACE_STD              \
+      _CCCL_PROLOGUE_INCLUDE_CHECK() namespace std \
+      {
+#    define _CCCL_END_NAMESPACE_STD \
+      }                             \
+      _CCCL_PROLOGUE_INCLUDE_CHECK()
+#  endif // ^^^ !_GLIBCXX_INLINE_VERSION ^^^
+#elif _CCCL_HOST_STD_LIB(LIBCXX)
+#  define _CCCL_BEGIN_NAMESPACE_STD _CCCL_PROLOGUE_INCLUDE_CHECK() _LIBCPP_BEGIN_NAMESPACE_STD
+#  define _CCCL_END_NAMESPACE_STD   _LIBCPP_END_NAMESPACE_STD _CCCL_PROLOGUE_INCLUDE_CHECK()
+#elif _CCCL_HOST_STD_LIB(STL)
+#  define _CCCL_BEGIN_NAMESPACE_STD _CCCL_PROLOGUE_INCLUDE_CHECK() _STD_BEGIN
+#  define _CCCL_END_NAMESPACE_STD   _STD_END _CCCL_PROLOGUE_INCLUDE_CHECK()
+#else
+#  define _CCCL_BEGIN_NAMESPACE_STD              \
+    _CCCL_PROLOGUE_INCLUDE_CHECK() namespace std \
+    {
+#  define _CCCL_END_NAMESPACE_STD \
+    }                             \
+    _CCCL_PROLOGUE_INCLUDE_CHECK()
+#endif
+
+#endif // _CUDA_STD___INTERNAL_NAMESPACES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/pstl_config.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/pstl_config.h
new file mode 100644
index 00000000..1ba12811
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/pstl_config.h
@@ -0,0 +1,34 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_PSTL_CONFIG_H
+#define _CUDA_STD___INTERNAL_PSTL_CONFIG_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_HAS_BACKEND_CUDA() _CCCL_CUDA_COMPILATION() && !_CCCL_COMPILER(NVRTC)
+#define _CCCL_HAS_BACKEND_OMP()  0
+#define _CCCL_HAS_BACKEND_TBB()  0
+
+#define _CCCL_HAS_PSTL_BACKEND() (_CCCL_HAS_BACKEND_CUDA() || _CCCL_HAS_BACKEND_OMP() || _CCCL_HAS_BACKEND_TBB())
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___INTERNAL_PSTL_CONFIG_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/thread_api.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/thread_api.h
new file mode 100644
index 00000000..1731bb0d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/thread_api.h
@@ -0,0 +1,58 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_THREAD_API_H
+#define _CUDA_STD___INTERNAL_THREAD_API_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// Thread API
+#ifndef _CCCL_HAS_THREAD_API_EXTERNAL
+#  if _CCCL_COMPILER(NVRTC) || defined(__EMSCRIPTEN__)
+#    define _CCCL_HAS_THREAD_API_EXTERNAL
+#  endif
+#endif // _CCCL_HAS_THREAD_API_EXTERNAL
+
+#ifndef _CCCL_HAS_THREAD_API_CUDA
+#  if ((_CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC)) || defined(__EMSCRIPTEN__) || _CCCL_HOSTJIT())
+#    define _CCCL_HAS_THREAD_API_CUDA
+#  endif // ((_CCCL_DEVICE_COMPILATION() && !_CCCL_CUDA_COMPILER(NVHPC)) || defined(__EMSCRIPTEN__))
+#endif // _CCCL_HAS_THREAD_API_CUDA
+
+#ifndef _CCCL_HAS_THREAD_API_WIN32
+#  if _CCCL_COMPILER(MSVC) && !defined(_CCCL_HAS_THREAD_API_CUDA)
+#    define _CCCL_HAS_THREAD_API_WIN32
+#  endif // _CCCL_COMPILER(MSVC) && !defined(_CCCL_HAS_THREAD_API_CUDA)
+#endif // _CCCL_HAS_THREAD_API_WIN32
+
+#if !defined(_CCCL_HAS_THREAD_API_PTHREAD) && !defined(_CCCL_HAS_THREAD_API_WIN32) \
+  && !defined(_CCCL_HAS_THREAD_API_EXTERNAL)
+#  if defined(__GNU__) || _CCCL_OS(LINUX) || _CCCL_OS(APPLE) || _CCCL_OS(QNX) \
+    || (defined(__MINGW32__) && __has_include(<pthread.h>))
+#    define _CCCL_HAS_THREAD_API_PTHREAD
+#  elif defined(_WIN32)
+#    define _CCCL_HAS_THREAD_API_WIN32
+#  else
+#    define _CCCL_UNSUPPORTED_THREAD_API
+#  endif // _CCCL_HAS_THREAD_API
+#endif
+
+#ifndef __STDCPP_THREADS__
+#  define __STDCPP_THREADS__ 1
+#endif // __STDCPP_THREADS__
+
+#endif // _CUDA_STD___INTERNAL_THREAD_API_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/version.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/version.h
new file mode 100644
index 00000000..d7be937c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__internal/version.h
@@ -0,0 +1,52 @@
+//===---------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===---------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___INTERNAL_VERSION_H
+#define _CUDA_STD___INTERNAL_VERSION_H
+
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/version.h> // IWYU pragma: export
+
+#define _LIBCUDACXX_CUDA_API_VERSION       CCCL_VERSION
+#define _LIBCUDACXX_CUDA_API_VERSION_MAJOR CCCL_MAJOR_VERSION
+#define _LIBCUDACXX_CUDA_API_VERSION_MINOR CCCL_MINOR_VERSION
+#define _LIBCUDACXX_CUDA_API_VERSION_PATCH CCCL_PATCH_VERSION
+
+#ifndef _LIBCUDACXX_CUDA_ABI_VERSION_LATEST
+#  define _LIBCUDACXX_CUDA_ABI_VERSION_LATEST 4
+#endif
+
+#ifdef _LIBCUDACXX_CUDA_ABI_VERSION
+#  if _LIBCUDACXX_CUDA_ABI_VERSION != 4
+#    error Unsupported libcu++ ABI version requested. Only version 4 is allowed.
+#  endif
+#else
+#  define _LIBCUDACXX_CUDA_ABI_VERSION _LIBCUDACXX_CUDA_ABI_VERSION_LATEST
+#endif
+
+#if (_LIBCUDACXX_CUDA_ABI_VERSION < 4) && !defined(LIBCUDACXX_IGNORE_DEPRECATED_ABI)
+#  error "libcu++ ABIs older than version 4 are deprecated, define LIBCUDACXX_IGNORE_DEPRECATED_ABI to ignore"
+#endif
+
+#ifdef _LIBCUDACXX_PIPELINE_ASSUMED_ABI_VERSION
+#  if _LIBCUDACXX_PIPELINE_ASSUMED_ABI_VERSION != _LIBCUDACXX_CUDA_ABI_VERSION
+#    error cuda_pipeline.h has assumed a different libcu++ ABI version than provided by this library. To fix this, please include a libcu++ header before including cuda_pipeline.h, or upgrade to a version of the toolkit this version of libcu++ shipped in.
+#  endif
+#endif
+
+#endif // _CUDA_STD___INTERNAL_VERSION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/access.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/access.h
new file mode 100644
index 00000000..3059fe53
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/access.h
@@ -0,0 +1,140 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_ACCESS_H
+#define _CUDA_STD___ITERATOR_ACCESS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+namespace __begin
+{
+struct __fn
+{
+  template <class _Tp, size_t _Np>
+  _CCCL_API constexpr _Tp* operator()(_Tp (&__array)[_Np]) const noexcept
+  {
+    return __array;
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(_Cp& __c) const noexcept(noexcept(__c.begin())) -> decltype(__c.begin())
+  {
+    return __c.begin();
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(__c.begin())) -> decltype(__c.begin())
+  {
+    return __c.begin();
+  }
+};
+} // namespace __begin
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto begin = __begin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __begin_cpo = __begin::__fn;
+} // namespace __cpo
+
+namespace __end
+{
+struct __fn
+{
+  template <class _Tp, size_t _Np>
+  _CCCL_API constexpr _Tp* operator()(_Tp (&__array)[_Np]) const noexcept
+  {
+    return __array + _Np;
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(_Cp& __c) const noexcept(noexcept(__c.end())) -> decltype(__c.end())
+  {
+    return __c.end();
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(__c.end())) -> decltype(__c.end())
+  {
+    return __c.end();
+  }
+};
+} // namespace __end
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto end = __end::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __end_cpo = __end::__fn;
+} // namespace __cpo
+
+namespace __cbegin
+{
+struct __fn
+{
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(::cuda::std::__begin_cpo{}(__c)))
+    -> decltype(::cuda::std::__begin_cpo{}(__c))
+  {
+    return ::cuda::std::__begin_cpo{}(__c);
+  }
+};
+} // namespace __cbegin
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto cbegin = __cbegin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __cbegin_cpo = __cbegin::__fn;
+} // namespace __cpo
+
+namespace __cend
+{
+struct __fn
+{
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(::cuda::std::__end_cpo{}(__c)))
+    -> decltype(::cuda::std::__end_cpo{}(__c))
+  {
+    return ::cuda::std::__end_cpo{}(__c);
+  }
+};
+} // namespace __cend
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto cend = __cend::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __cend_cpo = __cend::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ACCESS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/advance.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/advance.h
new file mode 100644
index 00000000..2458dbfb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/advance.h
@@ -0,0 +1,230 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_ADVANCE_H
+#define _CUDA_STD___ITERATOR_ADVANCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/assignable.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__utility/convert_to_integral.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _InputIter,
+          class _Distance,
+          class _IntegralDistance = decltype(::cuda::std::__convert_to_integral(::cuda::std::declval<_Distance>())),
+          class                   = enable_if_t<is_integral_v<_IntegralDistance>>>
+_CCCL_API constexpr void advance(_InputIter& __i, _Distance __orig_n)
+{
+  using _Difference = typename iterator_traits<_InputIter>::difference_type;
+  _Difference __n   = static_cast<_Difference>(::cuda::std::__convert_to_integral(__orig_n));
+  if constexpr (__has_random_access_traversal<_InputIter>) // To support pointers to incomplete types
+  {
+    __i += __n;
+  }
+  else if constexpr (__has_bidirectional_traversal<_InputIter>)
+  {
+    if (__n >= 0)
+    {
+      for (; __n > 0; --__n)
+      {
+        ++__i;
+      }
+    }
+    else
+    {
+      for (; __n < 0; ++__n)
+      {
+        --__i;
+      }
+    }
+  }
+  else
+  {
+    _CCCL_ASSERT(__n >= 0, "Attempt to advance(it, n) with negative n on a non-bidirectional iterator");
+    for (; __n > 0; --__n)
+    {
+      ++__i;
+    }
+  }
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// [range.iter.op.advance]
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__advance)
+struct __fn
+{
+private:
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter_difference>
+  [[nodiscard]] _CCCL_API static constexpr auto __magnitude_geq(_Iter_difference __a, _Iter_difference __b) noexcept
+  {
+    return __a == 0 ? __b == 0 : //
+             __a > 0 ? __a >= __b
+                     : __a <= __b;
+  }
+
+public:
+  // Preconditions: If `I` does not model `bidirectional_iterator`, `n` is not negative.
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip>)
+  _CCCL_API constexpr void operator()(_Ip& __i, iter_difference_t<_Ip> __n) const
+  {
+    _CCCL_ASSERT(__n >= 0 || bidirectional_iterator<_Ip>, "If `n < 0`, then `bidirectional_iterator<I>` must be true.");
+
+    // If `I` models `random_access_iterator`, equivalent to `i += n`.
+    if constexpr (random_access_iterator<_Ip>)
+    {
+      __i += __n;
+      return;
+    }
+    else if constexpr (bidirectional_iterator<_Ip>)
+    {
+      // Otherwise, if `n` is non-negative, increments `i` by `n`.
+      while (__n > 0)
+      {
+        --__n;
+        ++__i;
+      }
+      // Otherwise, decrements `i` by `-n`.
+      while (__n < 0)
+      {
+        ++__n;
+        --__i;
+      }
+      return;
+    }
+    else
+    {
+      // Otherwise, if `n` is non-negative, increments `i` by `n`.
+      while (__n > 0)
+      {
+        --__n;
+        ++__i;
+      }
+      return;
+    }
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip> _CCCL_AND sentinel_for<_Sp, _Ip>)
+  _CCCL_API constexpr void operator()(_Ip& __i, _Sp __bound_sentinel) const
+  {
+    // If `I` and `S` model `assignable_from<I&, S>`, equivalent to `i = std::move(bound_sentinel)`.
+    if constexpr (assignable_from<_Ip&, _Sp>)
+    {
+      __i = ::cuda::std::move(__bound_sentinel);
+    }
+    // Otherwise, if `S` and `I` model `sized_sentinel_for<S, I>`,
+    // equivalent to `ranges::advance(i, bound_sentinel - i)`.
+    else if constexpr (sized_sentinel_for<_Sp, _Ip>)
+    {
+      (*this)(__i, __bound_sentinel - __i);
+    }
+    // Otherwise, while `bool(i != bound_sentinel)` is true, increments `i`.
+    else
+    {
+      while (__i != __bound_sentinel)
+      {
+        ++__i;
+      }
+    }
+  }
+
+  // Preconditions:
+  //   * If `n > 0`, [i, bound_sentinel) denotes a range.
+  //   * If `n == 0`, [i, bound_sentinel) or [bound_sentinel, i) denotes a range.
+  //   * If `n < 0`, [bound_sentinel, i) denotes a range, `I` models `bidirectional_iterator`,
+  //     and `I` and `S` model `same_as<I, S>`.
+  // Returns: `n - M`, where `M` is the difference between the ending and starting position.
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip> _CCCL_AND sentinel_for<_Sp, _Ip>)
+  _CCCL_API constexpr iter_difference_t<_Ip> operator()(_Ip& __i, iter_difference_t<_Ip> __n, _Sp __bound_sentinel) const
+  {
+    _CCCL_ASSERT((__n >= 0) || (bidirectional_iterator<_Ip> && same_as<_Ip, _Sp>),
+                 "If `n < 0`, then `bidirectional_iterator<I> && same_as<I, S>` must be true.");
+    // If `S` and `I` model `sized_sentinel_for<S, I>`:
+    if constexpr (sized_sentinel_for<_Sp, _Ip>)
+    {
+      // If |n| >= |bound_sentinel - i|, equivalent to `ranges::advance(i, bound_sentinel)`.
+      // __magnitude_geq(a, b) returns |a| >= |b|, assuming they have the same sign.
+      const auto __M = __bound_sentinel - __i;
+      if (__magnitude_geq(__n, __M))
+      {
+        (*this)(__i, __bound_sentinel);
+        return __n - __M;
+      }
+
+      // Otherwise, equivalent to `ranges::advance(i, n)`.
+      (*this)(__i, __n);
+      return 0;
+    }
+    else
+    {
+      // Otherwise, if `n` is non-negative, while `bool(i != bound_sentinel)` is true, increments `i` but at
+      // most `n` times.
+      while (__i != __bound_sentinel && __n > 0)
+      {
+        ++__i;
+        --__n;
+      }
+
+      // Otherwise, while `bool(i != bound_sentinel)` is true, decrements `i` but at most `-n` times.
+      if constexpr (bidirectional_iterator<_Ip> && same_as<_Ip, _Sp>)
+      {
+        while (__i != __bound_sentinel && __n < 0)
+        {
+          --__i;
+          ++__n;
+        }
+      }
+      return __n;
+    }
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto advance = __advance::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __advance_cpo = __advance::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ADVANCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/concepts.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/concepts.h
new file mode 100644
index 00000000..a346d1cb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/concepts.h
@@ -0,0 +1,718 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_CONCEPTS_H
+#define _CUDA_STD___ITERATOR_CONCEPTS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/arithmetic.h>
+#include <cuda/std/__concepts/assignable.h>
+#include <cuda/std/__concepts/common_reference_with.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/copyable.h>
+#include <cuda/std/__concepts/derived_from.h>
+#include <cuda/std/__concepts/equality_comparable.h>
+#include <cuda/std/__concepts/invocable.h>
+#include <cuda/std/__concepts/movable.h>
+#include <cuda/std/__concepts/predicate.h>
+#include <cuda/std/__concepts/regular.h>
+#include <cuda/std/__concepts/relation.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__concepts/semiregular.h>
+#include <cuda/std/__concepts/totally_ordered.h>
+#include <cuda/std/__functional/invoke.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iter_move.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__memory/pointer_traits.h>
+#include <cuda/std/__type_traits/add_pointer.h>
+#include <cuda/std/__type_traits/common_reference.h>
+#include <cuda/std/__type_traits/conjunction.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [iterator.concept.readable]
+template <class _In>
+concept __indirectly_readable_impl =
+  requires(const _In __i) {
+    typename iter_value_t<_In>;
+    typename iter_reference_t<_In>;
+    typename iter_rvalue_reference_t<_In>;
+    { *__i } -> same_as<iter_reference_t<_In>>;
+    { ::cuda::std::ranges::__iter_move_cpo{}(__i) } -> same_as<iter_rvalue_reference_t<_In>>;
+  } && common_reference_with<iter_reference_t<_In>&&, iter_value_t<_In>&>
+  && common_reference_with<iter_reference_t<_In>&&, iter_rvalue_reference_t<_In>&&>
+  && common_reference_with<iter_rvalue_reference_t<_In>&&, const iter_value_t<_In>&>;
+
+template <class _In>
+concept indirectly_readable = __indirectly_readable_impl<remove_cvref_t<_In>>;
+
+template <indirectly_readable _Tp>
+using iter_common_reference_t = common_reference_t<iter_reference_t<_Tp>, iter_value_t<_Tp>&>;
+
+// [iterator.concept.writable]
+template <class _Out, class _Tp>
+concept indirectly_writable = requires(_Out&& __o, _Tp&& __t) {
+  *__o                       = static_cast<_Tp &&>(__t); // not required to be equality-preserving
+  *static_cast<_Out &&>(__o) = static_cast<_Tp &&>(__t); // not required to be equality-preserving
+  const_cast<const iter_reference_t<_Out> &&>(*__o) = static_cast<_Tp &&>(__t); // not required to be
+                                                                                // equality-preserving
+  const_cast<const iter_reference_t<_Out> &&>(*static_cast<_Out &&>(__o)) =
+    static_cast<_Tp &&>(__t); // not required to be equality-preserving
+};
+
+// [iterator.concept.winc]
+template <class _Tp>
+concept __integer_like = integral<_Tp> && !same_as<_Tp, bool>;
+
+template <class _Tp>
+concept __signed_integer_like = signed_integral<_Tp>;
+
+template <class _Ip>
+concept weakly_incrementable =
+  // TODO: remove this once the clang bug is fixed (bugs.llvm.org/PR48173).
+  !same_as<_Ip, bool> && // Currently, clang does not handle bool correctly.
+  movable<_Ip> && requires(_Ip __i) {
+    typename iter_difference_t<_Ip>;
+    requires __signed_integer_like<iter_difference_t<_Ip>>;
+    { ++__i } -> same_as<_Ip&>; // not required to be equality-preserving
+    __i++; // not required to be equality-preserving
+  };
+
+// [iterator.concept.inc]
+template <class _Ip>
+concept incrementable = regular<_Ip> && weakly_incrementable<_Ip> && requires(_Ip __i) {
+  { __i++ } -> same_as<_Ip>;
+};
+
+// [iterator.concept.iterator]
+template <class _Ip>
+concept input_or_output_iterator = requires(_Ip __i) {
+  { *__i } -> __can_reference;
+} && weakly_incrementable<_Ip>;
+
+// [iterator.concept.sentinel]
+template <class _Sp, class _Ip>
+concept sentinel_for = semiregular<_Sp> && input_or_output_iterator<_Ip> && __weakly_equality_comparable_with<_Sp, _Ip>;
+
+template <class, class>
+inline constexpr bool disable_sized_sentinel_for = false;
+
+template <class _Sp, class _Ip>
+concept sized_sentinel_for =
+  sentinel_for<_Sp, _Ip> && !disable_sized_sentinel_for<remove_cv_t<_Sp>, remove_cv_t<_Ip>>
+  && requires(const _Ip& __i, const _Sp& __s) {
+       { __s - __i } -> same_as<iter_difference_t<_Ip>>;
+       { __i - __s } -> same_as<iter_difference_t<_Ip>>;
+     };
+
+// [iterator.concept.input]
+template <class _Ip>
+concept input_iterator = input_or_output_iterator<_Ip> && indirectly_readable<_Ip> && requires {
+  typename _ITER_CONCEPT<_Ip>;
+} && derived_from<_ITER_CONCEPT<_Ip>, input_iterator_tag>;
+
+// [iterator.concept.output]
+template <class _Ip, class _Tp>
+concept output_iterator =
+  input_or_output_iterator<_Ip> && indirectly_writable<_Ip, _Tp> && requires(_Ip __it, _Tp&& __t) {
+    *__it++ = static_cast<_Tp &&>(__t); // not required to be equality-preserving
+  };
+
+// [iterator.concept.forward]
+template <class _Ip>
+concept forward_iterator = input_iterator<_Ip> && derived_from<_ITER_CONCEPT<_Ip>, forward_iterator_tag>
+                        && incrementable<_Ip> && sentinel_for<_Ip, _Ip>;
+
+// [iterator.concept.bidir]
+template <class _Iter>
+concept __iter_can_decrement = requires(_Iter __iter) {
+  { --__iter } -> same_as<_Iter&>;
+  { __iter-- } -> same_as<_Iter>;
+};
+
+template <class _Iter>
+concept bidirectional_iterator =
+  forward_iterator<_Iter> && derived_from<_ITER_CONCEPT<_Iter>, bidirectional_iterator_tag>
+  && __iter_can_decrement<_Iter>;
+
+template <class _Iter>
+concept __iter_can_plus_equal = requires(_Iter __iter, const iter_difference_t<_Iter> __n) {
+  { __iter += __n } -> same_as<_Iter&>;
+};
+
+template <class _Iter>
+concept __iter_can_plus = requires(const _Iter __iter, const iter_difference_t<_Iter> __n) {
+  { __iter + __n } -> same_as<_Iter>;
+  { __n + __iter } -> same_as<_Iter>;
+};
+
+template <class _Iter>
+concept __iter_can_minus_equal = requires(_Iter __iter, const iter_difference_t<_Iter> __n) {
+  { __iter -= __n } -> same_as<_Iter&>;
+};
+
+template <class _Iter>
+concept __iter_can_minus = requires(const _Iter __iter, const iter_difference_t<_Iter> __n) {
+  { __iter - __n } -> same_as<_Iter>;
+};
+
+template <class _Iter>
+concept __iter_can_subscript = requires(const _Iter __iter, const iter_difference_t<_Iter> __n) {
+  { __iter[__n] } -> same_as<iter_reference_t<_Iter>>;
+};
+
+template <class _Iter>
+concept __random_access_operations =
+  __iter_can_plus_equal<_Iter> && __iter_can_plus<_Iter> && __iter_can_minus_equal<_Iter> && __iter_can_minus<_Iter>
+  && __iter_can_subscript<_Iter>;
+
+template <class _Iter>
+concept random_access_iterator =
+  bidirectional_iterator<_Iter> && derived_from<_ITER_CONCEPT<_Iter>, random_access_iterator_tag>
+  && totally_ordered<_Iter> && sized_sentinel_for<_Iter, _Iter> && __random_access_operations<_Iter>;
+
+template <class _Ip>
+concept contiguous_iterator =
+  random_access_iterator<_Ip> && derived_from<_ITER_CONCEPT<_Ip>, contiguous_iterator_tag>
+  && is_lvalue_reference_v<iter_reference_t<_Ip>> && same_as<iter_value_t<_Ip>, remove_cvref_t<iter_reference_t<_Ip>>>
+  && requires(const _Ip& __i) {
+       { ::cuda::std::to_address(__i) } -> same_as<add_pointer_t<iter_reference_t<_Ip>>>;
+     };
+
+template <class _Ip>
+concept __has_arrow = input_iterator<_Ip> && (is_pointer_v<_Ip> || requires(_Ip __i) { __i.operator->(); });
+
+template <class _Ip>
+concept __has_const_arrow = (is_pointer_v<_Ip> || requires(const _Ip __i) { __i.operator->(); });
+
+// [indirectcallable.indirectinvocable]
+template <class _Fp, class _It>
+concept indirectly_unary_invocable =
+  indirectly_readable<_It> && copy_constructible<_Fp> && invocable<_Fp&, iter_value_t<_It>&>
+  && invocable<_Fp&, iter_reference_t<_It>> && invocable<_Fp&, iter_common_reference_t<_It>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It>&>, invoke_result_t<_Fp&, iter_reference_t<_It>>>;
+
+template <class _Fp, class _It>
+concept indirectly_regular_unary_invocable =
+  indirectly_readable<_It> && copy_constructible<_Fp> && regular_invocable<_Fp&, iter_value_t<_It>&>
+  && regular_invocable<_Fp&, iter_reference_t<_It>> && regular_invocable<_Fp&, iter_common_reference_t<_It>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It>&>, invoke_result_t<_Fp&, iter_reference_t<_It>>>;
+
+template <class _Fp, class _It>
+concept indirect_unary_predicate =
+  indirectly_readable<_It> && copy_constructible<_Fp> && predicate<_Fp&, iter_value_t<_It>&>
+  && predicate<_Fp&, iter_reference_t<_It>> && predicate<_Fp&, iter_common_reference_t<_It>>;
+
+template <class _Fp, class _It1, class _It2>
+concept indirect_binary_predicate =
+  indirectly_readable<_It1> && indirectly_readable<_It2> && copy_constructible<_Fp>
+  && predicate<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>
+  && predicate<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>
+  && predicate<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>
+  && predicate<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>
+  && predicate<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>;
+
+template <class _Fp, class _It1, class _It2 = _It1>
+concept indirect_equivalence_relation =
+  indirectly_readable<_It1> && indirectly_readable<_It2> && copy_constructible<_Fp>
+  && equivalence_relation<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>
+  && equivalence_relation<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>
+  && equivalence_relation<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>
+  && equivalence_relation<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>
+  && equivalence_relation<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>;
+
+template <class _Fp, class _It1, class _It2 = _It1>
+concept indirect_strict_weak_order =
+  indirectly_readable<_It1> && indirectly_readable<_It2> && copy_constructible<_Fp>
+  && strict_weak_order<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>
+  && strict_weak_order<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>
+  && strict_weak_order<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>
+  && strict_weak_order<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>
+  && strict_weak_order<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>;
+
+template <class _Fp, class... _Its>
+  requires(indirectly_readable<_Its> && ...) && invocable<_Fp, iter_reference_t<_Its>...>
+using indirect_result_t = invoke_result_t<_Fp, iter_reference_t<_Its>...>;
+
+template <class _In, class _Out>
+concept indirectly_movable = indirectly_readable<_In> && indirectly_writable<_Out, iter_rvalue_reference_t<_In>>;
+
+template <class _In, class _Out>
+concept indirectly_movable_storable =
+  indirectly_movable<_In, _Out> && indirectly_writable<_Out, iter_value_t<_In>> && movable<iter_value_t<_In>>
+  && constructible_from<iter_value_t<_In>, iter_rvalue_reference_t<_In>>
+  && assignable_from<iter_value_t<_In>&, iter_rvalue_reference_t<_In>>;
+
+template <class _In, class _Out>
+concept indirectly_copyable = indirectly_readable<_In> && indirectly_writable<_Out, iter_reference_t<_In>>;
+
+template <class _In, class _Out>
+concept indirectly_copyable_storable =
+  indirectly_copyable<_In, _Out> && indirectly_writable<_Out, iter_value_t<_In>&>
+  && indirectly_writable<_Out, const iter_value_t<_In>&> && indirectly_writable<_Out, iter_value_t<_In>&&>
+  && indirectly_writable<_Out, const iter_value_t<_In>&&> && copyable<iter_value_t<_In>>
+  && constructible_from<iter_value_t<_In>, iter_reference_t<_In>>
+  && assignable_from<iter_value_t<_In>&, iter_reference_t<_In>>;
+
+// Note: indirectly_swappable is located in iter_swap.h to prevent a dependency cycle
+// (both iter_swap and indirectly_swappable require indirectly_readable).
+
+// Extension of indirectly_unary_invocable to binary operators
+template <class _Fp, class _It1, class _It2>
+concept __indirectly_binary_invocable =
+  indirectly_readable<_It1> && indirectly_readable<_It2> && copy_constructible<_Fp>
+  && invocable<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>
+  && invocable<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>
+  && invocable<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>
+  && invocable<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>
+  && invocable<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>>;
+
+// Extension of indirectly_regular_unary_invocable to binary operators
+template <class _Fp, class _It1, class _It2>
+concept __indirectly_regular_binary_invocable =
+  indirectly_readable<_It1> && indirectly_readable<_It2> && copy_constructible<_Fp>
+  && regular_invocable<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>
+  && regular_invocable<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>
+  && regular_invocable<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>
+  && regular_invocable<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>
+  && regular_invocable<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>>
+  && common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                           invoke_result_t<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>>;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [iterator.concept.readable]
+template <class _In>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_readable_impl_,
+  requires(const _In __i)(
+    typename(iter_value_t<_In>),
+    typename(iter_reference_t<_In>),
+    typename(iter_rvalue_reference_t<_In>),
+    requires(same_as<iter_reference_t<_In>, decltype(*__i)>),
+    requires(same_as<iter_rvalue_reference_t<_In>, decltype(::cuda::std::ranges::__iter_move_cpo{}(__i))>),
+    requires(common_reference_with<iter_reference_t<_In>&&, iter_value_t<_In>&>),
+    requires(common_reference_with<iter_reference_t<_In>&&, iter_rvalue_reference_t<_In>&&>),
+    requires(common_reference_with<iter_rvalue_reference_t<_In>&&, const iter_value_t<_In>&>)));
+
+template <class _In>
+_CCCL_CONCEPT indirectly_readable = _CCCL_FRAGMENT(__indirectly_readable_impl_, remove_cvref_t<_In>);
+
+template <class _Tp>
+using iter_common_reference_t =
+  enable_if_t<indirectly_readable<_Tp>, common_reference_t<iter_reference_t<_Tp>, iter_value_t<_Tp>&>>;
+
+// [iterator.concept.writable]
+template <class _Out, class _Tp>
+_CCCL_CONCEPT indirectly_writable = _CCCL_REQUIRES_EXPR((_Out, _Tp), _Out&& __o, _Tp&& __t)(
+  (*__o = static_cast<_Tp&&>(__t)),
+  (*static_cast<_Out&&>(__o) = static_cast<_Tp&&>(__t)),
+  (const_cast<const iter_reference_t<_Out>&&>(*__o) = static_cast<_Tp&&>(__t)),
+  (const_cast<const iter_reference_t<_Out>&&>(*static_cast<_Out&&>(__o)) = static_cast<_Tp&&>(__t)));
+
+// [iterator.concept.winc]
+template <class _Tp>
+_CCCL_CONCEPT __integer_like = integral<_Tp> && !same_as<_Tp, bool>;
+
+template <class _Tp>
+_CCCL_CONCEPT __signed_integer_like = signed_integral<_Tp>;
+
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __weakly_incrementable_,
+  requires(_Ip __i)(typename(iter_difference_t<_Ip>),
+                    requires(!same_as<_Ip, bool>),
+                    requires(movable<_Ip>),
+                    requires(__signed_integer_like<iter_difference_t<_Ip>>),
+                    requires(same_as<_Ip&, decltype(++__i)>),
+                    (__i++)));
+
+template <class _Ip>
+_CCCL_CONCEPT weakly_incrementable = _CCCL_FRAGMENT(__weakly_incrementable_, _Ip);
+
+// [iterator.concept.inc]
+template <class _Ip>
+_CCCL_CONCEPT incrementable = _CCCL_REQUIRES_EXPR((_Ip), _Ip __i)(
+  requires(regular<_Ip>),
+  requires(weakly_incrementable<_Ip>),
+  // Requirement is unevaluated, not sure why clang-tidy complains
+  // NOLINTNEXTLINE(bugprone-pointer-arithmetic-on-polymorphic-object)
+  requires(same_as<_Ip, decltype(__i++)>));
+
+// [iterator.concept.iterator]
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __input_or_output_iterator_,
+  requires(_Ip __i)(requires(weakly_incrementable<_Ip>), requires(__can_reference<decltype(*__i)>)));
+
+template <class _Ip>
+_CCCL_CONCEPT input_or_output_iterator = _CCCL_FRAGMENT(__input_or_output_iterator_, _Ip);
+
+// [iterator.concept.sentinel]
+template <class _Sp, class _Ip>
+_CCCL_CONCEPT_FRAGMENT(__sentinel_for_,
+                       requires()(requires(semiregular<_Sp>),
+                                  requires(input_or_output_iterator<_Ip>),
+                                  requires(__weakly_equality_comparable_with<_Sp, _Ip>)));
+
+template <class _Sp, class _Ip>
+_CCCL_CONCEPT sentinel_for = _CCCL_FRAGMENT(__sentinel_for_, _Sp, _Ip);
+
+template <class, class>
+inline constexpr bool disable_sized_sentinel_for = false;
+
+template <class _Sp, class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __sized_sentinel_for_,
+  requires(const _Ip& __i, const _Sp& __s)(
+    requires(sentinel_for<_Sp, _Ip>),
+    requires(!disable_sized_sentinel_for<remove_cv_t<_Sp>, remove_cv_t<_Ip>>),
+    requires(same_as<iter_difference_t<_Ip>, decltype(__s - __i)>),
+    requires(same_as<iter_difference_t<_Ip>, decltype(__i - __s)>)));
+
+template <class _Sp, class _Ip>
+_CCCL_CONCEPT sized_sentinel_for = _CCCL_FRAGMENT(__sized_sentinel_for_, _Sp, _Ip);
+
+// [iterator.concept.input]
+// NOTE: The ordering here is load bearing. MSVC has issues with finding iterator_traits
+//       We can work around this by checking other constraints first
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __input_iterator_,
+  requires()(requires(input_or_output_iterator<_Ip>),
+             requires(indirectly_readable<_Ip>),
+             typename(_ITER_CONCEPT<_Ip>),
+             requires(derived_from<_ITER_CONCEPT<_Ip>, input_iterator_tag>)));
+
+template <class _Ip>
+_CCCL_CONCEPT input_iterator = _CCCL_FRAGMENT(__input_iterator_, _Ip);
+
+// [iterator.concept.output]
+template <class _Ip, class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__output_iterator_,
+                       requires(_Ip __it, _Tp&& __t)(requires(input_or_output_iterator<_Ip>),
+                                                     requires(indirectly_writable<_Ip, _Tp>),
+                                                     (*__it++ = static_cast<_Tp&&>(__t))));
+
+template <class _Ip, class _Tp>
+_CCCL_CONCEPT output_iterator = _CCCL_FRAGMENT(__output_iterator_, _Ip, _Tp);
+
+// [iterator.concept.forward]
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __forward_iterator_,
+  requires()(requires(input_iterator<_Ip>),
+             requires(derived_from<_ITER_CONCEPT<_Ip>, forward_iterator_tag>),
+             requires(incrementable<_Ip>),
+             requires(sentinel_for<_Ip, _Ip>)));
+
+template <class _Ip>
+_CCCL_CONCEPT forward_iterator = _CCCL_FRAGMENT(__forward_iterator_, _Ip);
+
+// [iterator.concept.bidir]
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_decrement =
+  _CCCL_REQUIRES_EXPR((_Iter), _Iter __iter)(_Same_as(_Iter&)(--__iter), _Same_as(_Iter) __iter--);
+
+template <class _Iter>
+_CCCL_CONCEPT bidirectional_iterator = _CCCL_REQUIRES_EXPR((_Iter))(
+  requires(forward_iterator<_Iter>),
+  requires(derived_from<_ITER_CONCEPT<_Iter>, bidirectional_iterator_tag>),
+  requires(__iter_can_decrement<_Iter>));
+
+// [iterator.concept.random.access]
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_plus_equal =
+  _CCCL_REQUIRES_EXPR((_Iter), _Iter __iter, const iter_difference_t<_Iter> __n)(_Same_as(_Iter&) __iter += __n);
+
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_plus = _CCCL_REQUIRES_EXPR((_Iter), const _Iter __iter, const iter_difference_t<_Iter> __n)(
+  _Same_as(_Iter) __iter + __n, _Same_as(_Iter) __n + __iter);
+
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_minus_equal =
+  _CCCL_REQUIRES_EXPR((_Iter), _Iter __iter, const iter_difference_t<_Iter> __n)(_Same_as(_Iter&) __iter -= __n);
+
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_minus =
+  _CCCL_REQUIRES_EXPR((_Iter), const _Iter __iter, const iter_difference_t<_Iter> __n)(_Same_as(_Iter) __iter - __n);
+
+template <class _Iter>
+_CCCL_CONCEPT __iter_can_subscript = _CCCL_REQUIRES_EXPR(
+  (_Iter), const _Iter __iter, const iter_difference_t<_Iter> __n)(_Same_as(iter_reference_t<_Iter>) __iter[__n]);
+
+template <class _Iter>
+_CCCL_CONCEPT __random_access_iterator_operations = _CCCL_REQUIRES_EXPR((_Iter))(
+  requires(__iter_can_plus_equal<_Iter>),
+  requires(__iter_can_plus<_Iter>),
+  requires(__iter_can_minus_equal<_Iter>),
+  requires(__iter_can_minus<_Iter>),
+  requires(__iter_can_subscript<_Iter>));
+
+template <class _Iter>
+_CCCL_CONCEPT random_access_iterator = _CCCL_REQUIRES_EXPR((_Iter))(
+  requires(bidirectional_iterator<_Iter>),
+  requires(derived_from<_ITER_CONCEPT<_Iter>, random_access_iterator_tag>),
+  requires(totally_ordered<_Iter>),
+  requires(sized_sentinel_for<_Iter, _Iter>),
+  requires(__random_access_iterator_operations<_Iter>));
+
+// [iterator.concept.contiguous]
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(
+  __contiguous_iterator_,
+  requires(const _Ip& __i)(
+    requires(random_access_iterator<_Ip>),
+    requires(derived_from<_ITER_CONCEPT<_Ip>, contiguous_iterator_tag>),
+    requires(is_lvalue_reference_v<iter_reference_t<_Ip>>),
+    requires(same_as<iter_value_t<_Ip>, remove_cvref_t<iter_reference_t<_Ip>>>),
+    requires(same_as<add_pointer_t<iter_reference_t<_Ip>>, decltype(::cuda::std::to_address(__i))>)));
+
+template <class _Ip>
+_CCCL_CONCEPT contiguous_iterator = _CCCL_FRAGMENT(__contiguous_iterator_, _Ip);
+
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(__has_arrow_, requires(_Ip __i)((__i.operator->())));
+
+template <class _Ip>
+_CCCL_CONCEPT __has_arrow = input_iterator<_Ip> && (is_pointer_v<_Ip> || _CCCL_FRAGMENT(__has_arrow_, _Ip));
+
+template <class _Ip>
+_CCCL_CONCEPT_FRAGMENT(__has_const_arrow_, requires(const _Ip __i)((__i.operator->())));
+
+template <class _Ip>
+_CCCL_CONCEPT __has_const_arrow = (is_pointer_v<_Ip> || _CCCL_FRAGMENT(__has_const_arrow_, _Ip));
+
+// [indirectcallable.indirectinvocable]
+template <class _Fp, class _It>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_unary_invocable,
+  requires()(
+    requires(indirectly_readable<_It>),
+    requires(copy_constructible<_Fp>),
+    requires(invocable<_Fp&, iter_value_t<_It>&>),
+    requires(invocable<_Fp&, iter_reference_t<_It>>),
+    requires(invocable<_Fp&, iter_common_reference_t<_It>>),
+    requires(
+      common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It>&>, invoke_result_t<_Fp&, iter_reference_t<_It>>>)));
+
+template <class _Fp, class _It>
+_CCCL_CONCEPT indirectly_unary_invocable = _CCCL_FRAGMENT(__indirectly_unary_invocable, _Fp, _It);
+
+template <class _Fp, class _It>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_regular_unary_invocable_,
+  requires()(
+    requires(indirectly_readable<_It>),
+    requires(copy_constructible<_Fp>),
+    requires(regular_invocable<_Fp&, iter_value_t<_It>&>),
+    requires(regular_invocable<_Fp&, iter_reference_t<_It>>),
+    requires(regular_invocable<_Fp&, iter_common_reference_t<_It>>),
+    requires(
+      common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It>&>, invoke_result_t<_Fp&, iter_reference_t<_It>>>)));
+
+template <class _Fp, class _It>
+_CCCL_CONCEPT indirectly_regular_unary_invocable = _CCCL_FRAGMENT(__indirectly_regular_unary_invocable_, _Fp, _It);
+
+template <class _Fp, class _It>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirect_unary_predicate_,
+  requires()(requires(indirectly_readable<_It>),
+             requires(copy_constructible<_Fp>),
+             requires(predicate<_Fp&, iter_value_t<_It>&>),
+             requires(predicate<_Fp&, iter_reference_t<_It>>),
+             requires(predicate<_Fp&, iter_common_reference_t<_It>>)));
+
+template <class _Fp, class _It>
+_CCCL_CONCEPT indirect_unary_predicate = _CCCL_FRAGMENT(__indirect_unary_predicate_, _Fp, _It);
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirect_binary_predicate_,
+  requires()(requires(indirectly_readable<_It1>),
+             requires(indirectly_readable<_It2>),
+             requires(copy_constructible<_Fp>),
+             requires(predicate<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>),
+             requires(predicate<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>),
+             requires(predicate<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>),
+             requires(predicate<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>),
+             requires(predicate<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>)));
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT indirect_binary_predicate = _CCCL_FRAGMENT(__indirect_binary_predicate_, _Fp, _It1, _It2);
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirect_equivalence_relation_,
+  requires()(requires(indirectly_readable<_It1>),
+             requires(indirectly_readable<_It2>),
+             requires(copy_constructible<_Fp>),
+             requires(equivalence_relation<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>),
+             requires(equivalence_relation<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>),
+             requires(equivalence_relation<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>),
+             requires(equivalence_relation<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>),
+             requires(equivalence_relation<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>)));
+
+template <class _Fp, class _It1, class _It2 = _It1>
+_CCCL_CONCEPT indirect_equivalence_relation = _CCCL_FRAGMENT(__indirect_equivalence_relation_, _Fp, _It1, _It2);
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirect_strict_weak_order_,
+  requires()(requires(indirectly_readable<_It1>),
+             requires(indirectly_readable<_It2>),
+             requires(copy_constructible<_Fp>),
+             requires(strict_weak_order<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>),
+             requires(strict_weak_order<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>),
+             requires(strict_weak_order<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>),
+             requires(strict_weak_order<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>),
+             requires(strict_weak_order<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>)));
+
+template <class _Fp, class _It1, class _It2 = _It1>
+_CCCL_CONCEPT indirect_strict_weak_order = _CCCL_FRAGMENT(__indirect_strict_weak_order_, _Fp, _It1, _It2);
+
+template <class _Fp, class... _Its>
+using indirect_result_t = enable_if_t<(indirectly_readable<_Its> && ...) && invocable<_Fp, iter_reference_t<_Its>...>,
+                                      invoke_result_t<_Fp, iter_reference_t<_Its>...>>;
+
+template <class _In, class _Out>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_movable_,
+  requires()(requires(indirectly_readable<_In>), requires(indirectly_writable<_Out, iter_rvalue_reference_t<_In>>)));
+
+template <class _In, class _Out>
+_CCCL_CONCEPT indirectly_movable = _CCCL_FRAGMENT(__indirectly_movable_, _In, _Out);
+
+template <class _In, class _Out>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_movable_storable_,
+  requires()(requires(indirectly_movable<_In, _Out>),
+             requires(indirectly_writable<_Out, iter_value_t<_In>>),
+             requires(movable<iter_value_t<_In>>),
+             requires(constructible_from<iter_value_t<_In>, iter_rvalue_reference_t<_In>>),
+             requires(assignable_from<iter_value_t<_In>&, iter_rvalue_reference_t<_In>>)));
+
+template <class _In, class _Out>
+_CCCL_CONCEPT indirectly_movable_storable = _CCCL_FRAGMENT(__indirectly_movable_storable_, _In, _Out);
+
+template <class _In, class _Out>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_copyable_,
+  requires()(requires(indirectly_readable<_In>), requires(indirectly_writable<_Out, iter_reference_t<_In>>)));
+
+template <class _In, class _Out>
+_CCCL_CONCEPT indirectly_copyable = _CCCL_FRAGMENT(__indirectly_copyable_, _In, _Out);
+
+template <class _In, class _Out>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_copyable_storable_,
+  requires()(requires(indirectly_copyable<_In, _Out>),
+             requires(indirectly_writable<_Out, iter_value_t<_In>&>),
+             requires(indirectly_writable<_Out, const iter_value_t<_In>&>),
+             requires(indirectly_writable<_Out, iter_value_t<_In>&&>),
+             requires(indirectly_writable<_Out, const iter_value_t<_In>&&>),
+             requires(copyable<iter_value_t<_In>>),
+             requires(constructible_from<iter_value_t<_In>, iter_reference_t<_In>>),
+             requires(assignable_from<iter_value_t<_In>&, iter_reference_t<_In>>)));
+
+template <class _In, class _Out>
+_CCCL_CONCEPT indirectly_copyable_storable = _CCCL_FRAGMENT(__indirectly_copyable_storable_, _In, _Out);
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT __indirectly_binary_invocable = _CCCL_REQUIRES_EXPR((_Fp, _It1, _It2))(
+  requires(indirectly_readable<_It1>),
+  requires(indirectly_readable<_It2>),
+  requires(copy_constructible<_Fp>),
+  requires(invocable<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>),
+  requires(invocable<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>),
+  requires(invocable<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>),
+  requires(invocable<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>),
+  requires(invocable<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>>));
+
+template <class _Fp, class _It1, class _It2>
+_CCCL_CONCEPT __indirectly_regular_binary_invocable = _CCCL_REQUIRES_EXPR((_Fp, _It1, _It2))(
+  requires(indirectly_readable<_It1>),
+  requires(indirectly_readable<_It2>),
+  requires(copy_constructible<_Fp>),
+  requires(regular_invocable<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>),
+  requires(regular_invocable<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>),
+  requires(regular_invocable<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>),
+  requires(regular_invocable<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>),
+  requires(regular_invocable<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_reference_t<_It2>>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_value_t<_It2>&>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_reference_t<_It1>, iter_reference_t<_It2>>>),
+  requires(common_reference_with<invoke_result_t<_Fp&, iter_value_t<_It1>&, iter_value_t<_It2>&>,
+                                 invoke_result_t<_Fp&, iter_common_reference_t<_It1>, iter_common_reference_t<_It2>>>));
+
+template <class _Ip, class = void>
+inline constexpr bool __has_iter_category = false;
+
+template <class _Ip>
+inline constexpr bool __has_iter_category<_Ip, void_t<typename _Ip::iterator_category>> = true;
+
+template <class _Ip, class = void>
+inline constexpr bool __has_iter_concept = false;
+
+template <class _Ip>
+inline constexpr bool __has_iter_concept<_Ip, void_t<typename _Ip::iterator_concept>> = true;
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_CONCEPTS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/data.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/data.h
new file mode 100644
index 00000000..a928fbfe
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/data.h
@@ -0,0 +1,61 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_DATA_H
+#define _CUDA_STD___ITERATOR_DATA_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Cont>
+[[nodiscard]] _CCCL_API constexpr auto data(_Cont& __c) noexcept(noexcept(__c.data())) -> decltype(__c.data())
+{
+  return __c.data();
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Cont>
+[[nodiscard]] _CCCL_API constexpr auto data(const _Cont& __c) noexcept(noexcept(__c.data())) -> decltype(__c.data())
+{
+  return __c.data();
+}
+
+template <class _Tp, size_t _Sz>
+_CCCL_API constexpr _Tp* data(_Tp (&__array)[_Sz]) noexcept
+{
+  return __array;
+}
+
+template <class _Ep>
+_CCCL_API constexpr const _Ep* data(initializer_list<_Ep> __il) noexcept
+{
+  return __il.begin();
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_DATA_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/distance.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/distance.h
new file mode 100644
index 00000000..73ca5dfb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/distance.h
@@ -0,0 +1,130 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_DISTANCE_H
+#define _CUDA_STD___ITERATOR_DISTANCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__ranges/access.h>
+#include <cuda/std/__ranges/concepts.h>
+#include <cuda/std/__ranges/size.h>
+#include <cuda/std/__type_traits/decay.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _InputIter>
+[[nodiscard]] _CCCL_API constexpr typename iterator_traits<_InputIter>::difference_type
+distance(_InputIter __first, _InputIter __last)
+{
+  // Must clone branches because sized_sentinel_for may require the type to be complete
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (__has_random_access_traversal<_InputIter>) // To support pointers to incomplete types
+  {
+    return __last - __first;
+  }
+  else if constexpr (sized_sentinel_for<_InputIter, _InputIter>)
+  {
+    return __last - __first;
+  }
+  else
+  {
+    typename iterator_traits<_InputIter>::difference_type __r(0);
+    for (; __first != __last; ++__first)
+    {
+      ++__r;
+    }
+    return __r;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// [range.iter.op.distance]
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__distance)
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES((sentinel_for<_Sp, _Ip> && !sized_sentinel_for<_Sp, _Ip>) )
+  [[nodiscard]] _CCCL_API constexpr iter_difference_t<_Ip> operator()(_Ip __first, _Sp __last) const
+  {
+    iter_difference_t<_Ip> __n = 0;
+    while (__first != __last)
+    {
+      ++__first;
+      ++__n;
+    }
+    return __n;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES((sized_sentinel_for<_Sp, decay_t<_Ip>>) )
+  [[nodiscard]] _CCCL_API constexpr iter_difference_t<_Ip> operator()(_Ip&& __first, _Sp __last) const
+  {
+    if constexpr (sized_sentinel_for<_Sp, remove_cvref_t<_Ip>>)
+    {
+      return __last - __first;
+    }
+    else
+    {
+      return __last - decay_t<_Ip>(__first);
+    }
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Rp)
+  _CCCL_REQUIRES((range<_Rp>) )
+  [[nodiscard]] _CCCL_API constexpr range_difference_t<_Rp> operator()(_Rp&& __r) const
+  {
+    if constexpr (sized_range<_Rp>)
+    {
+      return static_cast<range_difference_t<_Rp>>(::cuda::std::ranges::__size_cpo{}(__r));
+    }
+    else
+    {
+      return operator()(::cuda::std::ranges::__begin_cpo{}(__r), ::cuda::std::ranges::__end_cpo{}(__r));
+    }
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto distance = __distance::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __distance_cpo = __distance::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_DISTANCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/empty.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/empty.h
new file mode 100644
index 00000000..dc6ddd10
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/empty.h
@@ -0,0 +1,53 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_EMPTY_H
+#define _CUDA_STD___ITERATOR_EMPTY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Cont>
+[[nodiscard]] _CCCL_API constexpr auto empty(const _Cont& __c) noexcept(noexcept(__c.empty())) -> decltype(__c.empty())
+{
+  return __c.empty();
+}
+
+template <class _Tp, size_t _Sz>
+[[nodiscard]] _CCCL_API constexpr bool empty(const _Tp (&)[_Sz]) noexcept
+{
+  return false;
+}
+
+template <class _Ep>
+[[nodiscard]] _CCCL_API constexpr bool empty(initializer_list<_Ep> __il) noexcept
+{
+  return __il.size() == 0;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_EMPTY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/incrementable_traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/incrementable_traits.h
new file mode 100644
index 00000000..3fda10ea
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/incrementable_traits.h
@@ -0,0 +1,143 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_INCREMENTABLE_TRAITS_H
+#define _CUDA_STD___ITERATOR_INCREMENTABLE_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/arithmetic.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__fwd/iterator.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_const.h>
+#include <cuda/std/__type_traits/is_object.h>
+#include <cuda/std/__type_traits/is_pointer.h>
+#include <cuda/std/__type_traits/is_primary_template.h>
+#include <cuda/std/__type_traits/make_signed.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_difference_type = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::difference_type));
+
+template <class _Tp, class = void, class = void>
+inline constexpr bool __has_integral_minus_impl = false;
+
+// In C++17 we get issues trying to bind void* to a const& so special case it here
+template <class _Tp>
+inline constexpr bool
+  __has_integral_minus_impl<_Tp,
+                            enable_if_t<!same_as<_Tp, void*>>,
+                            void_t<decltype(::cuda::std::declval<const _Tp&>() - ::cuda::std::declval<const _Tp&>())>> =
+    integral<decltype(::cuda::std::declval<const _Tp&>() - ::cuda::std::declval<const _Tp&>())>;
+
+template <class _Tp>
+_CCCL_CONCEPT __has_integral_minus = _CCCL_REQUIRES_EXPR((_Tp))(requires(__has_integral_minus_impl<_Tp>));
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [incrementable.traits]
+template <class>
+struct incrementable_traits
+{};
+
+template <class _Tp>
+  requires is_object_v<_Tp>
+struct incrementable_traits<_Tp*>
+{
+  using difference_type = ptrdiff_t;
+};
+
+template <class _Ip>
+struct incrementable_traits<const _Ip> : incrementable_traits<_Ip>
+{};
+
+template <__has_member_difference_type _Tp>
+struct incrementable_traits<_Tp>
+{
+  using difference_type = typename _Tp::difference_type;
+};
+
+template <__has_integral_minus _Tp>
+  requires(!__has_member_difference_type<_Tp>)
+struct incrementable_traits<_Tp>
+{
+  using difference_type = make_signed_t<decltype(::cuda::std::declval<_Tp>() - ::cuda::std::declval<_Tp>())>;
+};
+
+// Let `RI` be `remove_cvref_t<I>`. The type `iter_difference_t<I>` denotes
+// `incrementable_traits<RI>::difference_type` if `iterator_traits<RI>` names a specialization
+// generated from the primary template, and `iterator_traits<RI>::difference_type` otherwise.
+template <class _Ip>
+using iter_difference_t =
+  typename __select_traits<remove_cvref_t<_Ip>, incrementable_traits<remove_cvref_t<_Ip>>>::difference_type;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [incrementable.traits]
+template <class, class = void>
+struct incrementable_traits
+{};
+
+template <class _Tp>
+struct incrementable_traits<_Tp*, enable_if_t<is_object_v<_Tp>>>
+{
+  using difference_type = ptrdiff_t;
+};
+
+template <class _Ip>
+struct incrementable_traits<const _Ip> : incrementable_traits<_Ip>
+{};
+
+template <class _Tp>
+struct incrementable_traits<_Tp, enable_if_t<!is_pointer_v<_Tp> && !is_const_v<_Tp> && __has_member_difference_type<_Tp>>>
+{
+  using difference_type = typename _Tp::difference_type;
+};
+
+template <class _Tp>
+struct incrementable_traits<
+  _Tp,
+  enable_if_t<!is_pointer_v<_Tp> && !is_const_v<_Tp> && !__has_member_difference_type<_Tp> && __has_integral_minus<_Tp>>>
+{
+  using difference_type = make_signed_t<decltype(::cuda::std::declval<_Tp>() - ::cuda::std::declval<_Tp>())>;
+};
+
+// Let `RI` be `remove_cvref_t<I>`. The type `iter_difference_t<I>` denotes
+// `incrementable_traits<RI>::difference_type` if `iterator_traits<RI>` names a specialization
+// generated from the primary template, and `iterator_traits<RI>::difference_type` otherwise.
+template <class _Ip>
+using iter_difference_t =
+  typename __select_traits<remove_cvref_t<_Ip>, incrementable_traits<remove_cvref_t<_Ip>>>::difference_type;
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_INCREMENTABLE_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_move.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_move.h
new file mode 100644
index 00000000..110d49ba
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_move.h
@@ -0,0 +1,165 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_ITER_MOVE_H
+#define _CUDA_STD___ITERATOR_ITER_MOVE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_CLANG("-Wvoid-ptr-dereference")
+
+// [iterator.cust.move]
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__iter_move)
+
+_CCCL_API void iter_move();
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __unqualified_iter_move =
+  __class_or_enum<remove_cvref_t<_Tp>> && requires(_Tp&& __t) { iter_move(::cuda::std::forward<_Tp>(__t)); };
+
+template <class _Tp>
+concept __move_deref = !__unqualified_iter_move<_Tp> && requires(_Tp&& __t) {
+  *__t;
+  requires is_lvalue_reference_v<decltype(*__t)>;
+};
+
+template <class _Tp>
+concept __just_deref = !__unqualified_iter_move<_Tp> && !__move_deref<_Tp> && requires(_Tp&& __t) {
+  *__t;
+  requires(!is_lvalue_reference_v<decltype(*__t)>);
+};
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__unqualified_iter_move_,
+                       requires(_Tp&& __t)(requires(__class_or_enum<remove_cvref_t<_Tp>>),
+                                           ((void) iter_move(::cuda::std::forward<_Tp>(__t)))));
+
+template <class _Tp>
+_CCCL_CONCEPT __unqualified_iter_move = _CCCL_FRAGMENT(__unqualified_iter_move_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __move_deref_,
+  requires(_Tp&& __t)(requires(!__unqualified_iter_move<_Tp>), requires(is_lvalue_reference_v<decltype(*__t)>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __move_deref = _CCCL_FRAGMENT(__move_deref_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__just_deref_,
+                       requires(_Tp&& __t)(requires(!__unqualified_iter_move<_Tp>),
+                                           requires(!__move_deref<_Tp>),
+                                           requires(!is_lvalue_reference_v<decltype(*__t)>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __just_deref = _CCCL_FRAGMENT(__just_deref_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+// [iterator.cust.move]
+
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(__unqualified_iter_move<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr decltype(auto) operator()(_Ip&& __i) const
+    noexcept(noexcept(iter_move(::cuda::std::forward<_Ip>(__i))))
+  {
+    return iter_move(::cuda::std::forward<_Ip>(__i));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(__move_deref<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Ip&& __i) const
+    noexcept(noexcept(::cuda::std::move(*::cuda::std::forward<_Ip>(__i))))
+      -> decltype(::cuda::std::move(*::cuda::std::forward<_Ip>(__i)))
+  {
+    return ::cuda::std::move(*::cuda::std::forward<_Ip>(__i));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(__just_deref<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Ip&& __i) const noexcept(noexcept(*::cuda::std::forward<_Ip>(__i)))
+    -> decltype(*::cuda::std::forward<_Ip>(__i))
+  {
+    return *::cuda::std::forward<_Ip>(__i);
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto iter_move = __iter_move::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __iter_move_cpo = __iter_move::__fn;
+} // namespace __cpo
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_CONCEPTS()
+template <__dereferenceable _Tp>
+  requires requires(_Tp& __t) {
+    { ::cuda::std::ranges::__iter_move_cpo{}(__t) } -> __can_reference;
+  }
+using iter_rvalue_reference_t = decltype(::cuda::std::ranges::__iter_move_cpo{}(::cuda::std::declval<_Tp&>()));
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __can_iter_rvalue_reference_t_,
+  requires(_Tp& __t)(requires(__dereferenceable<_Tp>),
+                     requires(__can_reference<decltype(::cuda::std::ranges::__iter_move_cpo{}(__t))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __can_iter_rvalue_reference_t = _CCCL_FRAGMENT(__can_iter_rvalue_reference_t_, _Tp);
+
+template <class _Tp>
+using __iter_rvalue_reference_t = decltype(::cuda::std::ranges::__iter_move_cpo{}(::cuda::std::declval<_Tp&>()));
+
+template <class _Tp>
+using iter_rvalue_reference_t = enable_if_t<__can_iter_rvalue_reference_t<_Tp>, __iter_rvalue_reference_t<_Tp>>;
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_DIAG_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ITER_MOVE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_swap.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_swap.h
new file mode 100644
index 00000000..0dd9d254
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iter_swap.h
@@ -0,0 +1,185 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES
+//
+//===----------------------------------------------------------------------===//
+#ifndef _CUDA_STD___ITERATOR_ITER_SWAP_H
+#define _CUDA_STD___ITERATOR_ITER_SWAP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__concepts/swappable.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/iter_move.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+// [iter.cust.swap]
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__iter_swap)
+template <class _I1, class _I2>
+void iter_swap(_I1, _I2) = delete;
+
+template <class _T1, class _T2>
+_CCCL_CONCEPT __unqualified_iter_swap = _CCCL_REQUIRES_EXPR((_T1, _T2), _T1&& __x, _T2&& __y)(
+  requires(__class_or_enum<remove_cvref_t<_T1>> || __class_or_enum<remove_cvref_t<_T2>>),
+  ((void) iter_swap(::cuda::std::forward<_T1>(__x), ::cuda::std::forward<_T2>(__y))));
+
+#if _CCCL_HAS_NOEXCEPT_MANGLING() // older GCC cannot use noexcept inside a requires clause
+template <class _T1, class _T2>
+_CCCL_CONCEPT __noexcept_unqualified_iter_swap = _CCCL_REQUIRES_EXPR((_T1, _T2), _T1&& __x, _T2&& __y)(
+  requires(__unqualified_iter_swap<_T1, _T2>),
+  noexcept(iter_swap(::cuda::std::forward<_T1>(__x), ::cuda::std::forward<_T2>(__y))));
+#else // ^^^ _CCCL_HAS_NOEXCEPT_MANGLING() ^^^ / vvv !_CCCL_HAS_NOEXCEPT_MANGLING() vvv
+template <class _T1, class _T2, bool = __unqualified_iter_swap<_T1, _T2>>
+inline constexpr bool __noexcept_unqualified_iter_swap = false;
+
+template <class _T1, class _T2>
+inline constexpr bool __noexcept_unqualified_iter_swap<_T1, _T2, true> =
+  noexcept(iter_swap(::cuda::std::declval<_T1>(), ::cuda::std::declval<_T2>()));
+#endif // !_CCCL_HAS_NOEXCEPT_MANGLING()
+
+template <class _T1, class _T2>
+_CCCL_CONCEPT __readable_swappable = _CCCL_REQUIRES_EXPR((_T1, _T2))(
+  requires(!__unqualified_iter_swap<_T1, _T2>),
+  requires(indirectly_readable<_T1>),
+  requires(indirectly_readable<_T2>),
+  requires(__can_reference<iter_reference_t<_T1>>),
+  requires(__can_reference<iter_reference_t<_T2>>),
+  requires(swappable_with<iter_reference_t<_T1>, iter_reference_t<_T2>>));
+
+#if _CCCL_HAS_NOEXCEPT_MANGLING() // older GCC cannot use noexcept inside a requires clause
+template <class _T1, class _T2>
+_CCCL_CONCEPT __noexcept_readable_swappable = _CCCL_REQUIRES_EXPR((_T1, _T2), _T1&& __x, _T2&& __y) //
+  (requires(__readable_swappable<_T1, _T2>),
+   noexcept(::cuda::std::ranges::__swap_cpo{}(*::cuda::std::forward<_T1>(__x), *::cuda::std::forward<_T2>(__y))));
+#else // ^^^ _CCCL_HAS_NOEXCEPT_MANGLING() ^^^ / vvv !_CCCL_HAS_NOEXCEPT_MANGLING() vvv
+template <class _T1, class _T2, bool = __readable_swappable<_T1, _T2>>
+inline constexpr bool __noexcept_readable_swappable = false;
+
+template <class _T1, class _T2>
+inline constexpr bool __noexcept_readable_swappable<_T1, _T2, true> =
+  noexcept(::cuda::std::ranges::__swap_cpo{}(*::cuda::std::declval<_T1>(), *::cuda::std::declval<_T2>()));
+#endif // !_CCCL_HAS_NOEXCEPT_MANGLING()
+
+template <class _T1, class _T2>
+_CCCL_CONCEPT __movable_storable = _CCCL_REQUIRES_EXPR((_T1, _T2))(
+  requires(!__unqualified_iter_swap<_T1, _T2>),
+  requires(!__readable_swappable<_T1, _T2>),
+  requires(indirectly_movable_storable<_T1, _T2>),
+  requires(indirectly_movable_storable<_T2, _T1>));
+
+#if _CCCL_HAS_NOEXCEPT_MANGLING() // older GCC cannot use noexcept inside a requires clause
+template <class _T1, class _T2>
+_CCCL_CONCEPT __noexcept_movable_storable =
+  _CCCL_REQUIRES_EXPR((_T1, _T2), _T1&& __x, _T2&& __y, iter_value_t<_T2> __old)(
+    requires(__movable_storable<_T1, _T2>),
+    noexcept(iter_value_t<_T2>(::cuda::std::ranges::__iter_move_cpo{}(__y))),
+    noexcept(*__y = ::cuda::std::ranges::__iter_move_cpo{}(__x)),
+    noexcept(*::cuda::std::forward<_T1>(__x) = ::cuda::std::move(__old)));
+#else // ^^^ _CCCL_HAS_NOEXCEPT_MANGLING() ^^^ / vvv !_CCCL_HAS_NOEXCEPT_MANGLING() vvv
+template <class _T1, class _T2, bool = __movable_storable<_T1, _T2>>
+inline constexpr bool __noexcept_movable_storable = false;
+
+template <class _T1, class _T2>
+inline constexpr bool __noexcept_movable_storable<_T1, _T2, true> =
+  noexcept(
+    iter_value_t<_T2>(::cuda::std::ranges::__iter_move_cpo{}(::cuda::std::declval<add_lvalue_reference_t<_T2>>())))
+  && noexcept(*::cuda::std::declval<add_lvalue_reference_t<_T2>>() =
+                ::cuda::std::ranges::__iter_move_cpo{}(::cuda::std::declval<add_lvalue_reference_t<_T1>>()))
+  && noexcept(*::cuda::std::declval<_T1>() = ::cuda::std::declval<iter_value_t<_T2>>());
+#endif // !_CCCL_HAS_NOEXCEPT_MANGLING()
+
+struct __fn
+{
+  _CCCL_TEMPLATE(class _T1, class _T2)
+  _CCCL_REQUIRES(__unqualified_iter_swap<_T1, _T2>)
+  _CCCL_API constexpr void operator()(_T1&& __x, _T2&& __y) const noexcept(__noexcept_unqualified_iter_swap<_T1, _T2>)
+  {
+    (void) iter_swap(::cuda::std::forward<_T1>(__x), ::cuda::std::forward<_T2>(__y));
+  }
+
+  _CCCL_TEMPLATE(class _T1, class _T2)
+  _CCCL_REQUIRES(__readable_swappable<_T1, _T2>)
+  _CCCL_API constexpr void operator()(_T1&& __x, _T2&& __y) const noexcept(__noexcept_readable_swappable<_T1, _T2>)
+  {
+    ::cuda::std::ranges::__swap_cpo{}(*::cuda::std::forward<_T1>(__x), *::cuda::std::forward<_T2>(__y));
+  }
+
+  _CCCL_TEMPLATE(class _T1, class _T2)
+  _CCCL_REQUIRES(__movable_storable<_T2, _T1>)
+  _CCCL_API constexpr void operator()(_T1&& __x, _T2&& __y) const noexcept(__noexcept_movable_storable<_T1, _T2>)
+  {
+    iter_value_t<_T2> __old(::cuda::std::ranges::__iter_move_cpo{}(__y));
+    *__y                            = ::cuda::std::ranges::__iter_move_cpo{}(__x);
+    *::cuda::std::forward<_T1>(__x) = ::cuda::std::move(__old);
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto iter_swap = __iter_swap::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __iter_swap_cpo = __iter_swap::__fn;
+} // namespace __cpo
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+#if _CCCL_HAS_CONCEPTS()
+template <class _I1, class _I2 = _I1>
+concept indirectly_swappable =
+  indirectly_readable<_I1> && indirectly_readable<_I2> && requires(const _I1 __i1, const _I2 __i2) {
+    ::cuda::std::ranges::__iter_swap_cpo{}(__i1, __i1);
+    ::cuda::std::ranges::__iter_swap_cpo{}(__i2, __i2);
+    ::cuda::std::ranges::__iter_swap_cpo{}(__i1, __i2);
+    ::cuda::std::ranges::__iter_swap_cpo{}(__i2, __i1);
+  };
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv _CCCL_HAS_CONCEPTS() vvv
+template <class _I1, class _I2>
+_CCCL_CONCEPT_FRAGMENT(
+  __indirectly_swappable_,
+  requires(const _I1 __i1, const _I2 __i2)(
+    requires(indirectly_readable<_I1>),
+    requires(indirectly_readable<_I2>),
+    (::cuda::std::ranges::__iter_swap_cpo{}(__i1, __i1)),
+    (::cuda::std::ranges::__iter_swap_cpo{}(__i2, __i2)),
+    (::cuda::std::ranges::__iter_swap_cpo{}(__i1, __i2)),
+    (::cuda::std::ranges::__iter_swap_cpo{}(__i2, __i1))));
+
+template <class _I1, class _I2 = _I1>
+_CCCL_CONCEPT indirectly_swappable = _CCCL_FRAGMENT(__indirectly_swappable_, _I1, _I2);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+template <class _I1, class _I2 = _I1, class = void>
+inline constexpr bool __noexcept_swappable = false;
+
+template <class _I1, class _I2>
+inline constexpr bool __noexcept_swappable<_I1, _I2, enable_if_t<indirectly_swappable<_I1, _I2>>> =
+  noexcept(::cuda::std::ranges::__iter_swap_cpo{}(::cuda::std::declval<_I1&>(), ::cuda::std::declval<_I2&>()));
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ITER_SWAP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator.h
new file mode 100644
index 00000000..0b865235
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator.h
@@ -0,0 +1,44 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_ITERATOR_H
+#define _CUDA_STD___ITERATOR_ITERATOR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Category, class _Tp, class _Distance = ptrdiff_t, class _Pointer = _Tp*, class _Reference = _Tp&>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT CCCL_DEPRECATED iterator
+{
+  using value_type        = _Tp;
+  using difference_type   = _Distance;
+  using pointer           = _Pointer;
+  using reference         = _Reference;
+  using iterator_category = _Category;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ITERATOR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator_traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator_traits.h
new file mode 100644
index 00000000..8dd0e366
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/iterator_traits.h
@@ -0,0 +1,618 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_ITERATOR_TRAITS_H
+#define _CUDA_STD___ITERATOR_ITERATOR_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/arithmetic.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__concepts/copyable.h>
+#include <cuda/std/__concepts/equality_comparable.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__concepts/totally_ordered.h>
+#include <cuda/std/__fwd/iterator.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__type_traits/add_const.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_primary_template.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/priority_tag.h>
+#include <cuda/std/cstddef>
+
+#if _CCCL_HOSTED()
+#  if _CCCL_COMPILER(MSVC)
+#    include <xutility> // for ::std::input_iterator_tag
+#  else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+#    include <iterator> // for ::std::input_iterator_tag
+#  endif // !_CCCL_COMPILER(MSVC)
+
+#  ifdef _GLIBCXX_DEBUG
+#    include <debug/safe_iterator.h>
+#  endif // _GLIBCXX_DEBUG
+
+#  if _CCCL_STD_VER >= 2020
+#    include <cuda/std/__cccl/prologue.h>
+template <class _Tp, class = void>
+struct __cccl_type_is_defined : ::cuda::std::false_type
+{};
+
+template <class _Tp>
+struct __cccl_type_is_defined<_Tp, ::cuda::std::void_t<decltype(sizeof(_Tp))>> : ::cuda::std::true_type
+{};
+
+// detect whether the used STL has contiguous_iterator_tag defined
+namespace std
+{
+struct __cccl_std_contiguous_iterator_tag_exists : __cccl_type_is_defined<struct contiguous_iterator_tag>
+{};
+} // namespace std
+
+#    include <cuda/std/__cccl/epilogue.h>
+#  endif // _CCCL_STD_VER >= 2020
+
+#endif // _CCCL_HOSTED()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+using __with_reference = _Tp&;
+
+template <class _Tp>
+_CCCL_CONCEPT __can_reference = _CCCL_REQUIRES_EXPR((_Tp))(typename(__with_reference<_Tp>));
+
+// [iterator.traits]
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __dereferenceable = requires(_Tp& __t) {
+  { *__t } -> __can_reference; // not required to be equality-preserving
+};
+
+template <__dereferenceable _Tp>
+using iter_reference_t = decltype(*::cuda::std::declval<_Tp&>());
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ // vvv _CCCL_HAS_CONCEPTS() vvv
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_CLANG("-Wvoid-ptr-dereference")
+
+template <class _Tp>
+_CCCL_CONCEPT __dereferenceable = _CCCL_REQUIRES_EXPR((_Tp), _Tp& __t)(requires(__can_reference<decltype(*__t)>));
+
+_CCCL_DIAG_POP
+
+template <class _Tp>
+using iter_reference_t = enable_if_t<__dereferenceable<_Tp>, decltype(*::cuda::std::declval<_Tp&>())>;
+#endif // _CCCL_HAS_CONCEPTS()
+
+#if _CCCL_FREESTANDING()
+
+struct _CCCL_TYPE_VISIBILITY_DEFAULT input_iterator_tag
+{};
+struct _CCCL_TYPE_VISIBILITY_DEFAULT output_iterator_tag
+{};
+struct _CCCL_TYPE_VISIBILITY_DEFAULT forward_iterator_tag : public input_iterator_tag
+{};
+struct _CCCL_TYPE_VISIBILITY_DEFAULT bidirectional_iterator_tag : public forward_iterator_tag
+{};
+struct _CCCL_TYPE_VISIBILITY_DEFAULT random_access_iterator_tag : public bidirectional_iterator_tag
+{};
+struct _CCCL_TYPE_VISIBILITY_DEFAULT contiguous_iterator_tag : public random_access_iterator_tag
+{};
+
+#else // ^^^ _CCCL_FREESTANDING() ^^^ / vvv _CCCL_HOSTED() vvv
+
+using input_iterator_tag         = ::std::input_iterator_tag;
+using output_iterator_tag        = ::std::output_iterator_tag;
+using forward_iterator_tag       = ::std::forward_iterator_tag;
+using bidirectional_iterator_tag = ::std::bidirectional_iterator_tag;
+using random_access_iterator_tag = ::std::random_access_iterator_tag;
+
+#  if _CCCL_STD_VER >= 2020
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __contiguous_iterator_tag_backfill : public ::std::random_access_iterator_tag
+{};
+using contiguous_iterator_tag =
+  _If<::std::__cccl_std_contiguous_iterator_tag_exists::value,
+      ::std::contiguous_iterator_tag,
+      __contiguous_iterator_tag_backfill>;
+#  else // ^^^ C++20 ^^^ / vvv C++17 vvv
+struct _CCCL_TYPE_VISIBILITY_DEFAULT contiguous_iterator_tag : public random_access_iterator_tag
+{};
+#  endif // _CCCL_STD_VER <= 2017
+
+#endif // _CCCL_HOSTED()
+
+template <class _Iter>
+struct __iter_traits_cache
+{
+  using type = __select_traits<remove_cvref_t<_Iter>, remove_cvref_t<_Iter>>;
+};
+template <class _Iter>
+using _ITER_TRAITS = typename __iter_traits_cache<_Iter>::type;
+
+#if _CCCL_HOSTED()
+#  if defined(_GLIBCXX_DEBUG)
+_CCCL_TEMPLATE(class _Iter, class _Ty, class _Range)
+_CCCL_REQUIRES(_IsSame<_Iter, ::__gnu_debug::_Safe_iterator<_Ty*, _Range>>::value)
+_CCCL_API inline auto __iter_concept_fn(::__gnu_debug::_Safe_iterator<_Ty*, _Range>, __priority_tag<3>)
+  -> contiguous_iterator_tag;
+#  endif // _GLIBCXX_DEBUG
+#  if _CCCL_HOST_STD_LIB(LIBSTDCXX)
+_CCCL_TEMPLATE(class _Iter, class _Ty, class _Range)
+_CCCL_REQUIRES(_IsSame<_Iter, ::__gnu_cxx::__normal_iterator<_Ty*, _Range>>::value)
+_CCCL_API inline auto __iter_concept_fn(::__gnu_cxx::__normal_iterator<_Ty*, _Range>, __priority_tag<3>)
+  -> contiguous_iterator_tag;
+#  endif // _CCCL_HOST_STD_LIB(LIBSTDCXX)
+#  if _CCCL_HOST_STD_LIB(LIBCXX)
+_CCCL_TEMPLATE(class _Iter, class _Ty)
+_CCCL_REQUIRES(_IsSame<_Iter, ::std::__wrap_iter<_Ty*>>::value)
+_CCCL_API inline auto __iter_concept_fn(::std::__wrap_iter<_Ty*>, __priority_tag<3>) -> contiguous_iterator_tag;
+#  elif _CCCL_HOST_STD_LIB(STL)
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_Array_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_Array_const_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_Vector_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_Vector_const_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_String_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_String_const_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_String_view_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+_CCCL_TEMPLATE(class _Iter)
+_CCCL_REQUIRES(_IsSame<_Iter, class _Iter::_Span_iterator>::value)
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<3>) -> contiguous_iterator_tag;
+#  endif // _CCCL_HOST_STD_LIB(STL)
+#endif // _CCCL_HOSTED()
+
+_CCCL_TEMPLATE(class _Iter, class _Ty)
+_CCCL_REQUIRES(_IsSame<_Iter, _Ty*>::value)
+_CCCL_API inline auto __iter_concept_fn(_Ty*, __priority_tag<3>) -> contiguous_iterator_tag;
+
+template <class _Iter>
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<2>) -> typename _ITER_TRAITS<_Iter>::iterator_concept;
+template <class _Iter>
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<1>) -> typename _ITER_TRAITS<_Iter>::iterator_category;
+template <class _Iter>
+_CCCL_API inline auto __iter_concept_fn(_Iter, __priority_tag<0>)
+  -> enable_if_t<__is_primary_cccl_template<_Iter>::value && __is_primary_std_template<_Iter>::value,
+                 random_access_iterator_tag>;
+
+template <class _Iter>
+using __iter_concept_t =
+  decltype(::cuda::std::__iter_concept_fn<_Iter>(::cuda::std::declval<_Iter>(), __priority_tag<3>{}));
+
+template <class _Iter, class = void>
+struct __iter_concept_cache
+{};
+
+template <class _Iter>
+struct __iter_concept_cache<_Iter, void_t<__iter_concept_t<_Iter>>>
+{
+  using type = __iter_concept_t<_Iter>;
+};
+
+template <class _Iter>
+using _ITER_CONCEPT = typename __iter_concept_cache<_Iter>::type;
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_reference = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::reference));
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_pointer = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::pointer));
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_iterator_category = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::iterator_category));
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_iterator_concept = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::iterator_concept));
+
+// The `cpp17-*-iterator` exposition-only concepts have very similar names to the `Cpp17*Iterator` named requirements
+// from `[iterator.cpp17]`. To avoid confusion between the two, the exposition-only concepts have been banished to
+// a "detail" namespace indicating they have a niche use-case.
+namespace __iterator_traits_detail
+{
+// [iterator.traits#concept:cpp17-iterator]
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_iterator = _CCCL_REQUIRES_EXPR((_Iter), _Iter __i)(
+  requires(copyable<_Iter>),
+  _Satisfies(__can_reference)(*__i),
+  _Same_as(_Iter&)(++__i),
+  _Satisfies(__can_reference)(*__i++));
+
+// [iterator.traits#concept:cpp17-input-iterator]
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_input_iterator = _CCCL_REQUIRES_EXPR((_Iter), _Iter __i)(
+  requires(__cpp17_iterator<_Iter>),
+  requires(equality_comparable<_Iter>),
+  typename(typename incrementable_traits<_Iter>::difference_type),
+  typename(typename indirectly_readable_traits<_Iter>::value_type),
+  typename(common_reference_t<iter_reference_t<_Iter>&&, typename indirectly_readable_traits<_Iter>::value_type&>),
+  typename(common_reference_t<decltype(*__i++)&&, typename indirectly_readable_traits<_Iter>::value_type&>),
+  requires(signed_integral<typename incrementable_traits<_Iter>::difference_type>));
+
+// [iterator.traits#concept:cpp17-forward-iterator]
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_forward_iterator = _CCCL_REQUIRES_EXPR((_Iter), _Iter __i)(
+  requires(__cpp17_input_iterator<_Iter>),
+  requires(constructible_from<_Iter>),
+  requires(is_lvalue_reference_v<iter_reference_t<_Iter>>),
+  requires(same_as<remove_cvref_t<iter_reference_t<_Iter>>, typename indirectly_readable_traits<_Iter>::value_type>),
+  requires(convertible_to<decltype(__i++), _Iter const&>),
+  _Same_as(iter_reference_t<_Iter>)(*__i++));
+
+// [iterator.traits#concept:cpp17-bidirectional-iterator]
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_bidirectional_iterator = _CCCL_REQUIRES_EXPR((_Iter), _Iter __i)(
+  requires(__cpp17_forward_iterator<_Iter>),
+  _Same_as(_Iter&)(--__i),
+  requires(convertible_to<decltype(__i--), _Iter const&>),
+  _Same_as(iter_reference_t<_Iter>)(*__i--));
+
+// [iterator.traits#concept:cpp17-random-access-iterator]
+// Needs to be its own concept, because we need `typename incrementable_traits<_Iter>::difference_type` to be valid
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_random_access_iterator_operations =
+  _CCCL_REQUIRES_EXPR((_Iter), _Iter __i, typename incrementable_traits<_Iter>::difference_type __n)(
+    _Same_as(_Iter&)(__i += __n),
+    _Same_as(_Iter&)(__i -= __n),
+    _Same_as(_Iter)(__i + __n),
+    _Same_as(_Iter)(__n + __i),
+    _Same_as(_Iter)(__i - __n),
+    _Same_as(decltype(__n))(__i - __i),
+    requires(convertible_to<decltype(__i[__n]), iter_reference_t<_Iter>>));
+
+template <class _Iter>
+_CCCL_CONCEPT __cpp17_random_access_iterator = _CCCL_REQUIRES_EXPR((_Iter))(
+  requires(__cpp17_bidirectional_iterator<_Iter>),
+  requires(totally_ordered<_Iter>),
+  requires(__cpp17_random_access_iterator_operations<_Iter>));
+} // namespace __iterator_traits_detail
+
+// [iterator.traits]#3.1
+// If the qualified-id I::pointer is valid and denotes a type, then pointer names that type;
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_pointer_or_void(int) -> typename _Iter::pointer;
+// Otherwise, it names void.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_pointer_or_void(...) -> void;
+
+template <class _Iter>
+using __iterator_traits_member_pointer_or_void =
+  decltype(::cuda::std::__iterator_traits_deduce_member_pointer_or_void<_Iter>(0));
+
+// [iterator.traits]#3.2
+// [iterator.traits]#3.2.1
+// If the qualified-id I::pointer is valid and denotes a type, then pointer names that type.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_pointer_or_arrow_or_void(int, __priority_tag<1>) ->
+  typename _Iter::pointer;
+
+// Otherwise, if decltype(declval<I&>().operator->()) is well-formed, then pointer names that type.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_pointer_or_arrow_or_void(int, __priority_tag<0>)
+  -> decltype(::cuda::std::declval<_Iter&>().operator->());
+
+// Otherwise, pointer names void.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_pointer_or_arrow_or_void(...) -> void;
+
+template <class _Iter>
+using __iterator_traits_member_pointer_or_arrow_or_void =
+  decltype(::cuda::std::__iterator_traits_deduce_member_pointer_or_arrow_or_void<_Iter>(0, __priority_tag<1>{}));
+
+// [iterator.traits]#3.2.2
+// If the qualified-id `I::reference` is valid and denotes a type, `reference` names that type.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_reference(int) -> typename _Iter::reference;
+// Otherwise, `reference` names `iter-reference-t<I>`.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_reference(...) -> iter_reference_t<_Iter>;
+
+template <class _Iter>
+using __iterator_traits_member_reference = decltype(::cuda::std::__iterator_traits_deduce_member_reference<_Iter>(0));
+
+// [iterator.traits]#3.2.3
+template <class _Iter>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL auto __iterator_traits_deduce_iterator_category() noexcept
+{
+  if constexpr (__has_member_iterator_category<_Iter>)
+  { // If the qualified-id `I::iterator-category` is valid and denotes a type, `iterator-category` names that type.
+    return typename _Iter::iterator_category{};
+  }
+  else if constexpr (__iterator_traits_detail::__cpp17_random_access_iterator<_Iter>)
+  { // Otherwise `random_access_iterator_tag` if `I` satisfies `cpp17-random-access-iterator`,
+    return random_access_iterator_tag{};
+  }
+  else if constexpr (__iterator_traits_detail::__cpp17_bidirectional_iterator<_Iter>)
+  { // or otherwise `bidirectional_iterator_tag` if `I` satisfies `cpp17-bidirectional-iterator`,
+    return bidirectional_iterator_tag{};
+  }
+  else if constexpr (__iterator_traits_detail::__cpp17_forward_iterator<_Iter>)
+  { // or otherwise `forward_iterator_tag` if `I` satisfies `cpp17-forward-iterator`,
+    return forward_iterator_tag{};
+  }
+  else
+  { // or otherwise input_iterator_tag
+    return input_iterator_tag{};
+  }
+}
+
+template <class _Iter>
+using __iterator_traits_iterator_category = decltype(::cuda::std::__iterator_traits_deduce_iterator_category<_Iter>());
+
+// [iterator.traits]#3.3
+// If the qualified-id `incrementable_traits<I>::difference_type` is valid and denotes a type, then
+// `difference_type` names that type;
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_difference(int) -> typename incrementable_traits<_Iter>::difference_type;
+// Otherwise, it names void.
+template <class _Iter>
+_CCCL_API auto __iterator_traits_deduce_member_difference(...) -> void;
+
+template <class _Iter>
+using __iterator_traits_difference_type = decltype(::cuda::std::__iterator_traits_deduce_member_difference<_Iter>(0));
+
+enum class __iterator_traits_selection
+{
+  __specialized_from_std,
+  __specifies_members,
+  __cpp17_input_iterator,
+  __cpp17_iterator,
+  __no_members,
+};
+
+// We need to consider if a user has specialized std::iterator_traits
+template <class _Iter>
+_CCCL_CONCEPT __specialized_from_std = !__is_primary_std_template<remove_cvref_t<_Iter>>::value;
+
+// If I has valid member types difference_type, value_type, reference, and iterator_category,
+template <class _Iter>
+_CCCL_CONCEPT __specifies_members = _CCCL_REQUIRES_EXPR((_Iter))(
+  typename(typename _Iter::value_type),
+  typename(typename _Iter::difference_type),
+  typename(typename _Iter::reference),
+  typename(typename _Iter::iterator_category));
+
+// [iterator.traits]#3.2.3
+template <class _Iter>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __iterator_traits_selection __select_iterator_traits_specialization() noexcept
+{
+  if constexpr (__specialized_from_std<_Iter>)
+  { // We need to consider if a user has specialized std::iterator_traits
+    return __iterator_traits_selection::__specialized_from_std;
+  }
+  if constexpr (__specifies_members<_Iter>)
+  { // If I has valid member types difference_type, value_type, reference, and iterator_category,
+    return __iterator_traits_selection::__specifies_members;
+  }
+  else if constexpr (__iterator_traits_detail::__cpp17_input_iterator<_Iter>)
+  { // Otherwise, if I satisfies the exposition-only concept cpp17-input-iterator,
+    return __iterator_traits_selection::__cpp17_input_iterator;
+  }
+  else if constexpr (__iterator_traits_detail::__cpp17_iterator<_Iter>)
+  { // Otherwise, if I satisfies the exposition-only concept cpp17-iterator,
+    return __iterator_traits_selection::__cpp17_iterator;
+  }
+  else
+  { // Otherwise, iterator_traits<I> has no members by any of the above names.
+    return __iterator_traits_selection::__no_members;
+  }
+}
+
+// [iterator.traits]#3
+template <class _Iter, __iterator_traits_selection = ::cuda::std::__select_iterator_traits_specialization<_Iter>()>
+struct __iterator_traits;
+
+#if _CCCL_HOSTED()
+// We need to properly accept specializations of `std::iterator_traits`
+template <class _Iter>
+struct __iterator_traits<_Iter, __iterator_traits_selection::__specialized_from_std>
+    : public ::std::iterator_traits<_Iter>
+{};
+#endif // _CCCL_HOSTED()
+
+// [iterator.traits]#3.1
+// If `I` has valid member types `difference-type`, `value-type`, `reference`, and
+// `iterator-category`, then `iterator-traits<I>` has the following publicly accessible members:
+template <class _Iter>
+struct __iterator_traits<_Iter, __iterator_traits_selection::__specifies_members>
+{
+  using iterator_category = typename _Iter::iterator_category;
+  using value_type        = typename _Iter::value_type;
+  using difference_type   = typename _Iter::difference_type;
+  using pointer           = __iterator_traits_member_pointer_or_void<_Iter>;
+  using reference         = typename _Iter::reference;
+};
+
+// [iterator.traits]#3.2
+// Otherwise, if `I` satisfies the exposition-only concept `cpp17-input-iterator`,
+// `iterator-traits<I>` has the following publicly accessible members:
+template <class _Iter>
+struct __iterator_traits<_Iter, __iterator_traits_selection::__cpp17_input_iterator>
+{
+  using iterator_category = __iterator_traits_iterator_category<_Iter>;
+  using value_type        = typename indirectly_readable_traits<_Iter>::value_type;
+  using difference_type   = typename incrementable_traits<_Iter>::difference_type;
+  using pointer           = __iterator_traits_member_pointer_or_arrow_or_void<_Iter>;
+  using reference         = __iterator_traits_member_reference<_Iter>;
+};
+
+// [iterator.traits]#3.3
+// Otherwise, if `I` satisfies the exposition-only concept `cpp17-iterator`, then
+// `iterator_traits<I>` has the following publicly accessible members:
+template <class _Iter>
+struct __iterator_traits<_Iter, __iterator_traits_selection::__cpp17_iterator>
+{
+  using iterator_category = output_iterator_tag;
+  using value_type        = void;
+  using difference_type   = __iterator_traits_difference_type<_Iter>;
+  using pointer           = void;
+  using reference         = void;
+};
+
+// [iterator.traits]#3.4
+// Otherwise, `iterator_traits<I>` has no members by any of the above names.
+template <class _Iter>
+struct __iterator_traits<_Iter, __iterator_traits_selection::__no_members>
+{};
+
+template <class _Iter, class>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT iterator_traits : __iterator_traits<_Iter>
+{
+  using __cccl_primary_template = iterator_traits;
+};
+
+// [iterator.traits]#5
+template <class _Tp>
+#if _CCCL_HAS_CONCEPTS()
+  requires is_object_v<_Tp>
+#endif // _CCCL_HAS_CONCEPTS()
+struct _CCCL_TYPE_VISIBILITY_DEFAULT iterator_traits<_Tp*>
+{
+  using difference_type   = ptrdiff_t;
+  using value_type        = remove_cv_t<_Tp>;
+  using pointer           = _Tp*;
+  using reference         = add_lvalue_reference_t<_Tp>;
+  using iterator_category = random_access_iterator_tag;
+  using iterator_concept  = contiguous_iterator_tag;
+};
+
+template <class _Iter, class _Tag>
+_CCCL_CONCEPT __has_iterator_category_convertible_to = _CCCL_REQUIRES_EXPR((_Iter, _Tag)) //
+  (typename(typename iterator_traits<_Iter>::iterator_category),
+   requires(is_convertible_v<typename iterator_traits<_Iter>::iterator_category, _Tag>));
+
+template <class _Iter, class _Tag>
+_CCCL_CONCEPT __has_iterator_concept_convertible_to = _CCCL_REQUIRES_EXPR((_Iter, _Tag)) //
+  (typename(typename _Iter::iterator_concept), requires(is_convertible_v<typename _Iter::iterator_concept, _Tag>));
+
+template <class _Iter>
+inline constexpr bool __has_input_traversal =
+  __has_iterator_category_convertible_to<_Iter, input_iterator_tag>
+  || __has_iterator_concept_convertible_to<_Iter, input_iterator_tag>;
+
+template <class _Iter>
+inline constexpr bool __has_forward_traversal =
+  __has_iterator_category_convertible_to<_Iter, forward_iterator_tag>
+  || __has_iterator_concept_convertible_to<_Iter, forward_iterator_tag>;
+
+template <class _Iter>
+inline constexpr bool __has_bidirectional_traversal =
+  __has_iterator_category_convertible_to<_Iter, bidirectional_iterator_tag>
+  || __has_iterator_concept_convertible_to<_Iter, bidirectional_iterator_tag>;
+
+template <class _Iter>
+inline constexpr bool __has_random_access_traversal =
+  __has_iterator_category_convertible_to<_Iter, random_access_iterator_tag>
+  || __has_iterator_concept_convertible_to<_Iter, random_access_iterator_tag>;
+
+// __has_contiguous_traversal determines if an iterator is known by
+// libc++ to be contiguous, either because it advertises itself as such
+// (in C++20) or because it is a pointer type or a known trivial wrapper
+// around a (possibly fancy) pointer type, such as __wrap_iter<T*>.
+// Such iterators receive special "contiguous" optimizations in
+// std::copy and std::sort.
+//
+template <class _Iter>
+inline constexpr bool __has_contiguous_traversal =
+  __has_iterator_category_convertible_to<_Iter, contiguous_iterator_tag>
+  || __has_iterator_concept_convertible_to<_Iter, contiguous_iterator_tag>;
+
+// Any native pointer which is an iterator is also a contiguous iterator.
+template <class _Tp>
+inline constexpr bool __has_contiguous_traversal<_Tp*> = true;
+
+template <class _Iter>
+using __iter_value_type = typename iterator_traits<_Iter>::value_type;
+
+template <class _Iter>
+using __iterator_category_type = typename iterator_traits<_Iter>::iterator_category;
+
+template <class _Iter>
+using __iterator_pointer_type = typename iterator_traits<_Iter>::pointer;
+
+template <class _Iter>
+using __iter_diff_t = typename iterator_traits<_Iter>::difference_type;
+
+template <class _Iter>
+using __iter_value_type = typename iterator_traits<_Iter>::value_type;
+
+// C++20 iterators do not play nicely with C++17 interfaces, because they commmonly use `iterator_concept` to
+// communicate their iterator category. Deduce the actual iterator category from the maximum of `iterator_concept` and
+// `iterator_category`
+template <class _Iter>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL auto __iterator_traits_category_or_concept() noexcept
+{
+  if constexpr (!__has_member_iterator_concept<_Iter>)
+  {
+    return __iterator_category_type<_Iter>{};
+  }
+  else if constexpr (__has_contiguous_traversal<_Iter>)
+  {
+    return contiguous_iterator_tag{};
+  }
+  else if constexpr (__has_random_access_traversal<_Iter>)
+  {
+    return random_access_iterator_tag{};
+  }
+  else if constexpr (__has_bidirectional_traversal<_Iter>)
+  {
+    return bidirectional_iterator_tag{};
+  }
+  else if constexpr (__has_forward_traversal<_Iter>)
+  {
+    return forward_iterator_tag{};
+  }
+  else if constexpr (__has_input_traversal<_Iter>)
+  {
+    return input_iterator_tag{};
+  }
+  else // if constexpr (__has_member_iterator_category<_Iter>)
+  {
+    return typename _Iter::iterator_category{};
+  }
+}
+
+template <class _Iter>
+using __iterator_traits_category_or_concept_t = decltype(::cuda::std::__iterator_traits_category_or_concept<_Iter>());
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_ITERATOR_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/next.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/next.h
new file mode 100644
index 00000000..f29ed0de
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/next.h
@@ -0,0 +1,104 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_NEXT_H
+#define _CUDA_STD___ITERATOR_NEXT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/advance.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/enable_if.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_TEMPLATE(class _InputIter)
+_CCCL_REQUIRES(__has_input_traversal<_InputIter>)
+[[nodiscard]] _CCCL_API constexpr _InputIter
+next(_InputIter __x, typename iterator_traits<_InputIter>::difference_type __n = 1)
+{
+  _CCCL_ASSERT(__n >= 0 || __has_bidirectional_traversal<_InputIter>,
+               "Attempt to next(it, n) with negative n on a non-bidirectional iterator");
+
+  ::cuda::std::advance(__x, __n);
+  return __x;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// [range.iter.op.next]
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__next)
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x) const
+  {
+    ++__x;
+    return __x;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x, iter_difference_t<_Ip> __n) const
+  {
+    ::cuda::std::ranges::__advance_cpo{}(__x, __n);
+    return __x;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip>&& sentinel_for<_Sp, _Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x, _Sp __bound_sentinel) const
+  {
+    ::cuda::std::ranges::__advance_cpo{}(__x, __bound_sentinel);
+    return __x;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip, class _Sp)
+  _CCCL_REQUIRES(input_or_output_iterator<_Ip>&& sentinel_for<_Sp, _Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x, iter_difference_t<_Ip> __n, _Sp __bound_sentinel) const
+  {
+    ::cuda::std::ranges::__advance_cpo{}(__x, __n, __bound_sentinel);
+    return __x;
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto next = __next::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __next_cpo = __next::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_NEXT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/prev.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/prev.h
new file mode 100644
index 00000000..dc9e5219
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/prev.h
@@ -0,0 +1,93 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_PREV_H
+#define _CUDA_STD___ITERATOR_PREV_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/advance.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__type_traits/enable_if.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_TEMPLATE(class _InputIter)
+_CCCL_REQUIRES(__has_input_traversal<_InputIter>)
+[[nodiscard]] _CCCL_API constexpr _InputIter
+prev(_InputIter __x, typename iterator_traits<_InputIter>::difference_type __n = 1)
+{
+  _CCCL_ASSERT(__n <= 0 || __has_bidirectional_traversal<_InputIter>, "Attempt to prev(it, +n) on a non-bidi iterator");
+  ::cuda::std::advance(__x, -__n);
+  return __x;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// [range.iter.op.prev]
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+_CCCL_BEGIN_NAMESPACE_CPO(__prev)
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(bidirectional_iterator<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x) const
+  {
+    --__x;
+    return __x;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(bidirectional_iterator<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x, iter_difference_t<_Ip> __n) const
+  {
+    ::cuda::std::ranges::__advance_cpo{}(__x, -__n);
+    return __x;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Ip)
+  _CCCL_REQUIRES(bidirectional_iterator<_Ip>)
+  [[nodiscard]] _CCCL_API constexpr _Ip operator()(_Ip __x, iter_difference_t<_Ip> __n, _Ip __bound_iter) const
+  {
+    ::cuda::std::ranges::__advance_cpo{}(__x, -__n, __bound_iter);
+    return __x;
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto prev = __prev::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __prev_cpo = __prev::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_PREV_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/readable_traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/readable_traits.h
new file mode 100644
index 00000000..17643d56
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/readable_traits.h
@@ -0,0 +1,156 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_READABLE_TRAITS_H
+#define _CUDA_STD___ITERATOR_READABLE_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__fwd/iterator.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_const.h>
+#include <cuda/std/__type_traits/is_object.h>
+#include <cuda/std/__type_traits/is_primary_template.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/remove_extent.h>
+#include <cuda/std/__type_traits/void_t.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_value_type = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::value_type));
+
+template <class _Tp>
+_CCCL_CONCEPT __has_member_element_type = _CCCL_REQUIRES_EXPR((_Tp))(typename(typename _Tp::element_type));
+
+template <class, class = void>
+struct __cond_value_type
+{};
+
+template <class _Tp>
+struct __cond_value_type<_Tp, enable_if_t<is_object_v<_Tp>>>
+{
+  using value_type = remove_cv_t<_Tp>;
+};
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [readable.traits]
+template <class>
+struct indirectly_readable_traits
+{};
+
+template <class _Ip>
+  requires is_array_v<_Ip>
+struct indirectly_readable_traits<_Ip>
+{
+  using value_type = remove_cv_t<remove_extent_t<_Ip>>;
+};
+
+template <class _Ip>
+struct indirectly_readable_traits<const _Ip> : indirectly_readable_traits<_Ip>
+{};
+
+template <class _Tp>
+struct indirectly_readable_traits<_Tp*> : __cond_value_type<_Tp>
+{};
+
+template <__has_member_value_type _Tp>
+struct indirectly_readable_traits<_Tp> : __cond_value_type<typename _Tp::value_type>
+{};
+
+template <__has_member_element_type _Tp>
+struct indirectly_readable_traits<_Tp> : __cond_value_type<typename _Tp::element_type>
+{};
+
+template <__has_member_value_type _Tp>
+  requires __has_member_element_type<_Tp>
+struct indirectly_readable_traits<_Tp>
+{};
+
+template <__has_member_value_type _Tp>
+  requires __has_member_element_type<_Tp>
+        && same_as<remove_cv_t<typename _Tp::element_type>, remove_cv_t<typename _Tp::value_type>>
+struct indirectly_readable_traits<_Tp> : __cond_value_type<typename _Tp::value_type>
+{};
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+// [readable.traits]
+template <class, class = void>
+struct indirectly_readable_traits
+{};
+
+template <class _Ip>
+struct indirectly_readable_traits<_Ip, enable_if_t<!is_const_v<_Ip> && is_array_v<_Ip>>>
+{
+  using value_type = remove_cv_t<remove_extent_t<_Ip>>;
+};
+
+template <class _Ip>
+struct indirectly_readable_traits<const _Ip> : indirectly_readable_traits<_Ip>
+{};
+
+template <class _Tp>
+struct indirectly_readable_traits<_Tp*> : __cond_value_type<_Tp>
+{};
+
+template <class _Tp>
+struct indirectly_readable_traits<
+  _Tp,
+  enable_if_t<!is_const_v<_Tp> && __has_member_value_type<_Tp> && !__has_member_element_type<_Tp>>>
+    : __cond_value_type<typename _Tp::value_type>
+{};
+
+template <class _Tp>
+struct indirectly_readable_traits<
+  _Tp,
+  enable_if_t<!is_const_v<_Tp> && !__has_member_value_type<_Tp> && __has_member_element_type<_Tp>>>
+    : __cond_value_type<typename _Tp::element_type>
+{};
+
+template <class _Tp>
+struct indirectly_readable_traits<
+  _Tp,
+  enable_if_t<!is_const_v<_Tp> && __has_member_value_type<_Tp> && __has_member_element_type<_Tp>
+              && same_as<remove_cv_t<typename _Tp::element_type>, remove_cv_t<typename _Tp::value_type>>>>
+    : __cond_value_type<typename _Tp::value_type>
+{};
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+// Let `RI` be `remove_cvref_t<I>`. The type `iter_value_t<I>` denotes
+// `indirectly_readable_traits<RI>::value_type` if `iterator_traits<RI>` names a specialization
+// generated from the primary template, and `iterator_traits<RI>::value_type` otherwise.
+template <class _Ip>
+using iter_value_t =
+  typename __select_traits<remove_cvref_t<_Ip>, indirectly_readable_traits<remove_cvref_t<_Ip>>>::value_type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_READABLE_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_access.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_access.h
new file mode 100644
index 00000000..c0d9813d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_access.h
@@ -0,0 +1,154 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_REVERSE_ACCESS_H
+#define _CUDA_STD___ITERATOR_REVERSE_ACCESS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__iterator/reverse_iterator.h>
+#include <cuda/std/cstddef>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+namespace __rbegin
+{
+struct __fn
+{
+  template <class _Tp, size_t _Np>
+  _CCCL_API constexpr reverse_iterator<_Tp*> operator()(_Tp (&__array)[_Np]) const noexcept
+  {
+    return reverse_iterator<_Tp*>(__array + _Np);
+  }
+
+  template <class _Ep>
+  _CCCL_API constexpr reverse_iterator<const _Ep*> operator()(initializer_list<_Ep> __il) const noexcept
+  {
+    return reverse_iterator<const _Ep*>(__il.end());
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(_Cp& __c) const noexcept(noexcept(__c.rbegin())) -> decltype(__c.rbegin())
+  {
+    return __c.rbegin();
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(__c.rbegin())) -> decltype(__c.rbegin())
+  {
+    return __c.rbegin();
+  }
+};
+} // namespace __rbegin
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto rbegin = __rbegin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __rbegin_cpo = __rbegin::__fn;
+} // namespace __cpo
+
+namespace __rend
+{
+struct __fn
+{
+  template <class _Tp, size_t _Np>
+  _CCCL_API constexpr reverse_iterator<_Tp*> operator()(_Tp (&__array)[_Np]) const noexcept
+  {
+    return reverse_iterator<_Tp*>(__array);
+  }
+
+  template <class _Ep>
+  _CCCL_API constexpr reverse_iterator<const _Ep*> operator()(initializer_list<_Ep> __il) const noexcept
+  {
+    return reverse_iterator<const _Ep*>(__il.begin());
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(_Cp& __c) const noexcept(noexcept(__c.rend())) -> decltype(__c.rend())
+  {
+    return __c.rend();
+  }
+
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(__c.rend())) -> decltype(__c.rend())
+  {
+    return __c.rend();
+  }
+};
+} // namespace __rend
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto rend = __rend::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __rend_cpo = __rend::__fn;
+} // namespace __cpo
+
+namespace __crbegin
+{
+struct __fn
+{
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(::cuda::std::__rbegin_cpo{}(__c)))
+    -> decltype(::cuda::std::__rbegin_cpo{}(__c))
+  {
+    return ::cuda::std::__rbegin_cpo{}(__c);
+  }
+};
+} // namespace __crbegin
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto crbegin = __crbegin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __crbegin_cpo = __crbegin::__fn;
+} // namespace __cpo
+
+namespace __crend
+{
+struct __fn
+{
+  template <class _Cp>
+  _CCCL_API constexpr auto operator()(const _Cp& __c) const noexcept(noexcept(::cuda::std::__rend_cpo{}(__c)))
+    -> decltype(::cuda::std::__rend_cpo{}(__c))
+  {
+    return ::cuda::std::__rend_cpo{}(__c);
+  }
+};
+} // namespace __crend
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto crend = __crend::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __crend_cpo = __crend::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_REVERSE_ACCESS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_iterator.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_iterator.h
new file mode 100644
index 00000000..1f1793da
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/reverse_iterator.h
@@ -0,0 +1,375 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_REVERSE_ITERATOR_H
+#define _CUDA_STD___ITERATOR_REVERSE_ITERATOR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/unwrap_iter.h>
+#include <cuda/std/__fwd/iterator.h>
+#if _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+#  include <cuda/std/__compare/compare_three_way_result.h>
+#  include <cuda/std/__compare/three_way_comparable.h>
+#endif // _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__iterator/advance.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iter_move.h>
+#include <cuda/std/__iterator/iter_swap.h>
+#include <cuda/std/__iterator/iterator.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__iterator/next.h>
+#include <cuda/std/__iterator/prev.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__memory/addressof.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_default_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_convertible.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_default_constructible.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Iter, class = void>
+inline constexpr bool __noexcept_rev_iter_iter_move = false;
+
+template <class _Iter>
+inline constexpr bool __noexcept_rev_iter_iter_move<_Iter, void_t<decltype(--::cuda::std::declval<_Iter&>())>> =
+  is_nothrow_copy_constructible_v<_Iter>
+  && noexcept(::cuda::std::ranges::__iter_move_cpo{}(--::cuda::std::declval<_Iter&>()));
+
+template <class _Iter, class _Iter2, class = void>
+inline constexpr bool __noexcept_rev_iter_iter_swap = false;
+
+template <class _Iter, class _Iter2>
+inline constexpr bool __noexcept_rev_iter_iter_swap<_Iter, _Iter2, enable_if_t<indirectly_swappable<_Iter, _Iter2>>> =
+  is_nothrow_copy_constructible_v<_Iter> && is_nothrow_copy_constructible_v<_Iter2>
+  && noexcept(::cuda::std::ranges::__iter_swap_cpo{}(--declval<_Iter&>(), --declval<_Iter2&>()));
+
+// MSVC has issues with `is_nothrow_convertible_v` sometimes, so do the noexcept expression
+template <class _Iter, class _Iter2, class = void>
+inline constexpr bool __noexcept_rev_iter_convertible = false;
+
+template <class _Iter, class _Iter2>
+inline constexpr bool
+  __noexcept_rev_iter_convertible<_Iter, _Iter2, enable_if_t<is_convertible_v<_Iter const&, _Iter2>>> =
+    noexcept(_Iter2(::cuda::std::declval<const _Iter&>()));
+
+_LIBCUDACXX_BEGIN_HIDDEN_FRIEND_NAMESPACE
+
+_CCCL_SUPPRESS_DEPRECATED_PUSH
+_CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+template <class _Iter>
+class _CCCL_TYPE_VISIBILITY_DEFAULT reverse_iterator
+{
+private:
+#if _CCCL_STD_VER > 2017
+  static_assert(__has_bidirectional_traversal<_Iter> || bidirectional_iterator<_Iter>,
+                "reverse_iterator<It> requires It to be a bidirectional iterator.");
+#endif // _CCCL_STD_VER > 2017
+
+protected:
+  _Iter current;
+
+public:
+  using iterator_type = _Iter;
+
+  using iterator_category =
+    _If<__has_random_access_traversal<_Iter>, random_access_iterator_tag, __iterator_traits_category_or_concept_t<_Iter>>;
+  using pointer          = typename iterator_traits<_Iter>::pointer;
+  using iterator_concept = _If<random_access_iterator<_Iter>, random_access_iterator_tag, bidirectional_iterator_tag>;
+  using value_type       = iter_value_t<_Iter>;
+  using difference_type  = iter_difference_t<_Iter>;
+  using reference        = iter_reference_t<_Iter>;
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _It2 = _Iter)
+  _CCCL_REQUIRES(is_default_constructible_v<_It2>)
+  _CCCL_API constexpr reverse_iterator() noexcept(is_nothrow_default_constructible_v<_It2>)
+      : current()
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr explicit reverse_iterator(_Iter __x) noexcept(is_nothrow_copy_constructible_v<_Iter>)
+      : current(__x)
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Up)
+  _CCCL_REQUIRES((!is_same_v<_Up, _Iter>) _CCCL_AND is_convertible_v<_Up const&, _Iter>)
+  _CCCL_API constexpr reverse_iterator(const reverse_iterator<_Up>& __u) noexcept(
+    __noexcept_rev_iter_convertible<_Up, _Iter>)
+      : current(__u.base())
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Up)
+  _CCCL_REQUIRES((
+    !is_same_v<_Up, _Iter>) _CCCL_AND is_convertible_v<_Up const&, _Iter> _CCCL_AND is_assignable_v<_Iter&, _Up const&>)
+  _CCCL_API constexpr reverse_iterator&
+  operator=(const reverse_iterator<_Up>& __u) noexcept(is_nothrow_assignable_v<_Iter&, _Up const&>)
+  {
+    current = __u.base();
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API constexpr _Iter base() const noexcept(is_nothrow_copy_constructible_v<_Iter>)
+  {
+    return current;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API constexpr reference operator*() const
+  {
+    return *::cuda::std::prev(current);
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Iter2 = _Iter)
+  _CCCL_REQUIRES(__has_const_arrow<_Iter2>)
+  _CCCL_API constexpr pointer operator->() const
+  {
+    if constexpr (is_pointer_v<_Iter>)
+    {
+      return ::cuda::std::prev(current);
+    }
+    else
+    {
+      return ::cuda::std::prev(current).operator->();
+    }
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator& operator++()
+  {
+    --current;
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator operator++(int)
+  {
+    reverse_iterator __tmp{*this};
+    --current;
+    return __tmp;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator& operator--()
+  {
+    ++current;
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator operator--(int)
+  {
+    reverse_iterator __tmp{*this};
+    ++current;
+    return __tmp;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator operator+(difference_type __n) const
+  {
+    return reverse_iterator{current - __n};
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API friend constexpr reverse_iterator operator+(difference_type __n, const reverse_iterator& __x)
+  {
+    return reverse_iterator{__x.base() - __n};
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator& operator+=(difference_type __n)
+  {
+    current -= __n;
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator operator-(difference_type __n) const
+  {
+    return reverse_iterator{current + __n};
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator-(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+    -> decltype(::cuda::std::declval<const _Iter2&>() - ::cuda::std::declval<const _Iter&>())
+  {
+    return __y.base() - __x.base();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr reverse_iterator& operator-=(difference_type __n)
+  {
+    current += __n;
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  [[nodiscard]] _CCCL_API constexpr reference operator[](difference_type __n) const
+  {
+    return *(*this + __n);
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2 = _Iter>
+  [[nodiscard]] _CCCL_API friend constexpr iter_rvalue_reference_t<_Iter2>
+  iter_move(const reverse_iterator& __i) noexcept(__noexcept_rev_iter_iter_move<_Iter2>)
+  {
+    auto __tmp = __i.base();
+    return ::cuda::std::ranges::__iter_move_cpo{}(--__tmp);
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  _CCCL_API friend constexpr auto iter_swap(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y) noexcept(
+    __noexcept_rev_iter_iter_swap<_Iter, _Iter2>) _CCCL_TRAILING_REQUIRES(void)(indirectly_swappable<_Iter2, _Iter>)
+  {
+    auto __xtmp = __x.base();
+    auto __ytmp = __y.base();
+    return ::cuda::std::ranges::__iter_swap_cpo{}(--__xtmp, --__ytmp);
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator==(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y) noexcept(
+    noexcept(bool(::cuda::std::declval<const _Iter&>() == ::cuda::std::declval<const _Iter2&>())))
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() == ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() == __y.base();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator!=(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y) noexcept(
+    noexcept(bool(::cuda::std::declval<const _Iter&>() != ::cuda::std::declval<const _Iter2&>())))
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() != ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() != __y.base();
+  }
+
+#if _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Iter2)
+  _CCCL_REQUIRES(three_way_comparable_with<_Iter, _Iter2>)
+  [[nodiscard]] _CCCL_API friend constexpr compare_three_way_result_t<_Iter, _Iter2>
+  operator<=>(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+  {
+    return __y.base() <=> __x.base();
+  }
+#else // ^^^ _LIBCUDACXX_HAS_SPACESHIP_OPERATOR() ^^^ / vvv !_LIBCUDACXX_HAS_SPACESHIP_OPERATOR() vvv
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator<(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() > ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() > __y.base();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator>(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() < ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() < __y.base();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator>=(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() <= ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() <= __y.base();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _Iter2>
+  [[nodiscard]] _CCCL_API friend constexpr auto
+  operator<=(const reverse_iterator& __x, const reverse_iterator<_Iter2>& __y)
+    -> decltype(static_cast<bool>(::cuda::std::declval<const _Iter&>() >= ::cuda::std::declval<const _Iter2&>()))
+  {
+    return __x.base() >= __y.base();
+  }
+#endif // !_LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+};
+_CCCL_SUPPRESS_DEPRECATED_POP
+
+_LIBCUDACXX_END_HIDDEN_FRIEND_NAMESPACE(reverse_iterator)
+
+template <class _Iter1, class _Iter2>
+inline constexpr bool disable_sized_sentinel_for<reverse_iterator<_Iter1>, reverse_iterator<_Iter2>> =
+  !sized_sentinel_for<_Iter1, _Iter2>;
+
+template <class _Iter>
+[[nodiscard]] _CCCL_API constexpr reverse_iterator<_Iter>
+make_reverse_iterator(_Iter __i) noexcept(is_nothrow_copy_constructible_v<_Iter>)
+{
+  return reverse_iterator<_Iter>{__i};
+}
+
+template <template <class> class _RevIter1, template <class> class _RevIter2, class _Iter>
+struct __unwrap_reverse_iter_impl
+{
+  using _UnwrappedIter  = decltype(__unwrap_iter_impl<_Iter>::__unwrap(::cuda::std::declval<_Iter>()));
+  using _ReverseWrapper = _RevIter1<_RevIter2<_Iter>>;
+
+  _CCCL_API static constexpr _ReverseWrapper __rewrap(_ReverseWrapper __orig_iter, _UnwrappedIter __unwrapped_iter)
+  {
+    return _ReverseWrapper(
+      _RevIter2<_Iter>(__unwrap_iter_impl<_Iter>::__rewrap(__orig_iter.base().base(), __unwrapped_iter)));
+  }
+
+  _CCCL_API static constexpr _UnwrappedIter __unwrap(_ReverseWrapper __i) noexcept
+  {
+    return __unwrap_iter_impl<_Iter>::__unwrap(__i.base().base());
+  }
+};
+
+template <class _Iter, bool __b>
+struct __unwrap_iter_impl<reverse_iterator<reverse_iterator<_Iter>>, __b>
+    : __unwrap_reverse_iter_impl<reverse_iterator, reverse_iterator, _Iter>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_REVERSE_ITERATOR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/size.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/size.h
new file mode 100644
index 00000000..c66a3c2f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__iterator/size.h
@@ -0,0 +1,69 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___ITERATOR_SIZE_H
+#define _CUDA_STD___ITERATOR_SIZE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/common_type.h>
+#include <cuda/std/__type_traits/make_signed.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Cont>
+_CCCL_API constexpr auto size(const _Cont& __c) noexcept(noexcept(__c.size())) -> decltype(__c.size())
+{
+  return __c.size();
+}
+
+template <class _Tp, size_t _Sz>
+_CCCL_API constexpr size_t size(const _Tp (&)[_Sz]) noexcept
+{
+  return _Sz;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Cont>
+_CCCL_API constexpr auto ssize(const _Cont& __c) noexcept(
+  noexcept(static_cast<common_type_t<ptrdiff_t, make_signed_t<decltype(__c.size())>>>(__c.size())))
+  -> common_type_t<ptrdiff_t, make_signed_t<decltype(__c.size())>>
+{
+  return static_cast<common_type_t<ptrdiff_t, make_signed_t<decltype(__c.size())>>>(__c.size());
+}
+
+// GCC complains about the implicit conversion from ptrdiff_t to size_t in
+// the array bound.
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_GCC("-Wsign-conversion")
+template <class _Tp, ptrdiff_t _Sz>
+_CCCL_API constexpr ptrdiff_t ssize(const _Tp (&)[_Sz]) noexcept
+{
+  return _Sz;
+}
+_CCCL_DIAG_POP
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___ITERATOR_SIZE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits.h
new file mode 100644
index 00000000..cec3dc02
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits.h
@@ -0,0 +1,612 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___LIMITS_NUMERIC_LIMITS_H
+#define _CUDA_STD___LIMITS_NUMERIC_LIMITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__bit/bit_cast.h>
+#include <cuda/std/__type_traits/is_extended_floating_point.h>
+#include <cuda/std/__type_traits/is_floating_point.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/make_unsigned.h>
+#include <cuda/std/cfloat>
+#include <cuda/std/climits>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+enum float_round_style
+{
+  round_indeterminate       = -1,
+  round_toward_zero         = 0,
+  round_to_nearest          = 1,
+  round_toward_infinity     = 2,
+  round_toward_neg_infinity = 3
+};
+
+enum _CCCL_DEPRECATED_IN_CXX23 float_denorm_style
+{
+  denorm_indeterminate = -1,
+  denorm_absent        = 0,
+  denorm_present       = 1
+};
+
+enum class __numeric_limits_type
+{
+  __integral,
+  __bool,
+  __floating_point,
+  __other,
+};
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __numeric_limits_type __make_numeric_limits_type() noexcept
+{
+  if constexpr (is_same_v<_Tp, bool>)
+  {
+    return __numeric_limits_type::__bool;
+  }
+  else if constexpr (is_integral_v<_Tp>)
+  {
+    return __numeric_limits_type::__integral;
+  }
+  else if constexpr (is_floating_point_v<_Tp> || __is_extended_floating_point_v<_Tp>)
+  {
+    return __numeric_limits_type::__floating_point;
+  }
+  else
+  {
+    return __numeric_limits_type::__other;
+  }
+}
+
+// To avoid including nvfp headers, we add the _Up type defaulted to _Tp which makes the specialization still be a
+// template, which won't be instantiated unless the numeric_limits<_Tp> class is instantiated. The specialization should
+// look as:
+//
+// template <class _Tp>
+// class __numeric_limits_impl<__nvfp_type, __numeric_limits_type::__floating_point, _Tp>
+// { ... };
+//
+// and _Tp should be used everywhere instead of __nvfp_type.
+template <class _Tp, __numeric_limits_type = __make_numeric_limits_type<_Tp>(), class _Up = _Tp>
+class __numeric_limits_impl
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = false;
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return type();
+  }
+
+  static constexpr int digits       = 0;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 0;
+  static constexpr bool is_signed   = false;
+  static constexpr bool is_integer  = false;
+  static constexpr bool is_exact    = false;
+  static constexpr int radix        = 0;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return type();
+  }
+
+  static constexpr int min_exponent   = 0;
+  static constexpr int min_exponent10 = 0;
+  static constexpr int max_exponent   = 0;
+  static constexpr int max_exponent10 = 0;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_absent;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return type();
+  }
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+    return type();
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = false;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_toward_zero;
+};
+
+template <class _Tp>
+class __numeric_limits_impl<_Tp, __numeric_limits_type::__integral>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = type(-1) < type(0);
+  static constexpr int digits       = static_cast<int>(sizeof(type) * CHAR_BIT - is_signed);
+  static constexpr int digits10     = digits * 3 / 10;
+  static constexpr int max_digits10 = 0;
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return static_cast<_Tp>(~max());
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    using _Up = make_unsigned_t<_Tp>;
+    return static_cast<_Tp>(static_cast<_Up>(~_Up{0}) >> static_cast<int>(is_signed));
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return min();
+  }
+
+  static constexpr bool is_integer = true;
+  static constexpr bool is_exact   = true;
+  static constexpr int radix       = 2;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return type(0);
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return type(0);
+  }
+
+  static constexpr int min_exponent   = 0;
+  static constexpr int min_exponent10 = 0;
+  static constexpr int max_exponent   = 0;
+  static constexpr int max_exponent10 = 0;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_absent;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return type(0);
+  }
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return type(0);
+  }
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return type(0);
+  }
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+    return type(0);
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = !is_signed;
+
+#if (_CCCL_HOST_ARCH(X86_64) && _CCCL_OS(LINUX)) || defined(__pnacl__) || defined(__wasm__)
+  static constexpr bool traps = true;
+#else
+  static constexpr bool traps = false;
+#endif
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_toward_zero;
+};
+
+template <>
+class __numeric_limits_impl<bool, __numeric_limits_type::__bool>
+{
+public:
+  using type = bool;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = false;
+  static constexpr int digits       = 1;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 0;
+
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return false;
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return true;
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return min();
+  }
+
+  static constexpr bool is_integer = true;
+  static constexpr bool is_exact   = true;
+  static constexpr int radix       = 2;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return false;
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return false;
+  }
+
+  static constexpr int min_exponent   = 0;
+  static constexpr int min_exponent10 = 0;
+  static constexpr int max_exponent   = 0;
+  static constexpr int max_exponent10 = 0;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_absent;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return false;
+  }
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return false;
+  }
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return false;
+  }
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+    return false;
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_toward_zero;
+};
+
+template <>
+class __numeric_limits_impl<float, __numeric_limits_type::__floating_point>
+{
+public:
+  using type = float;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = FLT_MANT_DIG;
+  static constexpr int digits10     = FLT_DIG;
+  static constexpr int max_digits10 = 2 + (digits * 30103l) / 100000l;
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return FLT_MIN;
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return FLT_MAX;
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return -max();
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return FLT_EPSILON;
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return 0.5F;
+  }
+
+  static constexpr int min_exponent   = FLT_MIN_EXP;
+  static constexpr int min_exponent10 = FLT_MIN_10_EXP;
+  static constexpr int max_exponent   = FLT_MAX_EXP;
+  static constexpr int max_exponent10 = FLT_MAX_10_EXP;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+
+#if defined(_CCCL_BUILTIN_HUGE_VALF)
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return _CCCL_BUILTIN_HUGE_VALF();
+  }
+#else // ^^^ _CCCL_BUILTIN_HUGE_VALF ^^^ // vvv !_CCCL_BUILTIN_HUGE_VALF vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type infinity() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7f800000);
+  }
+#endif // !_CCCL_BUILTIN_HUGE_VALF
+#if defined(_CCCL_BUILTIN_NANF)
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANF("");
+  }
+#else // ^^^ _CCCL_BUILTIN_NANF ^^^ // vvv !_CCCL_BUILTIN_NANF vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type quiet_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7fc00000);
+  }
+#endif // !_CCCL_BUILTIN_NANF
+#if defined(_CCCL_BUILTIN_NANSF)
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANSF("");
+  }
+#else // ^^^ _CCCL_BUILTIN_NANSF ^^^ // vvv !_CCCL_BUILTIN_NANSF vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type signaling_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7fa00000);
+  }
+#endif // !_CCCL_BUILTIN_NANSF
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+#if defined(FLT_TRUE_MIN)
+    return FLT_TRUE_MIN;
+#else // ^^^ FLT_TRUE_MIN ^^^ // vvv !FLT_TRUE_MIN vvv
+    return __FLT_DENORM_MIN__;
+#endif // ^^^ !FLT_TRUE_MIN ^^^
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+
+template <>
+class __numeric_limits_impl<double, __numeric_limits_type::__floating_point>
+{
+public:
+  using type = double;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = DBL_MANT_DIG;
+  static constexpr int digits10     = DBL_DIG;
+  static constexpr int max_digits10 = 2 + (digits * 30103l) / 100000l;
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return DBL_MIN;
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return DBL_MAX;
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return -max();
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return DBL_EPSILON;
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return 0.5;
+  }
+
+  static constexpr int min_exponent   = DBL_MIN_EXP;
+  static constexpr int min_exponent10 = DBL_MIN_10_EXP;
+  static constexpr int max_exponent   = DBL_MAX_EXP;
+  static constexpr int max_exponent10 = DBL_MAX_10_EXP;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+
+#if defined(_CCCL_BUILTIN_HUGE_VAL)
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return _CCCL_BUILTIN_HUGE_VAL();
+  }
+#else // ^^^ _CCCL_BUILTIN_HUGE_VAL ^^^ // vvv !_CCCL_BUILTIN_HUGE_VAL vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type infinity() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7ff0000000000000);
+  }
+#endif // !_CCCL_BUILTIN_HUGE_VAL
+#if defined(_CCCL_BUILTIN_NAN)
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NAN("");
+  }
+#else // ^^^ _CCCL_BUILTIN_NAN ^^^ // vvv !_CCCL_BUILTIN_NAN vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type quiet_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7ff8000000000000);
+  }
+#endif // !_CCCL_BUILTIN_NAN
+#if defined(_CCCL_BUILTIN_NANS)
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANS("");
+  }
+#else // ^^^ _CCCL_BUILTIN_NANS ^^^ // vvv !_CCCL_BUILTIN_NANS vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type signaling_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(0x7ff4000000000000);
+  }
+#endif // !_CCCL_BUILTIN_NANS
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+#if defined(DBL_TRUE_MIN)
+    return DBL_TRUE_MIN;
+#else // ^^^ DBL_TRUE_MIN ^^^ // vvv !DBL_TRUE_MIN vvv
+    return __DBL_DENORM_MIN__;
+#endif // ^^^ !DBL_TRUE_MIN ^^^
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+
+template <>
+class __numeric_limits_impl<long double, __numeric_limits_type::__floating_point>
+{
+#if _CCCL_HAS_LONG_DOUBLE()
+
+public:
+  using type = long double;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = LDBL_MANT_DIG;
+  static constexpr int digits10     = LDBL_DIG;
+  static constexpr int max_digits10 = 2 + (digits * 30103l) / 100000l;
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return LDBL_MIN;
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return LDBL_MAX;
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return -max();
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return LDBL_EPSILON;
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return 0.5L;
+  }
+
+  static constexpr int min_exponent   = LDBL_MIN_EXP;
+  static constexpr int min_exponent10 = LDBL_MIN_10_EXP;
+  static constexpr int max_exponent   = LDBL_MAX_EXP;
+  static constexpr int max_exponent10 = LDBL_MAX_10_EXP;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return _CCCL_BUILTIN_HUGE_VALL();
+  }
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANL("");
+  }
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANSL("");
+  }
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+#  if defined(LDBL_TRUE_MIN)
+    return LDBL_TRUE_MIN;
+#  else // ^^^ LDBL_TRUE_MIN ^^^ // vvv !LDBL_TRUE_MIN vvv
+    return __LDBL_DENORM_MIN__;
+#  endif // ^^^ !LDBL_TRUE_MIN ^^^
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+#endif // _CCCL_HAS_LONG_DOUBLE()
+};
+
+template <class _Tp>
+class numeric_limits : public __numeric_limits_impl<_Tp>
+{};
+
+template <class _Tp>
+class numeric_limits<const _Tp> : public numeric_limits<_Tp>
+{};
+
+template <class _Tp>
+class numeric_limits<volatile _Tp> : public numeric_limits<_Tp>
+{};
+
+template <class _Tp>
+class numeric_limits<const volatile _Tp> : public numeric_limits<_Tp>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___LIMITS_NUMERIC_LIMITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits_ext.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits_ext.h
new file mode 100644
index 00000000..b8e0d3f7
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__limits/numeric_limits_ext.h
@@ -0,0 +1,753 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___LIMITS_NUMERIC_LIMITS_EXT_H
+#define _CUDA_STD___LIMITS_NUMERIC_LIMITS_EXT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__bit/bit_cast.h>
+#include <cuda/std/__floating_point/storage.h>
+#include <cuda/std/__limits/numeric_limits.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// nvfp16
+
+#if _CCCL_HAS_NVFP16()
+template <class _Tp>
+class __numeric_limits_impl<__half, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 11;
+  static constexpr int digits10     = 3;
+  static constexpr int max_digits10 = 5;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x0400u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7bffu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0xfbffu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x1400u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x3800u));
+  }
+
+  static constexpr int min_exponent   = -13;
+  static constexpr int min_exponent10 = -4;
+  static constexpr int max_exponent   = 16;
+  static constexpr int max_exponent10 = 4;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7c00u));
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7e00u));
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7d00u));
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x0001u));
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP16
+
+// nvbf16
+
+#if _CCCL_HAS_NVBF16()
+template <class _Tp>
+class __numeric_limits_impl<__nv_bfloat16, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 8;
+  static constexpr int digits10     = 2;
+  static constexpr int max_digits10 = 4;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x0080u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7f7fu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0xff7fu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x3c00u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x3f00u));
+  }
+
+  static constexpr int min_exponent   = -125;
+  static constexpr int min_exponent10 = -37;
+  static constexpr int max_exponent   = 128;
+  static constexpr int max_exponent10 = 38;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7f80u));
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7fc0u));
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x7fa0u));
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint16_t(0x0001u));
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVBF16
+
+// nvfp8_e4m3
+
+#if _CCCL_HAS_NVFP8_E4M3()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp8_e4m3, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 4;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 3;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x08u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7eu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0xfeu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x20u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x30u));
+  }
+
+  static constexpr int min_exponent   = -6;
+  static constexpr int min_exponent10 = -2;
+  static constexpr int max_exponent   = 8;
+  static constexpr int max_exponent10 = 2;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7fu));
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x01u));
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP8_E4M3
+
+// nvfp8_e5m2
+
+#if _CCCL_HAS_NVFP8_E5M2()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp8_e5m2, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 3;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 2;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x04u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7bu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0xfbu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x34u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x38u));
+  }
+
+  static constexpr int min_exponent   = -15;
+  static constexpr int min_exponent10 = -5;
+  static constexpr int max_exponent   = 15;
+  static constexpr int max_exponent10 = 4;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7cu));
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7eu));
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7du));
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x01u));
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP8_E5M2
+
+// nvfp8_e8m0
+
+#if _CCCL_HAS_NVFP8_E8M0()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp8_e8m0, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = false;
+  static constexpr int digits       = 1;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 2;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x00u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0xfeu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x00u));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7fu));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7fu));
+  }
+
+  static constexpr int min_exponent   = -127;
+  static constexpr int min_exponent10 = -39;
+  static constexpr int max_exponent   = 127;
+  static constexpr int max_exponent10 = 38;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_absent;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0xffu));
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return _Tp{};
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_toward_zero;
+};
+#endif // _CCCL_HAS_NVFP8_E8M0()
+
+// nvfp6_e2m3
+
+#if _CCCL_HAS_NVFP6_E2M3()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp6_e2m3, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 4;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 3;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x08u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x1fu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x3fu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x01u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x04u));
+  }
+
+  static constexpr int min_exponent   = 0;
+  static constexpr int min_exponent10 = 0;
+  static constexpr int max_exponent   = 2;
+  static constexpr int max_exponent10 = 0;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x01u));
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP6_E2M3
+
+// nvfp6_e3m2
+
+#if _CCCL_HAS_NVFP6_E3M2()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp6_e3m2, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 3;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 2;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x04u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x1fu));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x3fu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x04u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x08u));
+  }
+
+  static constexpr int min_exponent   = -2;
+  static constexpr int min_exponent10 = -1;
+  static constexpr int max_exponent   = 4;
+  static constexpr int max_exponent10 = 1;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x01u));
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP6_E3M2()
+
+// nvfp4_e2m1
+
+#if _CCCL_HAS_NVFP4_E2M1()
+template <class _Tp>
+class __numeric_limits_impl<__nv_fp4_e2m1, __numeric_limits_type::__floating_point, _Tp>
+{
+public:
+  using type = _Tp;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 2;
+  static constexpr int digits10     = 0;
+  static constexpr int max_digits10 = 2;
+  _CCCL_API static constexpr _Tp min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x2u));
+  }
+  _CCCL_API static constexpr _Tp max() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x7u));
+  }
+  _CCCL_API static constexpr _Tp lowest() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0xfu));
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr _Tp epsilon() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x1u));
+  }
+  _CCCL_API static constexpr _Tp round_error() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x1u));
+  }
+
+  static constexpr int min_exponent   = 0;
+  static constexpr int min_exponent10 = 0;
+  static constexpr int max_exponent   = 2;
+  static constexpr int max_exponent10 = 0;
+
+  static constexpr bool has_infinity                                       = false;
+  static constexpr bool has_quiet_NaN                                      = false;
+  static constexpr bool has_signaling_NaN                                  = false;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+  _CCCL_API static constexpr _Tp infinity() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp quiet_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp signaling_NaN() noexcept
+  {
+    return _Tp{};
+  }
+  _CCCL_API static constexpr _Tp denorm_min() noexcept
+  {
+    return ::cuda::std::__fp_from_storage<_Tp>(uint8_t(0x1u));
+  }
+
+  static constexpr bool is_iec559  = false;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_NVFP4_E2M1()
+
+// __float128
+
+#if _CCCL_HAS_FLOAT128()
+template <>
+class __numeric_limits_impl<__float128, __numeric_limits_type::__floating_point>
+{
+public:
+  using type = __float128;
+
+  static constexpr bool is_specialized = true;
+
+  static constexpr bool is_signed   = true;
+  static constexpr int digits       = 113;
+  static constexpr int digits10     = 33;
+  static constexpr int max_digits10 = 36;
+
+  _CCCL_API static constexpr type min() noexcept
+  {
+    return 3.36210314311209350626267781732175260e-4932q;
+  }
+  _CCCL_API static constexpr type max() noexcept
+  {
+    return 1.18973149535723176508575932662800702e+4932q;
+  }
+  _CCCL_API static constexpr type lowest() noexcept
+  {
+    return -max();
+  }
+
+  static constexpr bool is_integer = false;
+  static constexpr bool is_exact   = false;
+  static constexpr int radix       = FLT_RADIX;
+  _CCCL_API static constexpr type epsilon() noexcept
+  {
+    return 1.92592994438723585305597794258492732e-34q;
+  }
+  _CCCL_API static constexpr type round_error() noexcept
+  {
+    return 0.5q;
+  }
+
+  static constexpr int min_exponent   = -16381;
+  static constexpr int min_exponent10 = -4931;
+  static constexpr int max_exponent   = 16384;
+  static constexpr int max_exponent10 = 4932;
+
+  static constexpr bool has_infinity                                       = true;
+  static constexpr bool has_quiet_NaN                                      = true;
+  static constexpr bool has_signaling_NaN                                  = true;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr float_denorm_style has_denorm = denorm_present;
+  _CCCL_DEPRECATED_IN_CXX23 static constexpr bool has_denorm_loss          = false;
+
+#  if defined(_CCCL_BUILTIN_HUGE_VALF128)
+  _CCCL_API static constexpr type infinity() noexcept
+  {
+    return _CCCL_BUILTIN_HUGE_VALF128();
+  }
+#  else // ^^^ _CCCL_BUILTIN_HUGE_VALF128 ^^^ // vvv !_CCCL_BUILTIN_HUGE_VALF128 vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type infinity() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(__uint128_t{0x7fff'0000'0000'0000} << 64);
+  }
+#  endif // ^^^ !_CCCL_BUILTIN_HUGE_VALF128 ^^^
+#  if defined(_CCCL_BUILTIN_NANF128)
+  _CCCL_API static constexpr type quiet_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANF128("");
+  }
+#  else // ^^^ _CCCL_BUILTIN_NANF128 ^^^ // vvv !_CCCL_BUILTIN_NANF128 vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type quiet_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(__uint128_t{0x7fff'8000'0000'0000} << 64);
+  }
+#  endif // ^^^ !_CCCL_BUILTIN_NANF128 ^^^
+#  if defined(_CCCL_BUILTIN_NANSF128)
+  _CCCL_API static constexpr type signaling_NaN() noexcept
+  {
+    return _CCCL_BUILTIN_NANSF128("");
+  }
+#  else // ^^^ _CCCL_BUILTIN_NANSF128 ^^^ // vvv !_CCCL_BUILTIN_NANSF128 vvv
+  _CCCL_API inline static _CCCL_CONSTEXPR_BIT_CAST type signaling_NaN() noexcept
+  {
+    return ::cuda::std::bit_cast<type>(__uint128_t{0x7fff'4000'0000'0000} << 64);
+  }
+#  endif // ^^^ !_CCCL_BUILTIN_NANSF128 ^^^
+  _CCCL_API static constexpr type denorm_min() noexcept
+  {
+    return 6.47517511943802511092443895822764655e-4966q;
+  }
+
+  static constexpr bool is_iec559  = true;
+  static constexpr bool is_bounded = true;
+  static constexpr bool is_modulo  = false;
+
+  static constexpr bool traps                    = false;
+  static constexpr bool tinyness_before          = false;
+  static constexpr float_round_style round_style = round_to_nearest;
+};
+#endif // _CCCL_HAS_FLOAT128()
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___LIMITS_NUMERIC_LIMITS_EXT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/addressof.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/addressof.h
new file mode 100644
index 00000000..aa650837
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/addressof.h
@@ -0,0 +1,96 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___MEMORY_ADDRESSOF_H
+#define _CUDA_STD___MEMORY_ADDRESSOF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 12)
+#  define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 1
+#else // ^^^ has builtin std::addressof ^^^ / vvv no builtin std::addressof vvv
+#  define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 0
+#endif // ^^^ no builtin std::addressof ^^^
+
+// nvcc warns about host only std::addressof being used in device code
+#if _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+#  undef _CCCL_HAS_BUILTIN_STD_ADDRESSOF
+#  define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 0
+#endif // _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+
+// We cannot use host features if we are building in freestanding
+#if _CCCL_FREESTANDING()
+#  undef _CCCL_HAS_BUILTIN_STD_ADDRESSOF
+#  define _CCCL_HAS_BUILTIN_STD_ADDRESSOF() 0
+#endif // _CCCL_FREESTANDING()
+
+// include minimal std:: headers
+#if _CCCL_HAS_BUILTIN_STD_ADDRESSOF()
+#  if _CCCL_HOST_STD_LIB(LIBSTDCXX) && __has_include(<bits/move.h>)
+#    include <bits/move.h>
+#  elif _CCCL_HOST_STD_LIB(LIBCXX) && __has_include(<__memory/addressof.h>)
+#    include <__memory/addressof.h>
+#  else
+#    include <cuda/std/__host_stdlib/memory>
+#  endif
+#endif // _CCCL_HAS_BUILTIN_STD_ADDRESSOF()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4312) // warning C4312: 'type cast': conversion from '_Tp' to '_Tp *' of greater size
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_BUILTIN_STD_ADDRESSOF()
+
+// The compiler treats ::std::addressof as a builtin function so it does not need to be
+// instantiated and will be compiled away even at -O0.
+using ::std::addressof;
+
+#elif defined(_CCCL_BUILTIN_ADDRESSOF)
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API _CCCL_NO_CFI constexpr _Tp* addressof(_Tp& __x) noexcept
+{
+  return _CCCL_BUILTIN_ADDRESSOF(__x);
+}
+
+template <class _Tp>
+_Tp* addressof(const _Tp&&) noexcept = delete;
+
+#else
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API _CCCL_NO_CFI _Tp* addressof(_Tp& __x) noexcept
+{
+  return reinterpret_cast<_Tp*>(const_cast<char*>(&reinterpret_cast<const volatile char&>(__x)));
+}
+
+template <class _Tp>
+_Tp* addressof(const _Tp&&) noexcept = delete;
+
+#endif // defined(_CCCL_BUILTIN_ADDRESSOF)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_DIAG_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___MEMORY_ADDRESSOF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/allocator_arg_t.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/allocator_arg_t.h
new file mode 100644
index 00000000..5b45d828
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/allocator_arg_t.h
@@ -0,0 +1,85 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___FUNCTIONAL_ALLOCATOR_ARG_T_H
+#define _CUDA_STD___FUNCTIONAL_ALLOCATOR_ARG_T_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__memory/uses_allocator.h>
+#include <cuda/std/__new/device_new.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/forward.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct _CCCL_TYPE_VISIBILITY_DEFAULT allocator_arg_t
+{
+  _CCCL_HIDE_FROM_ABI explicit allocator_arg_t() = default;
+};
+
+inline constexpr allocator_arg_t allocator_arg = allocator_arg_t();
+
+// allocator construction
+
+template <class _Tp, class _Alloc, class... _Args>
+struct __uses_alloc_ctor_imp
+{
+  using _RawAlloc _CCCL_NODEBUG_ALIAS = remove_cvref_t<_Alloc>;
+  static const bool __ua              = uses_allocator<_Tp, _RawAlloc>::value;
+  static const bool __ic              = is_constructible_v<_Tp, allocator_arg_t, _Alloc, _Args...>;
+  static const int value              = __ua ? 2 - __ic : 0;
+};
+
+template <class _Tp, class _Alloc, class... _Args>
+struct __uses_alloc_ctor : integral_constant<int, __uses_alloc_ctor_imp<_Tp, _Alloc, _Args...>::value>
+{};
+
+template <class _Tp, class _Allocator, class... _Args>
+_CCCL_API inline void
+__user_alloc_construct_impl(integral_constant<int, 0>, _Tp* __storage, const _Allocator&, _Args&&... __args)
+{
+  new (__storage) _Tp(::cuda::std::forward<_Args>(__args)...);
+}
+
+// FIXME: This should have a version which takes a non-const alloc.
+template <class _Tp, class _Allocator, class... _Args>
+_CCCL_API inline void
+__user_alloc_construct_impl(integral_constant<int, 1>, _Tp* __storage, const _Allocator& __a, _Args&&... __args)
+{
+  new (__storage) _Tp(allocator_arg, __a, ::cuda::std::forward<_Args>(__args)...);
+}
+
+// FIXME: This should have a version which takes a non-const alloc.
+template <class _Tp, class _Allocator, class... _Args>
+_CCCL_API inline void
+__user_alloc_construct_impl(integral_constant<int, 2>, _Tp* __storage, const _Allocator& __a, _Args&&... __args)
+{
+  new (__storage) _Tp(::cuda::std::forward<_Args>(__args)..., __a);
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___FUNCTIONAL_ALLOCATOR_ARG_T_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/pointer_traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/pointer_traits.h
new file mode 100644
index 00000000..b999bb31
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/pointer_traits.h
@@ -0,0 +1,247 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___MEMORY_POINTER_TRAITS_H
+#define _CUDA_STD___MEMORY_POINTER_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__memory/addressof.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/conjunction.h>
+#include <cuda/std/__type_traits/decay.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_class.h>
+#include <cuda/std/__type_traits/is_function.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, class = void>
+inline constexpr bool __has_element_type = false;
+
+template <class _Tp>
+inline constexpr bool __has_element_type<_Tp, void_t<typename _Tp::element_type>> = true;
+
+template <class _Ptr, bool = __has_element_type<_Ptr>>
+struct __pointer_traits_element_type;
+
+template <class _Ptr>
+struct __pointer_traits_element_type<_Ptr, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = typename _Ptr::element_type;
+};
+
+template <template <class, class...> class _Sp, class _Tp, class... _Args>
+struct __pointer_traits_element_type<_Sp<_Tp, _Args...>, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = typename _Sp<_Tp, _Args...>::element_type;
+};
+
+template <template <class, class...> class _Sp, class _Tp, class... _Args>
+struct __pointer_traits_element_type<_Sp<_Tp, _Args...>, false>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+template <class _Tp, class = void>
+inline constexpr bool __has_difference_type = false;
+
+template <class _Tp>
+inline constexpr bool __has_difference_type<_Tp, void_t<typename _Tp::difference_type>> = true;
+
+template <class _Ptr, bool = __has_difference_type<_Ptr>>
+struct __pointer_traits_difference_type
+{
+  using type _CCCL_NODEBUG_ALIAS = ptrdiff_t;
+};
+
+template <class _Ptr>
+struct __pointer_traits_difference_type<_Ptr, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = typename _Ptr::difference_type;
+};
+
+template <class _Tp, class _Up, class = void>
+inline constexpr bool __has_rebind = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __has_rebind<_Tp, _Up, void_t<typename _Tp::template rebind<_Up>>> = true;
+
+template <class _Tp, class _Up, bool = __has_rebind<_Tp, _Up>>
+struct __pointer_traits_rebind
+{
+  using type _CCCL_NODEBUG_ALIAS = typename _Tp::template rebind<_Up>;
+};
+
+template <template <class, class...> class _Sp, class _Tp, class... _Args, class _Up>
+struct __pointer_traits_rebind<_Sp<_Tp, _Args...>, _Up, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = typename _Sp<_Tp, _Args...>::template rebind<_Up>;
+};
+
+template <template <class, class...> class _Sp, class _Tp, class... _Args, class _Up>
+struct __pointer_traits_rebind<_Sp<_Tp, _Args...>, _Up, false>
+{
+  using type = _Sp<_Up, _Args...>;
+};
+
+template <class _Ptr, class = void>
+struct __pointer_traits_impl
+{};
+
+template <class _Ptr>
+struct __pointer_traits_impl<_Ptr, void_t<typename __pointer_traits_element_type<_Ptr>::type>>
+{
+  using pointer         = _Ptr;
+  using element_type    = typename __pointer_traits_element_type<pointer>::type;
+  using difference_type = typename __pointer_traits_difference_type<pointer>::type;
+
+  template <class _Up>
+  using rebind = typename __pointer_traits_rebind<pointer, _Up>::type;
+
+private:
+  struct __nat
+  {};
+
+public:
+  _CCCL_API inline _CCCL_CONSTEXPR_CXX20 static pointer
+  pointer_to(conditional_t<is_void_v<element_type>, __nat, element_type>& __r)
+  {
+    return pointer::pointer_to(__r);
+  }
+};
+
+template <class _Ptr, class = void>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT pointer_traits : __pointer_traits_impl<_Ptr>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT pointer_traits<_Tp*>
+{
+  using pointer         = _Tp*;
+  using element_type    = _Tp;
+  using difference_type = ptrdiff_t;
+
+  template <class _Up>
+  using rebind = _Up*;
+
+private:
+  struct __nat
+  {};
+
+public:
+  _CCCL_API inline _CCCL_CONSTEXPR_CXX20 static pointer
+  pointer_to(conditional_t<is_void_v<element_type>, __nat, element_type>& __r) noexcept
+  {
+    return ::cuda::std::addressof(__r);
+  }
+};
+
+// to_address
+
+template <class _Pointer, class = void>
+struct __to_address_helper;
+
+template <class _Tp>
+_CCCL_API constexpr _Tp* __to_address(_Tp* __p) noexcept
+{
+  static_assert(!is_function_v<_Tp>, "_Tp is a function type");
+  return __p;
+}
+
+template <class _Pointer, class = void>
+inline constexpr bool __has_toaddress = false;
+
+template <class _Pointer>
+inline constexpr bool
+  __has_toaddress<_Pointer,
+                  decltype((void) pointer_traits<_Pointer>::to_address(::cuda::std::declval<const _Pointer&>()))> =
+    true;
+
+template <class _Pointer, class = void>
+inline constexpr bool __has_const_operator_arrow = false;
+
+template <class _Pointer>
+inline constexpr bool
+  __has_const_operator_arrow<_Pointer, decltype((void) ::cuda::std::declval<const _Pointer&>().operator->())> = true;
+
+template <class _Pointer>
+inline constexpr bool __is_fancy_pointer = __has_const_operator_arrow<_Pointer> || __has_toaddress<_Pointer>;
+
+// enable_if is needed here to avoid instantiating checks for fancy pointers on raw pointers
+template <class _Pointer, class = enable_if_t<is_class_v<_Pointer>>, class = enable_if_t<__is_fancy_pointer<_Pointer>>>
+_CCCL_API constexpr decay_t<decltype(__to_address_helper<_Pointer>::__call(::cuda::std::declval<const _Pointer&>()))>
+__to_address(const _Pointer& __p) noexcept
+{
+  return __to_address_helper<_Pointer>::__call(__p);
+}
+
+template <class _Pointer, class>
+struct __to_address_helper
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr static decltype(::cuda::std::__to_address(::cuda::std::declval<const _Pointer&>().operator->()))
+  __call(const _Pointer& __p) noexcept
+  {
+    return ::cuda::std::__to_address(__p.operator->());
+  }
+};
+
+template <class _Pointer>
+struct __to_address_helper<_Pointer,
+                           decltype((void) pointer_traits<_Pointer>::to_address(::cuda::std::declval<const _Pointer&>()))>
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr static decltype(pointer_traits<_Pointer>::to_address(::cuda::std::declval<const _Pointer&>()))
+  __call(const _Pointer& __p) noexcept
+  {
+    return pointer_traits<_Pointer>::to_address(__p);
+  }
+};
+
+template <class _Tp>
+_CCCL_API constexpr auto to_address(_Tp* __p) noexcept
+{
+  return ::cuda::std::__to_address(__p);
+}
+
+template <class _Pointer>
+_CCCL_API constexpr auto to_address(const _Pointer& __p) noexcept -> decltype(::cuda::std::__to_address(__p))
+{
+  return ::cuda::std::__to_address(__p);
+}
+
+template <class _Iter, class = void>
+inline constexpr bool __can_to_address = false;
+
+template <class _Iter>
+inline constexpr bool
+  __can_to_address<_Iter, void_t<decltype(::cuda::std::to_address(::cuda::std::declval<_Iter&>()))>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___MEMORY_POINTER_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/uses_allocator.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/uses_allocator.h
new file mode 100644
index 00000000..37125d0d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__memory/uses_allocator.h
@@ -0,0 +1,59 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___MEMORY_USES_ALLOCATOR_H
+#define _CUDA_STD___MEMORY_USES_ALLOCATOR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, class = void>
+inline constexpr bool __has_allocator_type_v = false;
+template <class _Tp>
+inline constexpr bool __has_allocator_type_v<_Tp, void_t<typename _Tp::allocator_type>> = true;
+
+template <class _Tp, class _Alloc, bool = __has_allocator_type_v<_Tp>>
+inline constexpr bool __uses_allocator_v = false;
+template <class _Tp, class _Alloc>
+inline constexpr bool __uses_allocator_v<_Tp, _Alloc, true> = is_convertible_v<_Alloc, typename _Tp::allocator_type>;
+
+template <class _Tp, class _Alloc>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT uses_allocator : public integral_constant<bool, __uses_allocator_v<_Tp, _Alloc>>
+{};
+
+template <class _Tp, class _Alloc>
+inline constexpr bool uses_allocator_v = __uses_allocator_v<_Tp, _Alloc>;
+
+template <class... _Tp, class _Alloc>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT uses_allocator<tuple<_Tp...>, _Alloc> : true_type
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___MEMORY_USES_ALLOCATOR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__new/device_new.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__new/device_new.h
new file mode 100644
index 00000000..d9e0edea
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__new/device_new.h
@@ -0,0 +1,30 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___NEW_DEVICE_NEW_H
+#define _CUDA_STD___NEW_DEVICE_NEW_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// clang-cuda only provides device flavors of operator new if we included <new>
+#if _CCCL_CUDA_COMPILER(CLANG)
+#  include <cuda/std/__host_stdlib/new>
+#endif // _CCCL_CUDA_COMPILER(CLANG)
+
+#endif // _CUDA_STD___NEW_DEVICE_NEW_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/access.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/access.h
new file mode 100644
index 00000000..c93d0bb2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/access.h
@@ -0,0 +1,318 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+#ifndef _CUDA_STD___RANGES_ACCESS_H
+#define _CUDA_STD___RANGES_ACCESS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__ranges/enable_borrowed_range.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__utility/auto_cast.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+template <class _Tp>
+_CCCL_CONCEPT __can_borrow = is_lvalue_reference_v<_Tp> || enable_borrowed_range<remove_cvref_t<_Tp>>;
+
+// [range.access.begin]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__begin)
+template <class _Tp>
+void begin(_Tp&) = delete;
+template <class _Tp>
+void begin(const _Tp&) = delete;
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __member_begin = __can_borrow<_Tp> && __workaround_52970<_Tp> && requires(_Tp&& __t) {
+  { _LIBCUDACXX_AUTO_CAST(__t.begin()) } -> input_or_output_iterator;
+};
+
+template <class _Tp>
+concept __unqualified_begin =
+  !__member_begin<_Tp> && __can_borrow<_Tp> && __class_or_enum<remove_cvref_t<_Tp>> && requires(_Tp&& __t) {
+    { _LIBCUDACXX_AUTO_CAST(begin(__t)) } -> input_or_output_iterator;
+  };
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __member_begin_,
+  requires(_Tp&& __t)(requires(__can_borrow<_Tp>),
+                      requires(__workaround_52970<_Tp>),
+                      requires(input_or_output_iterator<decltype(_LIBCUDACXX_AUTO_CAST(__t.begin()))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __member_begin = _CCCL_FRAGMENT(__member_begin_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __unqualified_begin_,
+  requires(_Tp&& __t)(requires(!__member_begin<_Tp>),
+                      requires(__can_borrow<_Tp>),
+                      requires(__class_or_enum<remove_cvref_t<_Tp>>),
+                      requires(input_or_output_iterator<decltype(_LIBCUDACXX_AUTO_CAST(begin(__t)))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __unqualified_begin = _CCCL_FRAGMENT(__unqualified_begin_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+struct __fn
+{
+  // This has been made valid as a defect report for C++17 onwards, however gcc below 11.0 does not implement it
+#if (!_CCCL_COMPILER(GCC) || _CCCL_COMPILER(GCC, >=, 11))
+  _CCCL_TEMPLATE(class _Tp)
+  // NOLINTNEXTLINE(bugprone-sizeof-expression)
+  _CCCL_REQUIRES((sizeof(_Tp) >= 0)) // Disallow incomplete element types.
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp (&__t)[]) const noexcept
+  {
+    return __t + 0;
+  }
+#endif // (!_CCCL_COMPILER(GCC) || _CCCL_COMPILER(GCC, >=, 11))
+
+  _CCCL_TEMPLATE(class _Tp, size_t _Np)
+  // NOLINTNEXTLINE(bugprone-sizeof-expression)
+  _CCCL_REQUIRES((sizeof(_Tp) >= 0)) // Disallow incomplete element types.
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp (&__t)[_Np]) const noexcept
+  {
+    return __t + 0;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__member_begin<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(__t.begin())))
+  {
+    return _LIBCUDACXX_AUTO_CAST(__t.begin());
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__unqualified_begin<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(begin(__t))))
+  {
+    return _LIBCUDACXX_AUTO_CAST(begin(__t));
+  }
+
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES((!__member_begin<_Tp>) _CCCL_AND(!__unqualified_begin<_Tp>))
+  void operator()(_Tp&&) const = delete;
+
+#if _CCCL_COMPILER(MSVC, <, 19, 23)
+  template <class _Tp>
+  void operator()(_Tp (&&)[]) const = delete;
+
+  template <class _Tp, size_t _Np>
+  void operator()(_Tp (&&)[_Np]) const = delete;
+#endif // _CCCL_COMPILER(MSVC, <, 19, 23)
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto begin = __begin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __begin_cpo = __begin::__fn;
+} // namespace __cpo
+
+// [range.range]
+
+template <class _Tp>
+using iterator_t = decltype(::cuda::std::ranges::__begin_cpo{}(::cuda::std::declval<_Tp&>()));
+
+// [range.access.end]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__end)
+template <class _Tp>
+void end(_Tp&) = delete;
+template <class _Tp>
+void end(const _Tp&) = delete;
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __member_end = __can_borrow<_Tp> && __workaround_52970<_Tp> && requires(_Tp&& __t) {
+  typename iterator_t<_Tp>;
+  { _LIBCUDACXX_AUTO_CAST(__t.end()) } -> sentinel_for<iterator_t<_Tp>>;
+};
+
+template <class _Tp>
+concept __unqualified_end =
+  !__member_end<_Tp> && __can_borrow<_Tp> && __class_or_enum<remove_cvref_t<_Tp>> && requires(_Tp&& __t) {
+    typename iterator_t<_Tp>;
+    { _LIBCUDACXX_AUTO_CAST(end(__t)) } -> sentinel_for<iterator_t<_Tp>>;
+  };
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __member_end_,
+  requires(_Tp&& __t)(requires(__can_borrow<_Tp>),
+                      requires(__workaround_52970<_Tp>),
+                      typename(iterator_t<_Tp>),
+                      requires(sentinel_for<decltype(_LIBCUDACXX_AUTO_CAST(__t.end())), iterator_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __member_end = _CCCL_FRAGMENT(__member_end_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __unqualified_end_,
+  requires(_Tp&& __t)(requires(!__member_end<_Tp>),
+                      requires(__can_borrow<_Tp>),
+                      requires(__class_or_enum<remove_cvref_t<_Tp>>),
+                      typename(iterator_t<_Tp>),
+                      requires(sentinel_for<decltype(_LIBCUDACXX_AUTO_CAST(end(__t))), iterator_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __unqualified_end = _CCCL_FRAGMENT(__unqualified_end_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+struct __fn
+{
+  _CCCL_TEMPLATE(class _Tp, size_t _Np)
+  // NOLINTNEXTLINE(bugprone-sizeof-expression)
+  _CCCL_REQUIRES((sizeof(_Tp) >= 0)) // Disallow incomplete element types.
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp (&__t)[_Np]) const noexcept
+  {
+    return __t + _Np;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__member_end<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(__t.end())))
+  {
+    return _LIBCUDACXX_AUTO_CAST(__t.end());
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__unqualified_end<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(end(__t))))
+  {
+    return _LIBCUDACXX_AUTO_CAST(end(__t));
+  }
+
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES((!__member_end<_Tp>) _CCCL_AND(!__unqualified_end<_Tp>))
+  void operator()(_Tp&&) const = delete;
+
+#if _CCCL_COMPILER(MSVC, <, 19, 23)
+  template <class _Tp>
+  void operator()(_Tp (&&)[]) const = delete;
+
+  template <class _Tp, size_t _Np>
+  void operator()(_Tp (&&)[_Np]) const = delete;
+#endif // _CCCL_COMPILER(MSVC, <, 19, 23)
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto end = __end::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __end_cpo = __end::__fn;
+} // namespace __cpo
+
+// [range.access.cbegin]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__cbegin)
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_lvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__begin_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t))))
+      -> decltype(::cuda::std::ranges::__begin_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t)))
+  {
+    return ::cuda::std::ranges::__begin_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_rvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__begin_cpo{}(static_cast<const _Tp&&>(__t))))
+      -> decltype(::cuda::std::ranges::__begin_cpo{}(static_cast<const _Tp&&>(__t)))
+  {
+    return ::cuda::std::ranges::__begin_cpo{}(static_cast<const _Tp&&>(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto cbegin = __cbegin::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __cbegin_cpo = __cbegin::__fn;
+} // namespace __cpo
+
+// [range.access.cend]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__cend)
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_lvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__end_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t))))
+      -> decltype(::cuda::std::ranges::__end_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t)))
+  {
+    return ::cuda::std::ranges::__end_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t));
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_rvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__end_cpo{}(static_cast<const _Tp&&>(__t))))
+      -> decltype(::cuda::std::ranges::__end_cpo{}(static_cast<const _Tp&&>(__t)))
+  {
+    return ::cuda::std::ranges::__end_cpo{}(static_cast<const _Tp&&>(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto cend = __cend::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __cend_cpo = __cend::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_ACCESS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/concepts.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/concepts.h
new file mode 100644
index 00000000..21c56374
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/concepts.h
@@ -0,0 +1,303 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+#ifndef _CUDA_STD___RANGES_CONCEPTS_H
+#define _CUDA_STD___RANGES_CONCEPTS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/constructible.h>
+#include <cuda/std/__concepts/convertible_to.h>
+#include <cuda/std/__concepts/movable.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/incrementable_traits.h>
+#include <cuda/std/__iterator/iter_move.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__iterator/readable_traits.h>
+#include <cuda/std/__ranges/access.h>
+#include <cuda/std/__ranges/data.h>
+#include <cuda/std/__ranges/enable_borrowed_range.h>
+#include <cuda/std/__ranges/enable_view.h>
+#include <cuda/std/__ranges/size.h>
+#include <cuda/std/__type_traits/add_pointer.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/initializer_list>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+#if _CCCL_HAS_CONCEPTS()
+
+// [range.range]
+
+template <class _Tp>
+concept range = requires(_Tp& __t) {
+  ::cuda::std::ranges::__begin_cpo{}(__t); // sometimes equality-preserving
+  ::cuda::std::ranges::__end_cpo{}(__t);
+};
+
+template <class _Tp>
+concept input_range = range<_Tp> && input_iterator<iterator_t<_Tp>>;
+
+template <class _Range>
+concept borrowed_range =
+  range<_Range> && (is_lvalue_reference_v<_Range> || enable_borrowed_range<remove_cvref_t<_Range>>);
+
+// `iterator_t` defined in <__ranges/access.h>
+
+template <range _Rp>
+using sentinel_t = decltype(::cuda::std::ranges::__end_cpo{}(::cuda::std::declval<_Rp&>()));
+
+template <range _Rp>
+using range_difference_t = iter_difference_t<iterator_t<_Rp>>;
+
+template <range _Rp>
+using range_value_t = iter_value_t<iterator_t<_Rp>>;
+
+template <range _Rp>
+using range_reference_t = iter_reference_t<iterator_t<_Rp>>;
+
+template <range _Rp>
+using range_rvalue_reference_t = iter_rvalue_reference_t<iterator_t<_Rp>>;
+
+template <range _Rp>
+using range_common_reference_t = iter_common_reference_t<iterator_t<_Rp>>;
+
+// [range.sized]
+template <class _Tp>
+concept sized_range = range<_Tp> && requires(_Tp& __t) { ::cuda::std::ranges::__size_cpo{}(__t); };
+
+template <sized_range _Rp>
+using range_size_t = decltype(::cuda::std::ranges::__size_cpo{}(::cuda::std::declval<_Rp&>()));
+
+// `disable_sized_range` defined in `<__ranges/size.h>`
+
+// [range.view], views
+
+// `enable_view` defined in <__ranges/enable_view.h>
+// `view_base` defined in <__ranges/enable_view.h>
+
+template <class _Tp>
+concept view = range<_Tp> && movable<_Tp> && enable_view<_Tp>;
+
+template <class _Range>
+concept __simple_view = view<_Range> && range<const _Range> && same_as<iterator_t<_Range>, iterator_t<const _Range>>
+                     && same_as<sentinel_t<_Range>, sentinel_t<const _Range>>;
+
+// [range.refinements], other range refinements
+template <class _Rp, class _Tp>
+concept output_range = range<_Rp> && output_iterator<iterator_t<_Rp>, _Tp>;
+
+template <class _Tp>
+concept forward_range = input_range<_Tp> && forward_iterator<iterator_t<_Tp>>;
+
+template <class _Tp>
+concept bidirectional_range = forward_range<_Tp> && bidirectional_iterator<iterator_t<_Tp>>;
+
+template <class _Tp>
+concept random_access_range = bidirectional_range<_Tp> && random_access_iterator<iterator_t<_Tp>>;
+
+template <class _Tp>
+concept contiguous_range = random_access_range<_Tp> && contiguous_iterator<iterator_t<_Tp>> && requires(_Tp& __t) {
+  { ::cuda::std::ranges::__data_cpo{}(__t) } -> same_as<add_pointer_t<range_reference_t<_Tp>>>;
+};
+
+template <class _Tp>
+concept common_range = range<_Tp> && same_as<iterator_t<_Tp>, sentinel_t<_Tp>>;
+
+template <class _Tp>
+concept viewable_range =
+  range<_Tp>
+  && ((view<remove_cvref_t<_Tp>> && constructible_from<remove_cvref_t<_Tp>, _Tp>)
+      || (!view<remove_cvref_t<_Tp>>
+          && (is_lvalue_reference_v<_Tp>
+              || (movable<remove_reference_t<_Tp>> && !__is_cuda_std_initializer_list<remove_cvref_t<_Tp>>) )));
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+// [range.range]
+
+// clang-format off
+template <class _Tp>
+_CCCL_CONCEPT range =
+  _CCCL_REQUIRES_EXPR((_Tp), _Tp& __t)
+  (
+    void(::cuda::std::ranges::__begin_cpo{}(__t)),
+    void(::cuda::std::ranges::__end_cpo{}(__t))
+  );
+
+template <class _Tp>
+_CCCL_CONCEPT input_range =
+  _CCCL_REQUIRES_EXPR((_Tp))
+  (
+    requires(range<_Tp>),
+    requires(input_iterator<iterator_t<_Tp>>)
+  );
+// clang-format on
+
+template <class _Range>
+_CCCL_CONCEPT_FRAGMENT(
+  __borrowed_range_,
+  requires()(requires(range<_Range>),
+             requires((is_lvalue_reference_v<_Range> || enable_borrowed_range<remove_cvref_t<_Range>>) )));
+
+template <class _Range>
+_CCCL_CONCEPT borrowed_range = _CCCL_FRAGMENT(__borrowed_range_, _Range);
+
+// `iterator_t` defined in <__ranges/access.h>
+
+template <class _Rp>
+using sentinel_t = enable_if_t<range<_Rp>, decltype(::cuda::std::ranges::__end_cpo{}(::cuda::std::declval<_Rp&>()))>;
+
+template <class _Rp>
+using range_difference_t = enable_if_t<range<_Rp>, iter_difference_t<iterator_t<_Rp>>>;
+
+template <class _Rp>
+using range_value_t = enable_if_t<range<_Rp>, iter_value_t<iterator_t<_Rp>>>;
+
+template <class _Rp>
+using range_reference_t = enable_if_t<range<_Rp>, iter_reference_t<iterator_t<_Rp>>>;
+
+template <class _Rp>
+using range_rvalue_reference_t = enable_if_t<range<_Rp>, iter_rvalue_reference_t<iterator_t<_Rp>>>;
+
+template <class _Rp>
+using range_common_reference_t = enable_if_t<range<_Rp>, iter_common_reference_t<iterator_t<_Rp>>>;
+
+// [range.sized]
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __sized_range_, requires(_Tp& __t)(requires(range<_Tp>), typename(decltype(::cuda::std::ranges::__size_cpo{}(__t)))));
+
+template <class _Tp>
+_CCCL_CONCEPT sized_range = _CCCL_FRAGMENT(__sized_range_, _Tp);
+
+template <class _Rp>
+using range_size_t =
+  enable_if_t<sized_range<_Rp>, decltype(::cuda::std::ranges::__size_cpo{}(::cuda::std::declval<_Rp&>()))>;
+
+// `disable_sized_range` defined in `<__ranges/size.h>`
+
+// [range.view], views
+
+// `enable_view` defined in <__ranges/enable_view.h>
+// `view_base` defined in <__ranges/enable_view.h>
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__view_, requires()(requires(range<_Tp>), requires(movable<_Tp>), requires(enable_view<_Tp>)));
+
+template <class _Tp>
+_CCCL_CONCEPT view = _CCCL_FRAGMENT(__view_, _Tp);
+
+template <class _Range>
+_CCCL_CONCEPT_FRAGMENT(
+  __simple_view_,
+  requires()(requires(view<_Range>),
+             requires(range<const _Range>),
+             requires(same_as<iterator_t<_Range>, iterator_t<const _Range>>),
+             requires(same_as<sentinel_t<_Range>, sentinel_t<const _Range>>)));
+
+template <class _Range>
+_CCCL_CONCEPT __simple_view = _CCCL_FRAGMENT(__simple_view_, _Range);
+
+// [range.refinements], other range refinements
+template <class _Rp, class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__output_range_,
+                       requires()(requires(range<_Rp>), requires(output_iterator<iterator_t<_Rp>, _Tp>)));
+
+template <class _Rp, class _Tp>
+_CCCL_CONCEPT output_range = _CCCL_FRAGMENT(__output_range_, _Rp, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__forward_range_,
+                       requires()(requires(input_range<_Tp>), requires(forward_iterator<iterator_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT forward_range = _CCCL_FRAGMENT(__forward_range_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__bidirectional_range_,
+                       requires()(requires(forward_range<_Tp>), requires(bidirectional_iterator<iterator_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT bidirectional_range = _CCCL_FRAGMENT(__bidirectional_range_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __random_access_range_,
+  requires()(requires(bidirectional_range<_Tp>), requires(random_access_iterator<iterator_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT random_access_range = _CCCL_FRAGMENT(__random_access_range_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __contiguous_range_,
+  requires(_Tp& __t)(
+    requires(random_access_range<_Tp>),
+    requires(contiguous_iterator<iterator_t<_Tp>>),
+    requires(same_as<decltype(::cuda::std::ranges::__data_cpo{}(__t)), add_pointer_t<range_reference_t<_Tp>>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT contiguous_range = _CCCL_FRAGMENT(__contiguous_range_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__common_range_,
+                       requires()(requires(range<_Tp>), requires(same_as<iterator_t<_Tp>, sentinel_t<_Tp>>)));
+
+template <class _Tp>
+_CCCL_CONCEPT common_range = _CCCL_FRAGMENT(__common_range_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __viewable_range_,
+  requires()(requires(range<_Tp>),
+             requires(((view<remove_cvref_t<_Tp>> && constructible_from<remove_cvref_t<_Tp>, _Tp>)
+                       || (!view<remove_cvref_t<_Tp>>
+                           && (is_lvalue_reference_v<_Tp>
+                               || (movable<remove_reference_t<_Tp>>
+                                   && !__is_cuda_std_initializer_list<remove_cvref_t<_Tp>>) ))))));
+
+template <class _Tp>
+_CCCL_CONCEPT viewable_range = _CCCL_FRAGMENT(__viewable_range_, _Tp);
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+//[container.intro.reqmts]
+#if _CCCL_HAS_CONCEPTS()
+template <class _Range, class _Tp>
+concept __container_compatible_range = input_range<_Range> && convertible_to<range_reference_t<_Range>, _Tp>;
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Range, class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __container_compatible_range_,
+  requires()(requires(input_range<_Range>), requires(convertible_to<range_reference_t<_Range>, _Tp>)));
+
+template <class _Range, class _Tp>
+_CCCL_CONCEPT __container_compatible_range = _CCCL_FRAGMENT(__container_compatible_range_, _Range, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_CONCEPTS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/data.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/data.h
new file mode 100644
index 00000000..4802d62f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/data.h
@@ -0,0 +1,143 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+#ifndef _CUDA_STD___RANGES_DATA_H
+#define _CUDA_STD___RANGES_DATA_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/reverse_iterator.h>
+#include <cuda/std/__memory/pointer_traits.h>
+#include <cuda/std/__ranges/access.h>
+#include <cuda/std/__type_traits/is_object.h>
+#include <cuda/std/__type_traits/is_pointer.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_pointer.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__utility/auto_cast.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+// [range.prim.data]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__data)
+
+template <class _Tp>
+_CCCL_CONCEPT __ptr_to_object = is_pointer_v<_Tp> && is_object_v<remove_pointer_t<_Tp>>;
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __member_data = __can_borrow<_Tp> && __workaround_52970<_Tp> && requires(_Tp&& __t) {
+  { _LIBCUDACXX_AUTO_CAST(__t.data()) } -> __ptr_to_object;
+};
+
+template <class _Tp>
+concept __ranges_begin_invocable = !__member_data<_Tp> && __can_borrow<_Tp> && requires(_Tp&& __t) {
+  { ::cuda::std::ranges::__begin_cpo{}(__t) } -> contiguous_iterator;
+};
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__member_data_,
+                       requires(_Tp&& __t)(requires(__can_borrow<_Tp>),
+                                           requires(__workaround_52970<_Tp>),
+                                           requires(__ptr_to_object<decltype(_LIBCUDACXX_AUTO_CAST(__t.data()))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __member_data = _CCCL_FRAGMENT(__member_data_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __ranges_begin_invocable_,
+  requires(_Tp&& __t)(requires(!__member_data<_Tp>),
+                      requires(__can_borrow<_Tp>),
+                      requires(contiguous_iterator<decltype(::cuda::std::ranges::__begin_cpo{}(__t))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __ranges_begin_invocable = _CCCL_FRAGMENT(__ranges_begin_invocable_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+struct __fn
+{
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__member_data<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const noexcept(noexcept(__t.data()))
+  {
+    return __t.data();
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__ranges_begin_invocable<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::to_address(::cuda::std::ranges::__begin_cpo{}(__t))))
+  {
+    return ::cuda::std::to_address(::cuda::std::ranges::__begin_cpo{}(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto data = __data::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __data_cpo = __data::__fn;
+} // namespace __cpo
+
+// [range.prim.cdata]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__cdata)
+struct __fn
+{
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_lvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__data_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t))))
+      -> decltype(::cuda::std::ranges::__data_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t)))
+  {
+    return ::cuda::std::ranges::__data_cpo{}(static_cast<const remove_reference_t<_Tp>&>(__t));
+  }
+
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(is_rvalue_reference_v<_Tp&&>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__data_cpo{}(static_cast<const _Tp&&>(__t))))
+      -> decltype(::cuda::std::ranges::__data_cpo{}(static_cast<const _Tp&&>(__t)))
+  {
+    return ::cuda::std::ranges::__data_cpo{}(static_cast<const _Tp&&>(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto cdata = __cdata::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __cdata_cpo = __cdata::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_DATA_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_borrowed_range.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_borrowed_range.h
new file mode 100644
index 00000000..7abb492c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_borrowed_range.h
@@ -0,0 +1,41 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___RANGES_ENABLE_BORROWED_RANGE_H
+#define _CUDA_STD___RANGES_ENABLE_BORROWED_RANGE_H
+
+// These customization variables are used in <span> and <string_view>. The
+// separate header is used to avoid including the entire <ranges> header in
+// <span> and <string_view>.
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+// [range.range], ranges
+
+template <class>
+inline constexpr bool enable_borrowed_range = false;
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_ENABLE_BORROWED_RANGE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_view.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_view.h
new file mode 100644
index 00000000..27585bd5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/enable_view.h
@@ -0,0 +1,78 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___RANGES_ENABLE_VIEW_H
+#define _CUDA_STD___RANGES_ENABLE_VIEW_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/derived_from.h>
+#include <cuda/std/__concepts/same_as.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_class.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/void_t.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+struct view_base
+{};
+
+#if _CCCL_HAS_CONCEPTS()
+
+template <class _Derived>
+  requires is_class_v<_Derived> && same_as<_Derived, remove_cv_t<_Derived>>
+class view_interface;
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Derived, enable_if_t<is_class_v<_Derived> && same_as<_Derived, remove_cv_t<_Derived>>, int> = 0>
+class view_interface;
+
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_TEMPLATE(class _Op, class _Yp)
+_CCCL_REQUIRES(is_convertible_v<_Op*, view_interface<_Yp>*>)
+_CCCL_API inline void __is_derived_from_view_interface(const _Op*, const view_interface<_Yp>*);
+
+#if _CCCL_HAS_CONCEPTS()
+
+template <class _Tp>
+inline constexpr bool enable_view = derived_from<_Tp, view_base> || requires {
+  ::cuda::std::ranges::__is_derived_from_view_interface((_Tp*) nullptr, (_Tp*) nullptr);
+};
+
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+
+template <class _Tp, class = void>
+inline constexpr bool enable_view = derived_from<_Tp, view_base>;
+
+template <class _Tp>
+inline constexpr bool
+  enable_view<_Tp,
+              void_t<decltype(::cuda::std::ranges::__is_derived_from_view_interface((_Tp*) nullptr, (_Tp*) nullptr))>> =
+    true;
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_ENABLE_VIEW_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/size.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/size.h
new file mode 100644
index 00000000..d09a2703
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__ranges/size.h
@@ -0,0 +1,212 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+#ifndef _CUDA_STD___RANGES_SIZE_H
+#define _CUDA_STD___RANGES_SIZE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/arithmetic.h>
+#include <cuda/std/__concepts/class_or_enum.h>
+#include <cuda/std/__iterator/concepts.h>
+#include <cuda/std/__iterator/iterator_traits.h>
+#include <cuda/std/__memory/pointer_traits.h>
+#include <cuda/std/__ranges/access.h>
+#include <cuda/std/__type_traits/is_unbounded_array.h>
+#include <cuda/std/__type_traits/make_signed.h>
+#include <cuda/std/__type_traits/make_unsigned.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/auto_cast.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD_RANGES
+
+template <class>
+inline constexpr bool disable_sized_range = false;
+
+// [range.prim.size]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__size)
+template <class _Tp>
+void size(_Tp&) = delete;
+template <class _Tp>
+void size(const _Tp&) = delete;
+
+template <class _Tp>
+_CCCL_CONCEPT __size_enabled = !disable_sized_range<remove_cvref_t<_Tp>>;
+
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __member_size = __size_enabled<_Tp> && __workaround_52970<_Tp> && requires(_Tp&& __t) {
+  { _LIBCUDACXX_AUTO_CAST(__t.size()) } -> __integer_like;
+};
+
+template <class _Tp>
+concept __unqualified_size =
+  __size_enabled<_Tp> && !__member_size<_Tp> && __class_or_enum<remove_cvref_t<_Tp>> && requires(_Tp&& __t) {
+    { _LIBCUDACXX_AUTO_CAST(size(__t)) } -> __integer_like;
+  };
+
+template <class _Tp>
+concept __difference =
+  !__member_size<_Tp> && !__unqualified_size<_Tp> && __class_or_enum<remove_cvref_t<_Tp>> && requires(_Tp&& __t) {
+    { ::cuda::std::ranges::__begin_cpo{}(__t) } -> forward_iterator;
+    {
+      ::cuda::std::ranges::__end_cpo{}(__t)
+    } -> sized_sentinel_for<decltype(::cuda::std::ranges::__begin_cpo{}(::cuda::std::declval<_Tp>()))>;
+  };
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(__member_size_,
+                       requires(_Tp&& __t)(requires(__size_enabled<_Tp>),
+                                           requires(__workaround_52970<_Tp>),
+                                           requires(__integer_like<decltype(_LIBCUDACXX_AUTO_CAST(__t.size()))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __member_size = _CCCL_FRAGMENT(__member_size_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __unqualified_size_,
+  requires(_Tp&& __t)(requires(__size_enabled<_Tp>),
+                      requires(!__member_size<_Tp>),
+                      requires(__class_or_enum<remove_cvref_t<_Tp>>),
+                      requires(__integer_like<decltype(_LIBCUDACXX_AUTO_CAST(size(__t)))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __unqualified_size = _CCCL_FRAGMENT(__unqualified_size_, _Tp);
+
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __difference_,
+  requires(_Tp&& __t)(
+    requires(!__member_size<_Tp>),
+    requires(!__unqualified_size<_Tp>),
+    requires(__class_or_enum<remove_cvref_t<_Tp>>),
+    requires(forward_iterator<decltype(::cuda::std::ranges::__begin_cpo{}(__t))>),
+    requires(sized_sentinel_for<decltype(::cuda::std::ranges::__end_cpo{}(__t)),
+                                decltype(::cuda::std::ranges::__begin_cpo{}(::cuda::std::declval<_Tp>()))>)));
+
+template <class _Tp>
+_CCCL_CONCEPT __difference = _CCCL_FRAGMENT(__difference_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+struct __fn
+{
+  // `[range.prim.size]`: the array case (for rvalues).
+  template <class _Tp, size_t _Sz>
+  [[nodiscard]] _CCCL_API constexpr size_t operator()(_Tp (&&)[_Sz]) const noexcept
+  {
+    return _Sz;
+  }
+
+  // `[range.prim.size]`: the array case (for lvalues).
+  template <class _Tp, size_t _Sz>
+  [[nodiscard]] _CCCL_API constexpr size_t operator()(_Tp (&)[_Sz]) const noexcept
+  {
+    return _Sz;
+  }
+
+  // `[range.prim.size]`: `auto(t.size())` is a valid expression.
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__member_size<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(__t.size())))
+  {
+    return _LIBCUDACXX_AUTO_CAST(__t.size());
+  }
+
+  // `[range.prim.size]`: `auto(size(t))` is a valid expression.
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__unqualified_size<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(_LIBCUDACXX_AUTO_CAST(size(__t))))
+  {
+    return _LIBCUDACXX_AUTO_CAST(size(__t));
+  }
+
+  // [range.prim.size]: the `to-unsigned-like` case.
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__difference<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const noexcept(noexcept(
+    ::cuda::std::__to_unsigned_like(::cuda::std::ranges::__end_cpo{}(__t) - ::cuda::std::ranges::__begin_cpo{}(__t))))
+    -> decltype(::cuda::std::__to_unsigned_like(::cuda::std::ranges::__end_cpo{}(__t)
+                                                - ::cuda::std::ranges::__begin_cpo{}(__t)))
+  {
+    return ::cuda::std::__to_unsigned_like(
+      ::cuda::std::ranges::__end_cpo{}(__t) - ::cuda::std::ranges::__begin_cpo{}(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto size = __size::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __size_cpo = __size::__fn;
+} // namespace __cpo
+
+// [range.prim.ssize]
+
+_CCCL_BEGIN_NAMESPACE_CPO(__ssize)
+#if _CCCL_HAS_CONCEPTS()
+template <class _Tp>
+concept __can_ssize = requires(_Tp&& __t) { ::cuda::std::ranges::__size_cpo{}(__t); };
+#else // ^^^ _CCCL_HAS_CONCEPTS() ^^^ / vvv !_CCCL_HAS_CONCEPTS() vvv
+template <class _Tp>
+_CCCL_CONCEPT_FRAGMENT(
+  __can_ssize_,
+  requires(_Tp&& __t)(requires(!is_unbounded_array_v<_Tp>), ((void) ::cuda::std::ranges::__size_cpo{}(__t))));
+
+template <class _Tp>
+_CCCL_CONCEPT __can_ssize = _CCCL_FRAGMENT(__can_ssize_, _Tp);
+#endif // ^^^ !_CCCL_HAS_CONCEPTS() ^^^
+
+struct __fn
+{
+  _CCCL_TEMPLATE(class _Tp)
+  _CCCL_REQUIRES(__can_ssize<_Tp>)
+  [[nodiscard]] _CCCL_API constexpr auto operator()(_Tp&& __t) const
+    noexcept(noexcept(::cuda::std::ranges::__size_cpo{}(__t)))
+  {
+    using _Signed = make_signed_t<decltype(::cuda::std::ranges::__size_cpo{}(__t))>;
+    using _Result = conditional_t<(sizeof(ptrdiff_t) > sizeof(_Signed)), ptrdiff_t, _Signed>;
+    return static_cast<_Result>(::cuda::std::ranges::__size_cpo{}(__t));
+  }
+};
+_CCCL_END_NAMESPACE_CPO
+
+inline namespace __cpo
+{
+_CCCL_GLOBAL_CONSTANT auto ssize = __ssize::__fn{};
+
+// We want to avoid using the CPO internally because of __tile__ access
+using __ssize_cpo = __ssize::__fn;
+} // namespace __cpo
+
+_CCCL_END_NAMESPACE_CUDA_STD_RANGES
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___RANGES_SIZE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/make_tuple_types.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/make_tuple_types.h
new file mode 100644
index 00000000..48c0d60b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/make_tuple_types.h
@@ -0,0 +1,143 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_MAKE_TUPLE_TYPES_H
+#define _CUDA_STD___TUPLE_MAKE_TUPLE_TYPES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__fwd/complex.h>
+#include <cuda/std/__fwd/array.h>
+#include <cuda/std/__fwd/complex.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_indices.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__tuple_dir/tuple_types.h>
+#include <cuda/std/__type_traits/copy_cvref.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__type_traits/type_list.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// __make_tuple_types<_Tuple<_Types...>, _Ep, _Sp>::type is a
+// __tuple_types<_Types...> using only those _Types in the range [_Sp, _Ep).
+// _Sp defaults to 0 and _Ep defaults to tuple_size<_Tuple>.  If _Tuple is a
+// lvalue_reference type, then __tuple_types<_Types&...> is the result.
+
+template <class _TupleTypes, class _TupleIndices>
+struct __make_tuple_types_flat;
+
+template <template <class...> class _Tuple, class... _Types, size_t... _Idx>
+struct __make_tuple_types_flat<_Tuple<_Types...>, __tuple_indices<_Idx...>>
+{
+  using __tuple_types_list = __type_list<_Types...>;
+
+  // Specialization for pair, tuple, and __tuple_types
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS =
+    __tuple_types<__type_call<_ApplyFn, __type_at_c<_Idx, __tuple_types_list>>...>;
+};
+
+template <class _Vt, size_t _Np, size_t... _Idx>
+struct __make_tuple_types_flat<array<_Vt, _Np>, __tuple_indices<_Idx...>>
+{
+  template <size_t>
+  using __value_type _CCCL_NODEBUG_ALIAS = _Vt;
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS = __tuple_types<__type_call<_ApplyFn, __value_type<_Idx>>...>;
+};
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Vt, size_t _Np, size_t... _Idx>
+struct __make_tuple_types_flat<::std::array<_Vt, _Np>, __tuple_indices<_Idx...>>
+{
+  template <size_t>
+  using __value_type _CCCL_NODEBUG_ALIAS = _Vt;
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS = __tuple_types<__type_call<_ApplyFn, __value_type<_Idx>>...>;
+};
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Vt, size_t... _Idx>
+struct __make_tuple_types_flat<complex<_Vt>, __tuple_indices<_Idx...>>
+{
+  static_assert(sizeof...(_Idx) == 2, "__make_tuple_types: complex has only 2 members");
+  template <size_t>
+  using __value_type _CCCL_NODEBUG_ALIAS = _Vt;
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS = __tuple_types<__type_call<_ApplyFn, __value_type<_Idx>>...>;
+};
+
+template <class _Vt, size_t... _Idx>
+struct __make_tuple_types_flat<::cuda::complex<_Vt>, __tuple_indices<_Idx...>>
+{
+  static_assert(sizeof...(_Idx) == 2, "__make_tuple_types: complex has only 2 members");
+  template <size_t>
+  using __value_type _CCCL_NODEBUG_ALIAS = _Vt;
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS = __tuple_types<__type_call<_ApplyFn, __value_type<_Idx>>...>;
+};
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Vt, size_t... _Idx>
+struct __make_tuple_types_flat<::std::complex<_Vt>, __tuple_indices<_Idx...>>
+{
+  static_assert(sizeof...(_Idx) == 2, "__make_tuple_types: complex has only 2 members");
+  template <size_t>
+  using __value_type _CCCL_NODEBUG_ALIAS = _Vt;
+  template <class _Tp, class _ApplyFn = __apply_cvref_fn<_Tp>>
+  using __apply_quals _CCCL_NODEBUG_ALIAS = __tuple_types<__type_call<_ApplyFn, __value_type<_Idx>>...>;
+};
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Tp,
+          size_t _Ep     = tuple_size<remove_reference_t<_Tp>>::value,
+          size_t _Sp     = 0,
+          bool _SameSize = (_Ep == tuple_size<remove_reference_t<_Tp>>::value)>
+struct __make_tuple_types
+{
+  static_assert(_Sp <= _Ep, "__make_tuple_types input error");
+  using _RawTp _CCCL_NODEBUG_ALIAS = remove_cv_t<remove_reference_t<_Tp>>;
+  using _Maker _CCCL_NODEBUG_ALIAS = __make_tuple_types_flat<_RawTp, __make_tuple_indices_t<_Ep, _Sp>>;
+  using type _CCCL_NODEBUG_ALIAS   = typename _Maker::template __apply_quals<_Tp>;
+};
+
+template <class... _Types, size_t _Ep>
+struct __make_tuple_types<tuple<_Types...>, _Ep, 0, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = __tuple_types<_Types...>;
+};
+
+template <class... _Types, size_t _Ep>
+struct __make_tuple_types<__tuple_types<_Types...>, _Ep, 0, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = __tuple_types<_Types...>;
+};
+
+template <class _Tp, size_t _Ep = tuple_size<remove_reference_t<_Tp>>::value, size_t _Sp = 0>
+using __make_tuple_types_t = typename __make_tuple_types<_Tp, _Ep, _Sp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_MAKE_TUPLE_TYPES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/sfinae_helpers.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/sfinae_helpers.h
new file mode 100644
index 00000000..ac224bde
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/sfinae_helpers.h
@@ -0,0 +1,94 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_SFINAE_HELPERS_H
+#define _CUDA_STD___TUPLE_SFINAE_HELPERS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__type_traits/is_copy_assignable.h>
+#include <cuda/std/__type_traits/is_move_assignable.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+enum class __smf_availability
+{
+  __trivial,
+  __available,
+  __deleted,
+};
+
+// We need to synthesize the copy / move assignment if it would be implicitly deleted as a member of a class
+// In that case _Tp would be copy assignable but _TestSynthesizeAssignment<_Tp> would not
+// This happens e.g for reference types
+template <class _Tp>
+struct _TestSynthesizeAssignment
+{
+  _Tp __dummy;
+};
+
+template <class _Tp>
+inline constexpr bool __must_synthesize_assignment_v =
+  (is_copy_assignable_v<_Tp> && !is_copy_assignable_v<_TestSynthesizeAssignment<_Tp>>)
+  || (is_move_assignable_v<_Tp> && !is_move_assignable_v<_TestSynthesizeAssignment<_Tp>>);
+
+// We need to ensure that __tuple_impl_sfinae_helper is unique for every instantiation of __tuple_impl, so its templated
+// on the impl
+template <class _Impl, bool _AllCopyAssignable, bool _AllMoveAssignable>
+struct _CCCL_DECLSPEC_EMPTY_BASES __tuple_impl_sfinae_helper
+{};
+
+template <class _Impl>
+struct _CCCL_DECLSPEC_EMPTY_BASES __tuple_impl_sfinae_helper<_Impl, false, true>
+{
+  __tuple_impl_sfinae_helper()                                             = default;
+  __tuple_impl_sfinae_helper(const __tuple_impl_sfinae_helper&)            = default;
+  __tuple_impl_sfinae_helper(__tuple_impl_sfinae_helper&&)                 = default;
+  __tuple_impl_sfinae_helper& operator=(const __tuple_impl_sfinae_helper&) = delete;
+  __tuple_impl_sfinae_helper& operator=(__tuple_impl_sfinae_helper&&)      = default;
+};
+
+template <class _Impl>
+struct _CCCL_DECLSPEC_EMPTY_BASES __tuple_impl_sfinae_helper<_Impl, true, false>
+{
+  __tuple_impl_sfinae_helper()                                             = default;
+  __tuple_impl_sfinae_helper(const __tuple_impl_sfinae_helper&)            = default;
+  __tuple_impl_sfinae_helper(__tuple_impl_sfinae_helper&&)                 = default;
+  __tuple_impl_sfinae_helper& operator=(const __tuple_impl_sfinae_helper&) = default;
+  __tuple_impl_sfinae_helper& operator=(__tuple_impl_sfinae_helper&&)      = delete;
+};
+
+template <class _Impl>
+struct _CCCL_DECLSPEC_EMPTY_BASES __tuple_impl_sfinae_helper<_Impl, false, false>
+{
+  __tuple_impl_sfinae_helper()                                             = default;
+  __tuple_impl_sfinae_helper(const __tuple_impl_sfinae_helper&)            = default;
+  __tuple_impl_sfinae_helper(__tuple_impl_sfinae_helper&&)                 = default;
+  __tuple_impl_sfinae_helper& operator=(const __tuple_impl_sfinae_helper&) = delete;
+  __tuple_impl_sfinae_helper& operator=(__tuple_impl_sfinae_helper&&)      = delete;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_SFINAE_HELPERS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_constraints.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_constraints.h
new file mode 100644
index 00000000..df39b439
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_constraints.h
@@ -0,0 +1,643 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_TUPLE_CONSTRAINTS_H
+#define _CUDA_STD___TUPLE_TUPLE_CONSTRAINTS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/get.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__fwd/subrange.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__memory/allocator_arg_t.h>
+#include <cuda/std/__tuple_dir/make_tuple_types.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_indices.h>
+#include <cuda/std/__tuple_dir/tuple_like.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__tuple_dir/tuple_types.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/disjunction.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+#include <cuda/std/__type_traits/is_comparable.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_copy_assignable.h>
+#include <cuda/std/__type_traits/is_copy_constructible.h>
+#include <cuda/std/__type_traits/is_default_constructible.h>
+#include <cuda/std/__type_traits/is_implicitly_default_constructible.h>
+#include <cuda/std/__type_traits/is_move_assignable.h>
+#include <cuda/std/__type_traits/is_move_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_default_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/lazy.h>
+#include <cuda/std/__type_traits/reference_constructs_from_temporary.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__type_traits/sfinae_traits.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class... _Types>
+inline constexpr bool __tuple_all_nothrow_copy_constructible_v = (is_nothrow_copy_constructible_v<_Types> && ...);
+
+template <class... _Types>
+inline constexpr bool __tuple_all_nothrow_move_constructible_v = (is_nothrow_move_constructible_v<_Types> && ...);
+
+template <class, class>
+inline constexpr bool __tuple_all_nothrow_constructible_v = false;
+
+template <class... _Types, class... _UTypes>
+inline constexpr bool __tuple_all_nothrow_constructible_v<__tuple_types<_Types...>, __tuple_types<_UTypes...>> =
+  (is_nothrow_constructible_v<_Types, _UTypes> && ...);
+
+template <class... _Types>
+inline constexpr bool __tuple_all_copy_assignable_v = (is_copy_assignable_v<_Types> && ...);
+
+template <class... _Types>
+inline constexpr bool __tuple_all_move_assignable_v = (is_move_assignable_v<_Types> && ...);
+
+// __tuple_like_with_size
+template <class _Tuple, size_t _ExpectedSize, bool = __tuple_like<_Tuple>>
+inline constexpr bool __tuple_like_with_size = false;
+
+template <class _Tuple, size_t _ExpectedSize>
+inline constexpr bool __tuple_like_with_size<_Tuple, _ExpectedSize, true> =
+  _ExpectedSize == tuple_size<remove_cvref_t<_Tuple>>::value;
+
+template <class... _Types>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_default_constructible(__tuple_types<_Types...>) noexcept
+{
+  if constexpr (!(is_default_constructible_v<_Types> && ...))
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr ((__is_implicitly_default_constructible<_Types>::value && ...))
+  {
+    return __select_constructor::__implicit;
+  }
+  else
+  {
+    return __select_constructor::__explicit;
+  }
+}
+
+template <class... _Types>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_variadic_copy_constructible(__tuple_types<_Types...>) noexcept
+{
+  if constexpr (!(is_copy_constructible_v<_Types> && ...))
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr ((is_convertible_v<const _Types&, _Types> && ...))
+  {
+    return __select_constructor::__implicit;
+  }
+  else
+  {
+    return __select_constructor::__explicit;
+  }
+}
+
+template <class _TupleTypes>
+inline constexpr __select_constructor __tuple_select_variadic_copy_constructible_v =
+  ::cuda::std::__tuple_select_variadic_copy_constructible(_TupleTypes{});
+
+template <class... _Types>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_variadic_move_constructible(__tuple_types<_Types...>) noexcept
+{
+  if constexpr (!(is_move_constructible_v<_Types> && ...))
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr ((is_convertible_v<_Types&&, _Types> && ...))
+  {
+    return __select_constructor::__implicit;
+  }
+  else
+  {
+    return __select_constructor::__explicit;
+  }
+}
+
+template <class _TupleTypes>
+inline constexpr __select_constructor __tuple_select_variadic_move_constructible_v =
+  ::cuda::std::__tuple_select_variadic_move_constructible(_TupleTypes{});
+
+template <class... _Types, class... _UTypes>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_variadic_constructible(__tuple_types<_Types...>, __tuple_types<_UTypes...>) noexcept
+{
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (sizeof...(_Types) != sizeof...(_UTypes))
+  { // [tuple.cnstr]-13.1: sizeof...(Types) equals sizeof...(UTypes),
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (sizeof...(_Types) == 0)
+  { // [tuple.cnstr]-13.2: sizeof...(Types) >= 1,
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (sizeof...(_Types) == 2 || sizeof...(_Types) == 3)
+  { // [tuple.cnstr]-12.2: otherwise, if sizeof...(Types) is 2 or 3
+    using _U0 = __type_index_c<0, _UTypes...>;
+    using _T0 = __type_index_c<0, _Types...>;
+    if constexpr (!is_same_v<remove_cvref_t<_U0>, allocator_arg_t> || is_same_v<remove_cvref_t<_T0>, allocator_arg_t>)
+    { // [tuple.cnstr]-13.3: !is_same_v<remove_cvref_t<U0>, allocator_arg_t> || is_same_v<remove_cvref_t<T0>,
+      // allocator_arg_t>>
+      if constexpr (!(is_constructible_v<_Types, _UTypes> && ...))
+      { // [tuple.cnstr]-13.3: is_constructible<Types, UTypes>... is true
+        return __select_constructor::__invalid;
+      }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+      else if constexpr ((reference_constructs_from_temporary_v<_Types, _UTypes&&> || ...))
+      { // [tuple.cnstr]-15: This constructor is defined as deleted if
+        // (reference_constructs_from_temporary_v<Types, UTypes&&> || ...) is true
+        return __select_constructor::__deleted;
+      }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+      else if constexpr (!(is_convertible_v<_UTypes, _Types> && ...))
+      { // [tuple.cnstr]-15: !conjunction_v<is_convertible<UTypes, Types>...>
+        return __select_constructor::__explicit;
+      }
+      else
+      {
+        return __select_constructor::__implicit;
+      }
+    }
+    else
+    {
+      return __select_constructor::__invalid;
+    }
+  }
+  else if constexpr (!(is_constructible_v<_Types, _UTypes> && ...))
+  { // [tuple.cnstr]-13.3: is_constructible<Types, UTypes>... is true
+    return __select_constructor::__invalid;
+  }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  else if constexpr ((reference_constructs_from_temporary_v<_Types, _UTypes&&> || ...))
+  { // [tuple.cnstr]-15: This constructor is defined as deleted if
+    // (reference_constructs_from_temporary_v<Types, UTypes&&> || ...) is true
+    return __select_constructor::__deleted;
+  }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  else if constexpr (!(is_convertible_v<_UTypes, _Types> && ...))
+  { // [tuple.cnstr]-15: !conjunction_v<is_convertible<UTypes, Types>...>
+    return __select_constructor::__explicit;
+  }
+  else
+  {
+    return __select_constructor::__implicit;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <class _Type, class _UType>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_variadic_constructible(__tuple_types<_Type>, __tuple_types<_UType>) noexcept
+{
+  // NOLINTBEGIN(bugprone-branch-clone)
+  // Reject unpacking a tuple_of_iterator_references through the 1-arg variadic constructor so the
+  // dedicated constructor is used instead. Do allow it if the original type is a tuple_of_iterator_references though
+  if constexpr (__is_tuple_of_iterator_references_v<remove_cvref_t<_UType>>
+                && !__is_tuple_of_iterator_references_v<remove_cvref_t<_Type>>)
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (is_same_v<remove_cvref_t<_UType>, tuple<_Type>>)
+  { // [tuple.cnstr]-12.1: negation<is_same<remove_cvref_t<U0>, tuple>> if sizeof...(Types) is 1
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!is_constructible_v<_Type, _UType>)
+  { // [tuple.cnstr]-13.3: is_constructible<Types, UTypes>... is true
+    return __select_constructor::__invalid;
+  }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  else if constexpr (reference_constructs_from_temporary_v<_Type, _UType&&>)
+  { // [tuple.cnstr]-15: This constructor is defined as deleted if
+    // (reference_constructs_from_temporary_v<Types, UTypes&&> || ...) is true
+    return __select_constructor::__deleted;
+  }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  else if constexpr (!is_convertible_v<_UType, _Type>)
+  { // [tuple.cnstr]-15: !conjunction_v<is_convertible<UTypes, Types>...>
+    return __select_constructor::__explicit;
+  }
+  else
+  {
+    return __select_constructor::__implicit;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <class _TupleTypes, class _TupleUTypes>
+inline constexpr __select_constructor __tuple_select_variadic_constructible_v =
+  ::cuda::std::__tuple_select_variadic_constructible(_TupleTypes{}, _TupleUTypes{});
+
+template <class... _Types, class... _UTypes>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_variadic_constructible_less_rank(__tuple_types<_Types...>, __tuple_types<_UTypes...>) noexcept
+{
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (!(sizeof...(_UTypes) < sizeof...(_Types)))
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (sizeof...(_UTypes) == 0)
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (sizeof...(_UTypes) == 1 && (is_same_v<remove_cvref_t<_UTypes>, tuple<_Types...>> && ...))
+  { // Avoid this shadowing the copy / move constructors
+    return __select_constructor::__invalid;
+  }
+  else
+  { // MSVC has issues with constexpr variables here, so no `__can_construct<_Trait>` or constexpr variable
+    using __arg_list = __make_tuple_types_t<__tuple_types<_Types...>, sizeof...(_UTypes)>;
+    if constexpr (::cuda::std::__tuple_select_variadic_constructible(__arg_list{}, __tuple_types<_UTypes...>{})
+                    == __select_constructor::__invalid
+                  || ::cuda::std::__tuple_select_variadic_constructible(__arg_list{}, __tuple_types<_UTypes...>{})
+                       == __select_constructor::__deleted)
+    {
+      return __select_constructor::__invalid;
+    }
+    else
+    {
+      using __defaulted_list = __make_tuple_types_t<__tuple_types<_Types...>, sizeof...(_Types), sizeof...(_UTypes)>;
+      if constexpr (::cuda::std::__tuple_select_default_constructible(__defaulted_list{})
+                      == __select_constructor::__invalid
+                    || ::cuda::std::__tuple_select_default_constructible(__defaulted_list{})
+                         == __select_constructor::__deleted)
+      {
+        return __select_constructor::__invalid;
+      }
+      else
+      {
+        return __select_constructor::__explicit;
+      }
+    }
+  }
+  _CCCL_UNREACHABLE();
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <class _TupleTypes, class _TupleUTypes>
+inline constexpr __select_constructor __tuple_select_variadic_constructible_less_rank_v =
+  ::cuda::std::__tuple_select_variadic_constructible_less_rank(_TupleTypes{}, _TupleUTypes{});
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _UTuple, class... _Types, size_t... _Indices>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_tuple_like_constructible(__tuple_types<_Types...>, __tuple_indices<_Indices...>) noexcept
+{
+  using ::cuda::std::get;
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (__is_cuda_std_ranges_subrange_v<remove_cvref_t<_UTuple>>)
+  { // [tuple#cnstr]-29.2: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (is_same_v<_UTuple, const tuple<_Types...>&> || is_same_v<_UTuple, tuple<_Types...>&&>)
+  { // Prefers the copy/move constructor
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (sizeof...(_Types) == 0)
+  { // Avoids issues with the size 1 constructor below
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!__tuple_like_with_size<_UTuple, sizeof...(_Types)>)
+  { // [tuple#cnstr]-21.1: sizeof...(Types) equals sizeof...(UTypes), and
+    // [tuple#cnstr]-25.1: sizeof...(Types) is 2,
+    // [tuple#cnstr]-29.3: sizeof...(Types) equals sizeof...(UTypes), and
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!(is_constructible_v<_Types, decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))> && ...))
+  { // [tuple.cnstr]-21.2: is_constructible<Types, decltype(get<I>(std::forward<UTuple>(u)))>... is true
+    // [tuple.cnstr]-25.2: is_constructible<Types, decltype(get<I>(std::forward<UTuple>(u)))>... is true
+    // [tuple.cnstr]-29.4: is_constructible<Types, decltype(get<I>(std::forward<UTuple>(u)))>... is true
+    return __select_constructor::__invalid;
+  }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  else if constexpr ((reference_constructs_from_temporary_v<_Types,
+                                                            decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))>
+                      || ...))
+  { // [tuple.cnstr]-23: This constructor is defined as deleted if
+    // [tuple.cnstr]-27: This constructor is defined as deleted if
+    // [tuple.cnstr]-31: This constructor is defined as deleted if
+    // (reference_constructs_from_temporary_v<Types, decltype(get<I>(FWD(u)))> || ...) is true
+    return __select_constructor::__deleted;
+  }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  else if constexpr (!(is_convertible_v<decltype(get<_Indices>(::cuda::std::declval<_UTuple>())), _Types> && ...))
+  { // [tuple.cnstr]-15: The expression inside explicit is equivalent to:
+    // [tuple.cnstr]-23: The expression inside explicit is equivalent to:
+    // [tuple.cnstr]-31: The expression inside explicit is equivalent to:
+    // !(is_convertible_v<decltype(get<I>(FWD(u))), Types> && ...)
+    return __select_constructor::__explicit;
+  }
+  else
+  {
+    return __select_constructor::__implicit;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _UTuple, class _Type, size_t _Index>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor
+__tuple_select_tuple_like_constructible(__tuple_types<_Type>, __tuple_indices<_Index>) noexcept
+{
+  using ::cuda::std::get;
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (__is_cuda_std_ranges_subrange_v<remove_cvref_t<_UTuple>>)
+  { // [tuple#cnstr]-29.2: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (is_same_v<_UTuple, const tuple<_Type>&> || is_same_v<_UTuple, tuple<_Type>&&>)
+  { // Prefers the copy/move constructor
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!__tuple_like_with_size<_UTuple, 1>)
+  { // [tuple#cnstr]-21.1: sizeof...(Types) equals sizeof...(UTypes), and
+    // [tuple#cnstr]-29.3: sizeof...(Types) equals sizeof...(UTypes), and
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (__is_cuda_std_tuple<remove_cvref_t<_UTuple>>
+                     && is_same_v<_Type, tuple_element_t<_Index, remove_cvref_t<_UTuple>>>)
+  { // [tuple#cnstr]-21.3: either sizeof...(Types) is not 1
+    // [tuple#cnstr]-21.3: is_same_v<T, U> is false
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (is_constructible_v<_Type, _UTuple>)
+  { // [tuple#cnstr]-21.3: either sizeof...(Types) is not 1, or is_constructible_v<T, _UTuple> are false
+    // [tuple#cnstr]-29.5: either sizeof...(Types) is not 1, or is_constructible_v<T, _UTuple> are false
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (is_convertible_v<_UTuple, _Type>)
+  { // [tuple#cnstr]-21.3: either sizeof...(Types) is not 1, or is_convertible_v<_UTuple, T> are false
+    // [tuple#cnstr]-29.5: either sizeof...(Types) is not 1, or is_convertible_v<_UTuple, T> are false
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!is_constructible_v<_Type, decltype(get<_Index>(::cuda::std::declval<_UTuple>()))>)
+  { // [tuple.cnstr]-21.2: is_constructible<Types, decltype(get<I>(std::forward<UTuple>(u)))>... is true
+    // [tuple.cnstr]-29.4: is_constructible<Types, decltype(get<I>(std::forward<UTuple>(u)))>... is true
+    return __select_constructor::__invalid;
+  }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  else if constexpr (reference_constructs_from_temporary_v<_Type, decltype(get<_Index>(::cuda::std::declval<_UTuple>()))>)
+  { // [tuple.cnstr]-23: This constructor is defined as deleted if
+    // [tuple.cnstr]-31: This constructor is defined as deleted if
+    // (reference_constructs_from_temporary_v<Types, decltype(get<I>(FWD(u)))> || ...) is true
+    return __select_constructor::__deleted;
+  }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  else if constexpr (!is_convertible_v<decltype(get<_Index>(::cuda::std::declval<_UTuple>())), _Type>)
+  { // [tuple.cnstr]-15: The expression inside explicit is equivalent to:
+    // [tuple.cnstr]-23: The expression inside explicit is equivalent to:
+    // !(is_convertible_v<decltype(get<I>(FWD(u))), Types> && ...)
+    return __select_constructor::__explicit;
+  }
+  else
+  {
+    return __select_constructor::__implicit;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <class _UTuple, class _TupleTypes, class _TupleIndices>
+inline constexpr __select_constructor __tuple_select_tuple_like_constructible_v =
+  ::cuda::std::__tuple_select_tuple_like_constructible<_UTuple>(_TupleTypes{}, _TupleIndices{});
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _UTuple, class _Type, class... _Types, size_t _Index, size_t... _Indices>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL bool
+__tuple_nothrow_tuple_like_constructible(__tuple_types<_Type, _Types...>, __tuple_indices<_Index, _Indices...>) noexcept
+{
+  using ::cuda::std::get;
+  if constexpr (!is_nothrow_constructible_v<_Type, decltype(get<_Index>(::cuda::std::declval<_UTuple>()))>)
+  {
+    return false;
+  }
+  else if constexpr (sizeof...(_Types) != 0)
+  {
+    return ::cuda::std::__tuple_nothrow_tuple_like_constructible<_UTuple>(
+      __tuple_types<_Types...>{}, __tuple_indices<_Indices...>{});
+  }
+  else
+  {
+    return true;
+  }
+}
+
+template <class _UTuple>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL bool
+__tuple_nothrow_tuple_like_constructible(__tuple_types<>, __tuple_indices<>) noexcept
+{
+  return true;
+}
+
+template <class _UTuple, class _TupleTypes, class _TupleIndices>
+inline constexpr bool __tuple_nothrow_tuple_like_constructible_v =
+  ::cuda::std::__tuple_nothrow_tuple_like_constructible<_UTuple>(_TupleTypes{}, _TupleIndices{});
+
+struct _InvalidTupleComparison
+{
+  static constexpr bool __equality_comparable         = false;
+  static constexpr bool __nothrow_equality_comparable = false;
+
+  static constexpr bool __less_than_comparable         = false;
+  static constexpr bool __nothrow_less_than_comparable = false;
+};
+
+template <class, class>
+struct _TupleComparableTraits;
+
+template <class... _Types, class... _UTypes>
+struct _TupleComparableTraits<__tuple_types<_Types...>, __tuple_types<_UTypes...>>
+{
+  static constexpr bool __equality_comparable = (__is_cpp17_equality_comparable_v<_Types, _UTypes> && ...);
+  static constexpr bool __nothrow_equality_comparable =
+    (__is_cpp17_nothrow_equality_comparable_v<_Types, _UTypes> && ...);
+
+  static constexpr bool __less_than_comparable = (__is_cpp17_less_than_comparable_v<_Types, _UTypes> && ...);
+  static constexpr bool __nothrow_less_than_comparable =
+    (__is_cpp17_nothrow_less_than_comparable_v<_Types, _UTypes> && ...);
+};
+
+template <class... _Types, class... _UTypes, enable_if_t<(sizeof...(_Types) == sizeof...(_UTypes)), int> = 0>
+[[nodiscard]]
+_CCCL_API _CCCL_CONSTEVAL auto __tuple_is_comparable(__tuple_types<_Types...>, __tuple_types<_UTypes...>) noexcept
+  -> _TupleComparableTraits<__tuple_types<_Types...>, __tuple_types<_UTypes...>>;
+template <class>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL auto __tuple_is_comparable(...) noexcept -> _InvalidTupleComparison;
+
+template <class... _Types>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_assignment
+__tuple_select_const_copy_assignable(__tuple_types<_Types...>) noexcept
+{
+  if constexpr (!(is_copy_assignable_v<const _Types> && ...))
+  { // [tuple.assign]-5: is_copy_assignable_v<const Types> is true for all i.
+    return __select_assignment::__invalid;
+  }
+  else if constexpr ((is_nothrow_copy_assignable_v<const _Types> && ...))
+  {
+    return __select_assignment::__is_nothrow;
+  }
+  else
+  {
+    return __select_assignment::__may_throw;
+  }
+}
+
+template <class _TupleTypes>
+inline constexpr __select_assignment __tuple_select_const_copy_assignable_v =
+  ::cuda::std::__tuple_select_const_copy_assignable(_TupleTypes{});
+
+template <class... _Types>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_assignment
+__tuple_select_const_move_assignable(__tuple_types<_Types...>) noexcept
+{
+  if constexpr (!(is_assignable_v<const _Types&, _Types> && ...))
+  { // [tuple.assign]-12: is_assignable_v<const Types&, Types> is true for all i.
+    return __select_assignment::__invalid;
+  }
+  else if constexpr ((is_nothrow_assignable_v<const _Types&, _Types> && ...))
+  {
+    return __select_assignment::__is_nothrow;
+  }
+  else
+  {
+    return __select_assignment::__may_throw;
+  }
+}
+
+template <class _TupleTypes>
+inline constexpr __select_assignment __tuple_select_const_move_assignable_v =
+  ::cuda::std::__tuple_select_const_move_assignable(_TupleTypes{});
+
+template <bool _IsConst, class... _Types, class... _UTypes>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_assignment
+__tuple_select_converting_assignable(__tuple_types<_Types...>, __tuple_types<_UTypes...>) noexcept
+{
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (sizeof...(_Types) != sizeof...(_UTypes))
+  { // [tuple.assign]-15.1: sizeof...(Types) equals sizeof...(UTypes) and
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (is_same_v<tuple<_Types...>, tuple<_UTypes...>>)
+  { // Disambiguate the non-converting assignments
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (_IsConst)
+  {
+    if constexpr ((is_assignable_v<const _Types&, _UTypes> && ...))
+    { // [tuple.assign]-18.2: is_assignable_v<const Types&, const UTypes&> is true for all i.
+      // [tuple.assign]-24.2: is_assignable_v<const Types&, UTypes> is true for all i.
+      return (is_nothrow_assignable_v<const _Types&, _UTypes> && ...)
+             ? __select_assignment::__is_nothrow
+             : __select_assignment::__may_throw;
+    }
+    else
+    {
+      return __select_assignment::__invalid;
+    }
+  }
+  else if constexpr ((is_assignable_v<_Types&, _UTypes> && ...))
+  { // [tuple.assign]-15.2: is_assignable_v<Types&, const UTypes&> is true for all i.
+    // [tuple.assign]-21.2: is_assignable_v<Types&, UTypes> is true for all i.
+    return (is_nothrow_assignable_v<_Types&, _UTypes> && ...)
+           ? __select_assignment::__is_nothrow
+           : __select_assignment::__may_throw;
+  }
+  else
+  {
+    return __select_assignment::__invalid;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <bool _IsConst, class _TupleTypes, class _TupleUTypes>
+inline constexpr __select_assignment __tuple_select_converting_assignable_v =
+  ::cuda::std::__tuple_select_converting_assignable<_IsConst>(_TupleTypes{}, _TupleUTypes{});
+
+_CCCL_EXEC_CHECK_DISABLE
+template <bool _IsConst, class _UTuple, class... _Types, size_t... _Indices>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_assignment
+__tuple_select_tuple_like_assignable(__tuple_types<_Types...>, __tuple_indices<_Indices...>) noexcept
+{
+  using ::cuda::std::get;
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (is_same_v<remove_cvref_t<_UTuple>, tuple<_Types...>>)
+  { // [tuple.assign]-39.1: different-from<UTuple, tuple>
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (__is_cuda_std_ranges_subrange_v<remove_cvref_t<_UTuple>>)
+  { // [tuple.assign]-39.2: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (!__tuple_like_with_size<_UTuple, sizeof...(_Types)>)
+  { // [tuple.assign]-39.3: sizeof...(Types) equals tuple_size_v<remove_cvref_t<UTuple>>, and
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (_IsConst)
+  {
+    if constexpr ((is_assignable_v<const _Types&, decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))> && ...))
+    { // [tuple.assign]-42.4: is_assignable_v<const T_i&, decltype(get<i>(std::forward<UTuple>(u)))> is true for
+      // all i
+      return (is_nothrow_assignable_v<const _Types&, decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))> && ...)
+             ? __select_assignment::__is_nothrow
+             : __select_assignment::__may_throw;
+    }
+    else
+    {
+      return __select_assignment::__invalid;
+    }
+  }
+  else if constexpr ((is_assignable_v<_Types&, decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))> && ...))
+  { // [tuple.assign]-39.4: is_assignable_v<T_i&, decltype(get<i>(std::forward<UTuple>(u)))> is true for all i
+    return (is_nothrow_assignable_v<_Types&, decltype(get<_Indices>(::cuda::std::declval<_UTuple>()))> && ...)
+           ? __select_assignment::__is_nothrow
+           : __select_assignment::__may_throw;
+  }
+  else
+  {
+    return __select_assignment::__invalid;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <bool _IsConst, class _UTuple, class _TupleTypes, class _TupleIndices>
+inline constexpr __select_assignment __tuple_select_tuple_like_assignable_v =
+  ::cuda::std::__tuple_select_tuple_like_assignable<_IsConst, _UTuple>(_TupleTypes{}, _TupleIndices{});
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_TUPLE_CONSTRAINTS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_element.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_element.h
new file mode 100644
index 00000000..09a6db2a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_element.h
@@ -0,0 +1,88 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_TUPLE_ELEMENT_H
+#define _CUDA_STD___TUPLE_TUPLE_ELEMENT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/types.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+// cuda::std::tuple_element
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <size_t _Ip, class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element;
+
+template <size_t _Ip, class _Tp>
+using tuple_element_t _CCCL_NODEBUG_ALIAS = typename tuple_element<_Ip, _Tp>::type;
+
+template <size_t _Ip, class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, const _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = const tuple_element_t<_Ip, _Tp>;
+};
+
+template <size_t _Ip, class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, volatile _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = volatile tuple_element_t<_Ip, _Tp>;
+};
+
+template <size_t _Ip, class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, const volatile _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = const volatile tuple_element_t<_Ip, _Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// std::tuple_element
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <size_t _Ip, class _Tp>
+struct tuple_element;
+
+#if _CCCL_FREESTANDING()
+template <size_t _Ip, class _Tp>
+struct tuple_element<_Ip, const _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = const typename tuple_element<_Ip, _Tp>::type;
+};
+
+template <size_t _Ip, class _Tp>
+struct tuple_element<_Ip, volatile _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = volatile typename tuple_element<_Ip, _Tp>::type;
+};
+
+template <size_t _Ip, class _Tp>
+struct tuple_element<_Ip, const volatile _Tp>
+{
+  using type _CCCL_NODEBUG_ALIAS = const volatile typename tuple_element<_Ip, _Tp>::type;
+};
+#endif // _CCCL_FREESTANDING()
+
+_CCCL_END_NAMESPACE_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_TUPLE_ELEMENT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_indices.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_indices.h
new file mode 100644
index 00000000..b8250e68
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_indices.h
@@ -0,0 +1,44 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_MAKE_TUPLE_INDICES_H
+#define _CUDA_STD___TUPLE_MAKE_TUPLE_INDICES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__utility/integer_sequence.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <size_t _Ep, size_t _Sp = 0>
+struct __make_tuple_indices
+{
+  static_assert(_Sp <= _Ep, "__make_tuple_indices input error");
+  using type = __make_indices_imp<_Ep, _Sp>;
+};
+
+template <size_t _Ep, size_t _Sp = 0>
+using __make_tuple_indices_t = typename __make_tuple_indices<_Ep, _Sp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_MAKE_TUPLE_INDICES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_like.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_like.h
new file mode 100644
index 00000000..11b0989c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_like.h
@@ -0,0 +1,103 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_TUPLE_LIKE_H
+#define _CUDA_STD___TUPLE_TUPLE_LIKE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/__fwd/complex.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__fwd/array.h>
+#include <cuda/std/__fwd/complex.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__fwd/subrange.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__tuple_dir/tuple_types.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __tuple_like_ext = false;
+
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<const _Tp> = __tuple_like_ext<_Tp>;
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<volatile _Tp> = __tuple_like_ext<_Tp>;
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<const volatile _Tp> = __tuple_like_ext<_Tp>;
+
+template <class... _Tp>
+inline constexpr bool __tuple_like_ext<tuple<_Tp...>> = true;
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class... _Tp>
+inline constexpr bool __tuple_like_ext<::std::tuple<_Tp...>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _T1, class _T2>
+inline constexpr bool __tuple_like_ext<pair<_T1, _T2>> = true;
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _T1, class _T2>
+inline constexpr bool __tuple_like_ext<::std::pair<_T1, _T2>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Tp, size_t _Size>
+inline constexpr bool __tuple_like_ext<array<_Tp, _Size>> = true;
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Tp, size_t _Size>
+inline constexpr bool __tuple_like_ext<::std::array<_Tp, _Size>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<complex<_Tp>> = true;
+
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<::cuda::complex<_Tp>> = true;
+
+#if _CCCL_HAS_HOST_STD_LIB() && __cpp_lib_tuple_like >= 202311L
+template <class _Tp>
+inline constexpr bool __tuple_like_ext<::std::complex<_Tp>> = true;
+#endif // _CCCL_HAS_HOST_STD_LIB() && __cpp_lib_tuple_like >= 202311L
+
+template <class _Ip, class _Sp, ::cuda::std::ranges::subrange_kind _Kp>
+inline constexpr bool __tuple_like_ext<::cuda::std::ranges::subrange<_Ip, _Sp, _Kp>> = true;
+
+template <class... _Tp>
+inline constexpr bool __tuple_like_ext<__tuple_types<_Tp...>> = true;
+
+template <class _Tp>
+_CCCL_CONCEPT __tuple_like = __tuple_like_ext<remove_cvref_t<_Tp>>;
+
+// Not on line 74 because of __COUNTER__ missing in NVRTC
+template <class _Tp>
+_CCCL_CONCEPT __pair_like = _CCCL_REQUIRES_EXPR((_Tp)) //
+  (requires(__tuple_like<_Tp>), requires(tuple_size<remove_cvref_t<_Tp>>::value == 2));
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_TUPLE_LIKE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_size.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_size.h
new file mode 100644
index 00000000..a3ba5276
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_size.h
@@ -0,0 +1,101 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_TUPLE_SIZE_H
+#define _CUDA_STD___TUPLE_TUPLE_SIZE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_const.h>
+#include <cuda/std/__type_traits/is_volatile.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size;
+
+template <class _Tp, class...>
+using __enable_if_tuple_size_imp = _Tp;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+tuple_size<__enable_if_tuple_size_imp<const _Tp,
+                                      enable_if_t<!is_volatile_v<_Tp>>,
+                                      integral_constant<size_t, sizeof(tuple_size<_Tp>)>>>
+    : integral_constant<size_t, tuple_size<_Tp>::value>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+tuple_size<__enable_if_tuple_size_imp<volatile _Tp,
+                                      enable_if_t<!is_const_v<_Tp>>,
+                                      integral_constant<size_t, sizeof(tuple_size<_Tp>)>>>
+    : integral_constant<size_t, tuple_size<_Tp>::value>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+tuple_size<__enable_if_tuple_size_imp<const volatile _Tp, integral_constant<size_t, sizeof(tuple_size<_Tp>)>>>
+    : integral_constant<size_t, tuple_size<_Tp>::value>
+{};
+
+template <class _Tp>
+inline constexpr size_t tuple_size_v = tuple_size<_Tp>::value;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _Tp>
+struct tuple_size;
+
+#if _CCCL_FREESTANDING()
+
+template <class _Tp>
+struct tuple_size<
+  ::cuda::std::__enable_if_tuple_size_imp<const _Tp,
+                                          ::cuda::std::enable_if_t<!::cuda::std::is_volatile_v<_Tp>>,
+                                          ::cuda::std::integral_constant<::cuda::std::size_t, sizeof(tuple_size<_Tp>)>>>
+    : ::cuda::std::integral_constant<::cuda::std::size_t, tuple_size<_Tp>::value>
+{};
+
+template <class _Tp>
+struct tuple_size<
+  ::cuda::std::__enable_if_tuple_size_imp<volatile _Tp,
+                                          ::cuda::std::enable_if_t<!::cuda::std::is_const_v<_Tp>>,
+                                          ::cuda::std::integral_constant<::cuda::std::size_t, sizeof(tuple_size<_Tp>)>>>
+    : ::cuda::std::integral_constant<::cuda::std::size_t, tuple_size<_Tp>::value>
+{};
+
+template <class _Tp>
+struct tuple_size<
+  ::cuda::std::__enable_if_tuple_size_imp<const volatile _Tp,
+                                          ::cuda::std::integral_constant<::cuda::std::size_t, sizeof(tuple_size<_Tp>)>>>
+    : ::cuda::std::integral_constant<::cuda::std::size_t, tuple_size<_Tp>::value>
+{};
+#endif // _CCCL_FREESTANDING()
+
+_CCCL_END_NAMESPACE_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_TUPLE_SIZE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_types.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_types.h
new file mode 100644
index 00000000..da4f6043
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__tuple_dir/tuple_types.h
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TUPLE_TUPLE_TYPES_H
+#define _CUDA_STD___TUPLE_TUPLE_TYPES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/type_list.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class... _Tp>
+struct __tuple_types
+{};
+
+// specialize cuda::std::tuple_size and cuda::std::tuple_element for cuda::std::__tuple_types
+
+template <class... _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<__tuple_types<_Tp...>> : integral_constant<size_t, sizeof...(_Tp)>
+{};
+
+template <size_t _Ip, class... _Types>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+tuple_element<_Ip, __tuple_types<_Types...>> : enable_if<(_Ip < sizeof...(_Types)), __type_index_c<_Ip, _Types...>>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TUPLE_TUPLE_TYPES_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_const.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_const.h
new file mode 100644
index 00000000..fe5d4106
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_const.h
@@ -0,0 +1,40 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ADD_CONST_H
+#define _CUDA_STD___TYPE_TRAITS_ADD_CONST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT add_const
+{
+  using type _CCCL_NODEBUG_ALIAS = const _Tp;
+};
+
+template <class _Tp>
+using add_const_t _CCCL_NODEBUG_ALIAS = const _Tp;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ADD_CONST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_lvalue_reference.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_lvalue_reference.h
new file mode 100644
index 00000000..07be75ff
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_lvalue_reference.h
@@ -0,0 +1,62 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ADD_LVALUE_REFERENCE_H
+#define _CUDA_STD___TYPE_TRAITS_ADD_LVALUE_REFERENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_referenceable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_ADD_LVALUE_REFERENCE) && !defined(_LIBCUDACXX_USE_ADD_LVALUE_REFERENCE_FALLBACK)
+
+template <class _Tp>
+using add_lvalue_reference_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_ADD_LVALUE_REFERENCE(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_ADD_LVALUE_REFERENCE ^^^ / vvv !_CCCL_BUILTIN_ADD_LVALUE_REFERENCE vvv
+
+template <class _Tp, bool = __cccl_is_referenceable<_Tp>::value>
+struct __add_lvalue_reference_impl
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct __add_lvalue_reference_impl<_Tp, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp&;
+};
+
+template <class _Tp>
+using add_lvalue_reference_t _CCCL_NODEBUG_ALIAS = typename __add_lvalue_reference_impl<_Tp>::type;
+
+#endif // !_CCCL_BUILTIN_ADD_LVALUE_REFERENCE
+
+template <class _Tp>
+struct add_lvalue_reference
+{
+  using type _CCCL_NODEBUG_ALIAS = add_lvalue_reference_t<_Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ADD_LVALUE_REFERENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_pointer.h
new file mode 100644
index 00000000..6409382c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_pointer.h
@@ -0,0 +1,65 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ADD_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_ADD_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_referenceable.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_ADD_POINTER) && !defined(_LIBCUDACXX_USE_ADD_POINTER_FALLBACK)
+
+template <class _Tp>
+using add_pointer_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_ADD_POINTER(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_ADD_POINTER ^^^ / vvv !_CCCL_BUILTIN_ADD_POINTER vvv
+template <class _Tp, bool = __cccl_is_referenceable<_Tp>::value || is_void<_Tp>::value>
+struct __add_pointer_impl
+{
+  using type _CCCL_NODEBUG_ALIAS = remove_reference_t<_Tp>*;
+};
+template <class _Tp>
+struct __add_pointer_impl<_Tp, false>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+template <class _Tp>
+using add_pointer_t _CCCL_NODEBUG_ALIAS = typename __add_pointer_impl<_Tp>::type;
+
+#endif // !_CCCL_BUILTIN_ADD_POINTER
+
+template <class _Tp>
+struct add_pointer
+{
+  using type _CCCL_NODEBUG_ALIAS = add_pointer_t<_Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ADD_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_rvalue_reference.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_rvalue_reference.h
new file mode 100644
index 00000000..67580e8c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/add_rvalue_reference.h
@@ -0,0 +1,62 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ADD_RVALUE_REFERENCE_H
+#define _CUDA_STD___TYPE_TRAITS_ADD_RVALUE_REFERENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_referenceable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_ADD_RVALUE_REFERENCE) && !defined(_LIBCUDACXX_USE_ADD_RVALUE_REFERENCE_FALLBACK)
+
+template <class _Tp>
+using add_rvalue_reference_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_ADD_RVALUE_REFERENCE(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_ADD_RVALUE_REFERENCE ^^^ / vvv !_CCCL_BUILTIN_ADD_RVALUE_REFERENCE vvv
+
+template <class _Tp, bool = __cccl_is_referenceable<_Tp>::value>
+struct __add_rvalue_reference_impl
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct __add_rvalue_reference_impl<_Tp, true>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp&&;
+};
+
+template <class _Tp>
+using add_rvalue_reference_t _CCCL_NODEBUG_ALIAS = typename __add_rvalue_reference_impl<_Tp>::type;
+
+#endif // _CCCL_BUILTIN_ADD_RVALUE_REFERENCE
+
+template <class _Tp>
+struct add_rvalue_reference
+{
+  using type = add_rvalue_reference_t<_Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ADD_RVALUE_REFERENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/aggregate_members_all_of.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/aggregate_members_all_of.h
new file mode 100644
index 00000000..faf859c9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/aggregate_members_all_of.h
@@ -0,0 +1,135 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_AGGREGATE_MEMBERS_ALL_OF_H
+#define _CUDA_STD___TYPE_TRAITS_AGGREGATE_MEMBERS_ALL_OF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_aggregate.h>
+#include <cuda/std/__type_traits/is_empty.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/integer_sequence.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_CLANG("-Wmissing-field-initializers")
+_CCCL_DIAG_SUPPRESS_CLANG("-Wmissing-braces")
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// provide a generic way to initialize an aggregate member
+struct __any_aggregate_member
+{
+  template <typename _Tp>
+  _CCCL_API constexpr operator _Tp&&() const;
+};
+
+// Computes the number of aggregate members
+template <typename _Tp>
+struct __aggregate_arity_impl
+{
+  // N == 0: skip _Up{} to avoid inconsistencies across compilers (deleted default constructor)
+  template <typename _Self = __aggregate_arity_impl>
+  _CCCL_API auto operator()() -> decltype(_Self{}(__any_aggregate_member{}));
+
+  // N >= 1: SFINAE on _Up{_A0{}, _Args{}...}
+  template <typename _A0,
+            typename... _Args,
+            typename _Up   = _Tp,
+            typename       = decltype(_Up{_A0{}, (_Args{})...}),
+            typename _Self = __aggregate_arity_impl>
+  _CCCL_API auto operator()(_A0, _Args... __args) -> decltype(_Self{}(_A0{}, __args..., __any_aggregate_member{}));
+
+  template <typename... _Args>
+  _CCCL_API auto operator()(_Args...) const -> char (*)[sizeof...(_Args) + 1]; // return the number of members + 1
+};
+
+// Returns the number of aggregate members. Only meaningful when is_aggregate_v<_Tp> is true.
+template <typename _Tp>
+inline constexpr int __aggregate_arity_v = int{sizeof(*__aggregate_arity_impl<_Tp>{}())} - 2;
+
+// Apply a Predicate to every aggregate member
+
+// provide a generic way to initialize an aggregate member but only if the Predicate is true
+template <template <typename> class _Predicate>
+struct __aggregate_member_if
+{
+  template <typename _Tp, typename = enable_if_t<_Predicate<remove_cvref_t<_Tp>>::value>>
+  _CCCL_API constexpr operator _Tp&&() const;
+};
+
+inline constexpr int __aggregate_max_arity = 8;
+
+// Apply the Predicate to every member
+template <int _Arity>
+struct __aggregate_all_of_fn
+{
+  template <template <typename> class _Predicate, typename _Tp, size_t... _Is>
+  _CCCL_API static auto __test(index_sequence<_Is...>, int)
+    -> decltype(_Tp{((void) _Is, __aggregate_member_if<_Predicate>{})...}, true_type{});
+
+  template <template <typename> class, typename>
+  _CCCL_API static auto __test(...) -> false_type;
+
+  template <template <typename> class _Predicate, typename _Tp>
+  _CCCL_API static auto __call(int) -> decltype(__test<_Predicate, _Tp>(make_index_sequence<_Arity>{}, 0));
+};
+
+// (aggregate-only trait) Dispatch based on the aggregate arity for non-empty aggregates. Returns:
+// - false: _Arity out of [1, __aggregate_max_arity]
+// - true:  if the unary predicate is true for all members
+template <template <typename> class _Predicate,
+          typename _Tp,
+          int _Arity = __aggregate_arity_v<_Tp>,
+          bool       = (_Arity > 0) && (_Arity <= __aggregate_max_arity)>
+inline constexpr bool __aggregate_all_of_dispatch_v = false;
+
+template <template <typename> class _Predicate, typename _Tp, int _Arity>
+inline constexpr bool __aggregate_all_of_dispatch_v<_Predicate, _Tp, _Arity, true> =
+  decltype(__aggregate_all_of_fn<_Arity>::template __call<_Predicate, _Tp>(0))::value;
+
+// if _Tp is not an aggregate, return false.
+// Empty aggregates are true for any predicate.
+// The specialization is needed to skip computing the arity and save compile time.
+template <template <typename> class _Predicate,
+          typename _Tp,
+          bool = is_aggregate_v<_Tp>,
+          bool = is_empty_v<_Tp>> // required for nvc++ and NVCC+clang
+inline constexpr bool __aggregate_all_of_v = false;
+
+// (non-empty) aggregate
+template <template <typename> class _Predicate, typename _Tp>
+inline constexpr bool __aggregate_all_of_v<_Predicate, _Tp, true, false> =
+  __aggregate_all_of_dispatch_v<_Predicate, _Tp>;
+
+// Empty aggregate (true for any predicate)
+template <template <typename> class _Predicate, typename _Tp>
+inline constexpr bool __aggregate_all_of_v<_Predicate, _Tp, true, true> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_DIAG_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_AGGREGATE_MEMBERS_ALL_OF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/always_false.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/always_false.h
new file mode 100644
index 00000000..10ef14af
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/always_false.h
@@ -0,0 +1,35 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ALWAYS_FALSE_H
+#define _CUDA_STD___TYPE_TRAITS_ALWAYS_FALSE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+inline constexpr bool __always_false_v = false;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ALWAYS_FALSE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_reference.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_reference.h
new file mode 100644
index 00000000..94e2360c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_reference.h
@@ -0,0 +1,262 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+// SPDX-FileCopyrightText: Copyright (c) Microsoft Corporation.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_COMMON_REFERENCE_H
+#define _CUDA_STD___TYPE_TRAITS_COMMON_REFERENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/common_type.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/copy_cv.h>
+#include <cuda/std/__type_traits/copy_cvref.h>
+#include <cuda/std/__type_traits/disjunction.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_scalar.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(1384) // warning: pointer converted to bool
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// common_reference
+
+// Let COND_RES(X, Y) be:
+#if _CCCL_COMPILER(MSVC) // Workaround for DevCom-1627396
+template <class _Tp>
+_Tp __returns_exactly() noexcept; // not defined
+
+template <class _Xp, class _Yp>
+using __cond_res_if_right = decltype(false ? __returns_exactly<_Xp>() : __returns_exactly<_Yp>());
+
+template <class _Tp, class _Up, class = void>
+struct __cond_res_workaround
+{};
+
+template <class _Tp, class _Up>
+struct __cond_res_workaround<_Tp, _Up, void_t<__cond_res_if_right<_Tp, _Up>>>
+{
+  using _RTp = remove_cvref_t<_Tp>;
+  using type =
+    conditional_t<is_same_v<_RTp, remove_cvref_t<_Up>> && (is_scalar_v<_RTp> || is_array_v<_RTp>)
+                    && ((is_lvalue_reference_v<_Tp> && is_rvalue_reference_v<_Up>)
+                        || (is_rvalue_reference_v<_Tp> && is_lvalue_reference_v<_Up>) ),
+                  decay_t<__copy_cv_t<remove_reference_t<_Tp>, remove_reference_t<_Up>>>,
+                  __cond_res_if_right<_Tp, _Up>>;
+};
+
+template <class _Xp, class _Yp>
+using __cond_res = typename __cond_res_workaround<_Xp, _Yp>::type;
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+template <class _Xp, class _Yp>
+using __cond_res = decltype(false ? ::cuda::std::declval<_Xp (&)()>()() : ::cuda::std::declval<_Yp (&)()>()());
+#endif // !_CCCL_COMPILER(MSVC)
+
+// Let `XREF(A)` denote a unary alias template `T` such that `T<U>` denotes the same type as `U`
+// with the addition of `A`'s cv and reference qualifiers, for a non-reference cv-unqualified type
+// `U`.
+// [Note: `XREF(A)` is `__xref<A>::template __call`]
+template <class _Tp>
+using __xref = __apply_cvref_fn<_Tp>;
+
+// Given types A and B, let X be remove_reference_t<A>, let Y be remove_reference_t<B>,
+// and let COMMON-REF(A, B) be:
+template <class _Ap, class _Bp, class = void>
+struct __common_ref;
+
+template <class _Xp, class _Yp>
+using __common_ref_t _CCCL_NODEBUG_ALIAS = typename __common_ref<_Xp, _Yp>::__type;
+
+template <class _Xp, class _Yp>
+using __cv_cond_res = __cond_res<__copy_cv_t<_Xp, _Yp>&, __copy_cv_t<_Yp, _Xp>&>;
+
+//    If A and B are both lvalue reference types, COMMON-REF(A, B) is
+//    COND-RES(COPYCV(X, Y)&, COPYCV(Y, X)&) if that type exists and is a reference type.
+template <class _Ap, class _Bp>
+struct __common_ref<_Ap&, _Bp&, enable_if_t<is_reference_v<__cv_cond_res<_Ap, _Bp>>>>
+{
+  using __type = __cv_cond_res<_Ap, _Bp>;
+};
+
+//    Otherwise, let C be remove_reference_t<COMMON-REF(X&, Y&)>&&. ...
+template <class _Xp, class _Yp>
+using __common_ref_C = remove_reference_t<__common_ref_t<_Xp&, _Yp&>>&&;
+
+//    .... If A and B are both rvalue reference types, C is well-formed, and
+//    is_convertible_v<A, C> && is_convertible_v<B, C> is true, then COMMON-REF(A, B) is C.
+template <class _Ap, class _Bp, class = void>
+struct __common_ref_rr
+{};
+
+template <class _Ap, class _Bp>
+struct __common_ref_rr<
+  _Ap&&,
+  _Bp&&,
+  enable_if_t<is_convertible_v<_Ap&&, __common_ref_C<_Ap, _Bp>> && is_convertible_v<_Bp&&, __common_ref_C<_Ap, _Bp>>>>
+{
+  using __type = __common_ref_C<_Ap, _Bp>;
+};
+
+template <class _Ap, class _Bp>
+struct __common_ref<_Ap&&, _Bp&&> : __common_ref_rr<_Ap&&, _Bp&&>
+{};
+
+//    Otherwise, let D be COMMON-REF(const X&, Y&). ...
+template <class _Tp, class _Up>
+using __common_ref_D = __common_ref_t<const _Tp&, _Up&>;
+
+//    ... If A is an rvalue reference and B is an lvalue reference and D is well-formed and
+//    is_convertible_v<A, D> is true, then COMMON-REF(A, B) is D.
+template <class _Ap, class _Bp, class = void>
+struct __common_ref_lr
+{};
+
+template <class _Ap, class _Bp>
+struct __common_ref_lr<_Ap&&, _Bp&, enable_if_t<is_convertible_v<_Ap&&, __common_ref_D<_Ap, _Bp>>>>
+{
+  using __type = __common_ref_D<_Ap, _Bp>;
+};
+
+template <class _Ap, class _Bp>
+struct __common_ref<_Ap&&, _Bp&> : __common_ref_lr<_Ap&&, _Bp&>
+{};
+
+//    Otherwise, if A is an lvalue reference and B is an rvalue reference, then
+//    COMMON-REF(A, B) is COMMON-REF(B, A).
+template <class _Ap, class _Bp>
+struct __common_ref<_Ap&, _Bp&&> : __common_ref_lr<_Bp&&, _Ap&>
+{};
+
+//    Otherwise, COMMON-REF(A, B) is ill-formed.
+template <class _Ap, class _Bp, class>
+struct __common_ref
+{};
+
+// Note C: For the common_reference trait applied to a parameter pack [...]
+
+template <class...>
+struct common_reference;
+
+template <class... _Types>
+using common_reference_t _CCCL_NODEBUG_ALIAS = typename common_reference<_Types...>::type;
+
+template <class, class, class = void>
+inline constexpr bool __has_common_reference = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __has_common_reference<_Tp, _Up, void_t<common_reference_t<_Tp, _Up>>> = true;
+
+// bullet 1 - sizeof...(T) == 0
+template <>
+struct common_reference<>
+{};
+
+// bullet 2 - sizeof...(T) == 1
+template <class _Tp>
+struct common_reference<_Tp>
+{
+  using type = _Tp;
+};
+
+// bullet 3 - sizeof...(T) == 2
+template <class _Tp, class _Up, class = void>
+struct __common_reference_sub_bullet3;
+template <class _Tp, class _Up, class = void>
+struct __common_reference_sub_bullet2 : __common_reference_sub_bullet3<_Tp, _Up>
+{};
+template <class _Tp, class _Up, class = void>
+struct __common_reference_sub_bullet1 : __common_reference_sub_bullet2<_Tp, _Up>
+{};
+
+// sub-bullet 1 - If T1 and T2 are reference types and COMMON-REF(T1, T2) is well-formed, then
+// the member typedef `type` denotes that type.
+template <class _Tp, class _Up>
+struct common_reference<_Tp, _Up> : __common_reference_sub_bullet1<_Tp, _Up>
+{};
+
+template <class _Tp, class _Up>
+struct __common_reference_sub_bullet1<
+  _Tp,
+  _Up,
+  void_t<__common_ref_t<_Tp, _Up>, enable_if_t<is_reference_v<_Tp> && is_reference_v<_Up>>>>
+{
+  using type = __common_ref_t<_Tp, _Up>;
+};
+
+// sub-bullet 2 - Otherwise, if basic_common_reference<remove_cvref_t<T1>, remove_cvref_t<T2>, XREF(T1), XREF(T2)>::type
+// is well-formed, then the member typedef `type` denotes that type.
+template <class, class, template <class> class, template <class> class, class = void>
+struct basic_common_reference
+{};
+
+template <class _Tp, class _Up>
+using __basic_common_reference_t _CCCL_NODEBUG_ALIAS =
+  typename basic_common_reference<remove_cvref_t<_Tp>,
+                                  remove_cvref_t<_Up>,
+                                  __xref<_Tp>::template __call,
+                                  __xref<_Up>::template __call>::type;
+
+template <class _Tp, class _Up>
+struct __common_reference_sub_bullet2<_Tp, _Up, void_t<__basic_common_reference_t<_Tp, _Up>>>
+{
+  using type = __basic_common_reference_t<_Tp, _Up>;
+};
+
+// sub-bullet 3 - Otherwise, if COND-RES(T1, T2) is well-formed,
+// then the member typedef `type` denotes that type.
+template <class _Tp, class _Up>
+struct __common_reference_sub_bullet3<_Tp, _Up, void_t<__cond_res<_Tp, _Up>>>
+{
+  using type = __cond_res<_Tp, _Up>;
+};
+
+// sub-bullet 4 & 5 - Otherwise, if common_type_t<T1, T2> is well-formed,
+//                    then the member typedef `type` denotes that type.
+//                  - Otherwise, there shall be no member `type`.
+template <class _Tp, class _Up, class>
+struct __common_reference_sub_bullet3 : common_type<_Tp, _Up>
+{};
+
+// bullet 4 - If there is such a type `C`, the member typedef type shall denote the same type, if
+//            any, as `common_reference_t<C, Rest...>`.
+template <class _Tp, class _Up, class _Vp, class... _Rest>
+struct common_reference<_Tp, _Up, _Vp, void_t<common_reference_t<_Tp, _Up>>, _Rest...>
+    : common_reference<common_reference_t<_Tp, _Up>, _Vp, _Rest...>
+{};
+
+// bullet 5 - Otherwise, there shall be no member `type`.
+template <class...>
+struct common_reference
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_END_NV_DIAG_SUPPRESS()
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_COMMON_REFERENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_type.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_type.h
new file mode 100644
index 00000000..14685592
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/common_type.h
@@ -0,0 +1,173 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_COMMON_TYPE_H
+#define _CUDA_STD___TYPE_TRAITS_COMMON_TYPE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/copy_cvref.h>
+#include <cuda/std/__type_traits/decay.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_arithmetic.h>
+#include <cuda/std/__type_traits/is_extended_floating_point.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class... _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT common_type;
+
+template <class... _Tp>
+using common_type_t _CCCL_NODEBUG_ALIAS = typename common_type<_Tp...>::type;
+
+// Let COND_RES(X, Y) be:
+template <class _Tp, class _Up>
+using __cond_type = decltype(false ? declval<_Tp>() : declval<_Up>());
+
+// We need to ensure that extended floating point types like __half and __nv bfloat16 have a common type with real
+// floating point types
+template <class _Tp, class _Up, class = void>
+struct __common_type_extended_floating_point
+{};
+
+#if !defined(__CUDA_NO_HALF_CONVERSIONS__) && !defined(__CUDA_NO_HALF_OPERATORS__) \
+  && !defined(__CUDA_NO_BFLOAT16_CONVERSIONS__) && !defined(__CUDA_NO_BFLOAT16_OPERATORS__)
+template <class _Tp, class _Up>
+struct __common_type_extended_floating_point<
+  _Tp,
+  _Up,
+  enable_if_t<__is_extended_floating_point_v<remove_cvref_t<_Tp>> && is_arithmetic_v<remove_cvref_t<_Up>>>>
+{
+  using type = common_type_t<__copy_cvref_t<_Tp, float>, _Up>;
+};
+
+template <class _Tp, class _Up>
+struct __common_type_extended_floating_point<
+  _Tp,
+  _Up,
+  enable_if_t<is_arithmetic_v<remove_cvref_t<_Tp>> && __is_extended_floating_point_v<remove_cvref_t<_Up>>>>
+{
+  using type = common_type_t<_Tp, __copy_cvref_t<_Up, float>>;
+};
+#endif // extended floating point as arithmetic type
+
+template <class _Tp, class _Up, class = void>
+struct __common_type3 : __common_type_extended_floating_point<_Tp, _Up>
+{};
+
+#if _CCCL_STD_VER >= 2020
+// sub-bullet 4 - "if COND_RES(CREF(D1), CREF(D2)) denotes a type..."
+template <class _Tp, class _Up>
+struct __common_type3<_Tp, _Up, void_t<__cond_type<const _Tp&, const _Up&>>>
+{
+  using type = remove_cvref_t<__cond_type<const _Tp&, const _Up&>>;
+};
+#endif // _CCCL_STD_VER >= 2020
+
+template <class _Tp, class _Up, class = void>
+struct __common_type2_imp : __common_type3<_Tp, _Up>
+{};
+
+// MSVC has a bug in its declval handling, where it happily accepts __cond_type<_Tp, _Up>, even though both
+// branches have diverging return types, this happens for extended floating point types
+template <class _Tp, class _Up>
+using __msvc_declval_workaround =
+#if _CCCL_COMPILER(MSVC)
+  enable_if_t<is_same_v<__cond_type<_Tp, _Up>, __cond_type<_Up, _Tp>>>;
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+  void;
+#endif // !_CCCL_COMPILER(MSVC)
+
+// sub-bullet 3 - "if decay_t<decltype(false ? declval<D1>() : declval<D2>())> ..."
+template <class _Tp, class _Up>
+struct __common_type2_imp<_Tp, _Up, void_t<__cond_type<_Tp, _Up>, __msvc_declval_workaround<_Tp, _Up>>>
+{
+  using type _CCCL_NODEBUG_ALIAS = decay_t<__cond_type<_Tp, _Up>>;
+};
+
+template <class, class = void>
+struct __common_type_impl
+{};
+
+template <class... _Tp>
+struct __common_types;
+
+template <class _Tp, class _Up>
+struct __common_type_impl<__common_types<_Tp, _Up>, void_t<common_type_t<_Tp, _Up>>>
+{
+  using type = common_type_t<_Tp, _Up>;
+};
+
+template <class _Tp, class _Up, class _Vp, class... _Rest>
+struct __common_type_impl<__common_types<_Tp, _Up, _Vp, _Rest...>, void_t<common_type_t<_Tp, _Up>>>
+    : __common_type_impl<__common_types<common_type_t<_Tp, _Up>, _Vp, _Rest...>>
+{};
+
+// bullet 1 - sizeof...(Tp) == 0
+
+template <>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT common_type<>
+{};
+
+// bullet 2 - sizeof...(Tp) == 1
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT common_type<_Tp> : public common_type<_Tp, _Tp>
+{};
+
+// bullet 3 - sizeof...(Tp) == 2
+
+// sub-bullet 1 - "If is_same_v<T1, D1> is false or ..."
+template <class _Tp, class _Up, class _D1 = decay_t<_Tp>, class _D2 = decay_t<_Up>>
+struct __common_type2 : common_type<_D1, _D2>
+{};
+
+template <class _Tp, class _Up>
+struct __common_type2<_Tp, _Up, _Tp, _Up> : __common_type2_imp<_Tp, _Up>
+{};
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT common_type<_Tp, _Up> : __common_type2<_Tp, _Up>
+{};
+
+// bullet 4 - sizeof...(Tp) > 2
+
+template <class _Tp, class _Up, class _Vp, class... _Rest>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+common_type<_Tp, _Up, _Vp, _Rest...> : __common_type_impl<__common_types<_Tp, _Up, _Vp, _Rest...>>
+{};
+
+template <class... _Tp>
+using common_type_t _CCCL_NODEBUG_ALIAS = typename common_type<_Tp...>::type;
+
+template <class, class, class = void>
+inline constexpr bool __has_common_type = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __has_common_type<_Tp, _Up, void_t<common_type_t<_Tp, _Up>>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_COMMON_TYPE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conditional.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conditional.h
new file mode 100644
index 00000000..a6329f83
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conditional.h
@@ -0,0 +1,65 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_CONDITIONAL_H
+#define _CUDA_STD___TYPE_TRAITS_CONDITIONAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <bool>
+struct _IfImpl;
+
+template <>
+struct _IfImpl<true>
+{
+  template <class _IfRes, class _ElseRes>
+  using _Select _CCCL_NODEBUG_ALIAS = _IfRes;
+};
+
+template <>
+struct _IfImpl<false>
+{
+  template <class _IfRes, class _ElseRes>
+  using _Select _CCCL_NODEBUG_ALIAS = _ElseRes;
+};
+
+template <bool _Cond, class _IfRes, class _ElseRes>
+using _If _CCCL_NODEBUG_ALIAS = typename _IfImpl<_Cond>::template _Select<_IfRes, _ElseRes>;
+
+template <bool _Bp, class _If, class _Then>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT conditional
+{
+  using type = _If;
+};
+template <class _If, class _Then>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT conditional<false, _If, _Then>
+{
+  using type = _Then;
+};
+
+template <bool _Bp, class _If, class _Then>
+using conditional_t _CCCL_NODEBUG_ALIAS = typename conditional<_Bp, _If, _Then>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_CONDITIONAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conjunction.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conjunction.h
new file mode 100644
index 00000000..996cb6b6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/conjunction.h
@@ -0,0 +1,67 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_CONJUNCTION_H
+#define _CUDA_STD___TYPE_TRAITS_CONJUNCTION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+using __expand_to_true = true_type;
+
+template <class... _Pred>
+_CCCL_HOST_DEVICE __expand_to_true<enable_if_t<_Pred::value>...> __and_helper(int);
+
+template <class...>
+_CCCL_HOST_DEVICE false_type __and_helper(...);
+
+// _And always performs lazy evaluation of its arguments.
+//
+// However, `_And<_Pred...>` itself will evaluate its result immediately (without having to
+// be instantiated) since it is an alias, unlike `conjunction<_Pred...>`, which is a struct.
+// If you want to defer the evaluation of `_And<_Pred...>` itself, use `_Lazy<_And, _Pred...>`.
+template <class... _Pred>
+using _And _CCCL_NODEBUG_ALIAS = decltype(__and_helper<_Pred...>(0));
+
+template <class...>
+struct conjunction : true_type
+{};
+
+template <class _Arg>
+struct conjunction<_Arg> : _Arg
+{};
+
+template <class _Arg, class... _Args>
+struct conjunction<_Arg, _Args...> : _If<!bool(_Arg::value), _Arg, conjunction<_Args...>>
+{};
+
+template <class... _Args>
+inline constexpr bool conjunction_v = conjunction<_Args...>::value;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_CONJUNCTION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cv.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cv.h
new file mode 100644
index 00000000..278245e7
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cv.h
@@ -0,0 +1,50 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_COPY_CV_H
+#define _CUDA_STD___TYPE_TRAITS_COPY_CV_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/copy_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// Let COPYCV(FROM, TO) be an alias for type TO with the addition of FROM's
+// top-level cv-qualifiers.
+template <class>
+extern __apply_cvref_ __apply_cv;
+template <class _Tp>
+extern __apply_cvref_c __apply_cv<const _Tp>;
+template <class _Tp>
+extern __apply_cvref_v __apply_cv<volatile _Tp>;
+template <class _Tp>
+extern __apply_cvref_cv __apply_cv<const volatile _Tp>;
+
+template <class _Tp>
+using __apply_cv_fn = decltype(__apply_cv<_Tp>);
+
+template <class _From, class _To>
+using __copy_cv_t = typename __apply_cv_fn<_From>::template __call<_To>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_COPY_CV_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cvref.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cvref.h
new file mode 100644
index 00000000..de59e463
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/copy_cvref.h
@@ -0,0 +1,148 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_COPY_CVREF_H
+#define _CUDA_STD___TYPE_TRAITS_COPY_CVREF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/add_rvalue_reference.h>
+
+#if _CCCL_COMPILER(GCC, <, 7)
+#  define _CCCL_ADD_LVALUE_REFERENCE_WAR(_Tp) add_lvalue_reference_t<_Tp>
+#  define _CCCL_ADD_RVALUE_REFERENCE_WAR(_Tp) add_rvalue_reference_t<_Tp>
+#else
+#  define _CCCL_ADD_LVALUE_REFERENCE_WAR(_Tp) _Tp&
+#  define _CCCL_ADD_RVALUE_REFERENCE_WAR(_Tp) _Tp&&
+#endif
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct __apply_cvref_
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+struct __apply_cvref_c
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = const _Tp;
+};
+
+struct __apply_cvref_v
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = volatile _Tp;
+};
+
+struct __apply_cvref_cv
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = const volatile _Tp;
+};
+
+struct __apply_cvref_lr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_LVALUE_REFERENCE_WAR(_Tp);
+};
+
+struct __apply_cvref_clr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_LVALUE_REFERENCE_WAR(const _Tp);
+};
+
+struct __apply_cvref_vlr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_LVALUE_REFERENCE_WAR(volatile _Tp);
+};
+
+struct __apply_cvref_cvlr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_LVALUE_REFERENCE_WAR(const volatile _Tp);
+};
+
+struct __apply_cvref_rr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_RVALUE_REFERENCE_WAR(_Tp);
+};
+
+struct __apply_cvref_crr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_RVALUE_REFERENCE_WAR(const _Tp);
+};
+
+struct __apply_cvref_vrr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_RVALUE_REFERENCE_WAR(volatile _Tp);
+};
+
+struct __apply_cvref_cvrr
+{
+  template <class _Tp>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_ADD_RVALUE_REFERENCE_WAR(const volatile _Tp);
+};
+
+template <class>
+extern __apply_cvref_ __apply_cvref;
+template <class _Tp>
+extern __apply_cvref_c __apply_cvref<const _Tp>;
+template <class _Tp>
+extern __apply_cvref_v __apply_cvref<volatile _Tp>;
+template <class _Tp>
+extern __apply_cvref_cv __apply_cvref<const volatile _Tp>;
+template <class _Tp>
+extern __apply_cvref_lr __apply_cvref<_Tp&>;
+template <class _Tp>
+extern __apply_cvref_clr __apply_cvref<const _Tp&>;
+template <class _Tp>
+extern __apply_cvref_vlr __apply_cvref<volatile _Tp&>;
+template <class _Tp>
+extern __apply_cvref_cvlr __apply_cvref<const volatile _Tp&>;
+template <class _Tp>
+extern __apply_cvref_rr __apply_cvref<_Tp&&>;
+template <class _Tp>
+extern __apply_cvref_crr __apply_cvref<const _Tp&&>;
+template <class _Tp>
+extern __apply_cvref_vrr __apply_cvref<volatile _Tp&&>;
+template <class _Tp>
+extern __apply_cvref_cvrr __apply_cvref<const volatile _Tp&&>;
+
+template <class _Tp>
+using __apply_cvref_fn = decltype(__apply_cvref<_Tp>);
+
+template <class _From, class _To>
+using __copy_cvref_t = typename __apply_cvref_fn<_From>::template __call<_To>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#undef _CCCL_ADD_RVALUE_REFERENCE_WAR
+#undef _CCCL_ADD_LVALUE_REFERENCE_WAR
+
+#endif // _CUDA_STD___TYPE_TRAITS_COPY_CVREF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/decay.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/decay.h
new file mode 100644
index 00000000..e53949e5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/decay.h
@@ -0,0 +1,83 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_DECAY_H
+#define _CUDA_STD___TYPE_TRAITS_DECAY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_pointer.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_function.h>
+#include <cuda/std/__type_traits/is_referenceable.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_extent.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_DECAY) && !defined(_LIBCUDACXX_USE_DECAY_FALLBACK)
+template <class _Tp>
+struct decay
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_DECAY(_Tp);
+};
+
+template <class _Tp>
+using decay_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_DECAY(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_DECAY ^^^ / vvv !_CCCL_BUILTIN_DECAY vvv
+
+template <class _Up, bool>
+struct __decay_impl
+{
+  using type _CCCL_NODEBUG_ALIAS = remove_cv_t<_Up>;
+};
+
+template <class _Up>
+struct __decay_impl<_Up, true>
+{
+public:
+  using type _CCCL_NODEBUG_ALIAS =
+    conditional_t<is_array_v<_Up>,
+                  remove_extent_t<_Up>*,
+                  conditional_t<is_function_v<_Up>, add_pointer_t<_Up>, remove_cv_t<_Up>>>;
+};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT decay
+{
+private:
+  using _Up _CCCL_NODEBUG_ALIAS = remove_reference_t<_Tp>;
+
+public:
+  using type _CCCL_NODEBUG_ALIAS = typename __decay_impl<_Up, __cccl_is_referenceable<_Up>::value>::type;
+};
+
+template <class _Tp>
+using decay_t _CCCL_NODEBUG_ALIAS = typename decay<_Tp>::type;
+
+#endif // !_CCCL_BUILTIN_DECAY
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_DECAY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/disjunction.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/disjunction.h
new file mode 100644
index 00000000..8464ffd2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/disjunction.h
@@ -0,0 +1,77 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_DISJUNCTION_H
+#define _CUDA_STD___TYPE_TRAITS_DISJUNCTION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <bool>
+struct _OrImpl;
+
+template <>
+struct _OrImpl<true>
+{
+  template <class _Res, class _First, class... _Rest>
+  using _Result _CCCL_NODEBUG_ALIAS =
+    typename _OrImpl<!bool(_First::value) && sizeof...(_Rest) != 0>::template _Result<_First, _Rest...>;
+};
+
+template <>
+struct _OrImpl<false>
+{
+  template <class _Res, class...>
+  using _Result = _Res;
+};
+
+// _Or always performs lazy evaluation of its arguments.
+//
+// However, `_Or<_Pred...>` itself will evaluate its result immediately (without having to
+// be instantiated) since it is an alias, unlike `disjunction<_Pred...>`, which is a struct.
+// If you want to defer the evaluation of `_Or<_Pred...>` itself, use `_Lazy<_Or, _Pred...>`
+// or `disjunction<_Pred...>` directly.
+template <class... _Args>
+using _Or _CCCL_NODEBUG_ALIAS = typename _OrImpl<sizeof...(_Args) != 0>::template _Result<false_type, _Args...>;
+
+#if _CCCL_COMPILER(MSVC)
+template <class... _Args>
+struct disjunction : false_type
+{};
+
+template <class _First, class... _Rest>
+struct disjunction<_First, _Rest...> : _OrImpl<true>::template _Result<false_type, _First, _Rest...>
+{};
+#else
+template <class... _Args>
+struct disjunction : _Or<_Args...>
+{};
+#endif // !_CCCL_COMPILER(MSVC)
+
+template <class... _Args>
+inline constexpr bool disjunction_v = _Or<_Args...>::value;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_DISJUNCTION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/enable_if.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/enable_if.h
new file mode 100644
index 00000000..0ac2c9bb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/enable_if.h
@@ -0,0 +1,43 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_ENABLE_IF_H
+#define _CUDA_STD___TYPE_TRAITS_ENABLE_IF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <bool, class _Tp = void>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT enable_if
+{};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT enable_if<true, _Tp>
+{
+  using type = _Tp;
+};
+
+template <bool _Bp, class _Tp = void>
+using enable_if_t _CCCL_NODEBUG_ALIAS = typename enable_if<_Bp, _Tp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_ENABLE_IF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/extent.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/extent.h
new file mode 100644
index 00000000..1f06e84b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/extent.h
@@ -0,0 +1,67 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_EXTENT_H
+#define _CUDA_STD___TYPE_TRAITS_EXTENT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(array_extent)
+#  define _CCCL_BUILTIN_ARRAY_EXTENT(...) __array_extent(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(array_extent)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_ARRAY_EXTENT)
+
+template <class _Tp, size_t _Ip = 0>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT extent : integral_constant<size_t, _CCCL_BUILTIN_ARRAY_EXTENT(_Tp, _Ip)>
+{};
+
+template <class _Tp, unsigned _Ip = 0>
+inline constexpr size_t extent_v = _CCCL_BUILTIN_ARRAY_EXTENT(_Tp, _Ip);
+
+#else // ^^^ _CCCL_BUILTIN_ARRAY_EXTENT ^^^ / vvv !_CCCL_BUILTIN_ARRAY_EXTENT vvv
+
+template <class _Tp, unsigned _Ip = 0>
+inline constexpr size_t extent_v = 0;
+template <class _Tp>
+inline constexpr size_t extent_v<_Tp[], 0> = 0;
+template <class _Tp, unsigned _Ip>
+inline constexpr size_t extent_v<_Tp[], _Ip> = extent_v<_Tp, _Ip - 1>;
+template <class _Tp, size_t _Np>
+inline constexpr size_t extent_v<_Tp[_Np], 0> = _Np;
+template <class _Tp, size_t _Np, unsigned _Ip>
+inline constexpr size_t extent_v<_Tp[_Np], _Ip> = extent_v<_Tp, _Ip - 1>;
+
+template <class _Tp, unsigned _Ip = 0>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT extent : integral_constant<size_t, extent_v<_Tp, _Ip>>
+{};
+
+#endif // !_CCCL_BUILTIN_ARRAY_EXTENT
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_EXTENT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/fold.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/fold.h
new file mode 100644
index 00000000..ce5a497b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/fold.h
@@ -0,0 +1,47 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_FOLD_H
+#define _CUDA_STD___TYPE_TRAITS_FOLD_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// Use fold expressions when possible to implement __fold_and[_v] and
+// __fold_or[_v].
+template <bool... _Preds>
+inline constexpr bool __fold_and_v = (_Preds && ...);
+
+template <bool... _Preds>
+inline constexpr bool __fold_or_v = (_Preds || ...);
+
+template <bool... _Preds>
+using __fold_and = bool_constant<bool((_Preds && ...))>; // cast to bool to avoid error from gcc < 8
+
+template <bool... _Preds>
+using __fold_or = bool_constant<bool((_Preds || ...))>; // cast to bool to avoid error from gcc < 8
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_FOLD_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/integral_constant.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/integral_constant.h
new file mode 100644
index 00000000..e226a993
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/integral_constant.h
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_INTEGRAL_CONSTANT_H
+#define _CUDA_STD___TYPE_TRAITS_INTEGRAL_CONSTANT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, _Tp __v>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT integral_constant
+{
+  static constexpr const _Tp value = __v;
+  using value_type                 = _Tp;
+  using type                       = integral_constant;
+  _CCCL_API constexpr operator value_type() const noexcept
+  {
+    return value;
+  }
+  _CCCL_API constexpr value_type operator()() const noexcept
+  {
+    return value;
+  }
+};
+
+using true_type  = integral_constant<bool, true>;
+using false_type = integral_constant<bool, false>;
+
+template <bool __b>
+using bool_constant = integral_constant<bool, __b>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_INTEGRAL_CONSTANT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_aggregate.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_aggregate.h
new file mode 100644
index 00000000..635f5f5f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_aggregate.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_AGGREGATE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_AGGREGATE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_AGGREGATE(...) __is_aggregate(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_aggregate : bool_constant<_CCCL_BUILTIN_IS_AGGREGATE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_aggregate_v = _CCCL_BUILTIN_IS_AGGREGATE(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_AGGREGATE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_arithmetic.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_arithmetic.h
new file mode 100644
index 00000000..36d91ed4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_arithmetic.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_ARITHMETIC_H
+#define _CUDA_STD___TYPE_TRAITS_IS_ARITHMETIC_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_floating_point.h>
+#include <cuda/std/__type_traits/is_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool is_arithmetic_v = is_integral_v<_Tp> || is_floating_point_v<_Tp>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_arithmetic : public bool_constant<is_arithmetic_v<_Tp>>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_ARITHMETIC_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_array.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_array.h
new file mode 100644
index 00000000..7dae5fd0
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_array.h
@@ -0,0 +1,62 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_ARRAY_H
+#define _CUDA_STD___TYPE_TRAITS_IS_ARRAY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// TODO: Clang incorrectly reports that __is_array is true for T[0].
+//       Re-enable the branch once https://llvm.org/PR54705 is fixed.
+#if defined(_CCCL_BUILTIN_IS_ARRAY) && !defined(_LIBCUDACXX_USE_IS_ARRAY_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_array : public bool_constant<_CCCL_BUILTIN_IS_ARRAY(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_array_v = _CCCL_BUILTIN_IS_ARRAY(_Tp);
+
+#else
+
+template <class _Tp>
+inline constexpr bool is_array_v = false;
+
+template <class _Tp>
+inline constexpr bool is_array_v<_Tp[]> = true;
+
+template <class _Tp, size_t _Np>
+inline constexpr bool is_array_v<_Tp[_Np]> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_array : public bool_constant<is_array_v<_Tp>>
+{};
+
+#endif // defined(_CCCL_BUILTIN_IS_ARRAY) && !defined(_LIBCUDACXX_USE_IS_ARRAY_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_ARRAY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_assignable.h
new file mode 100644
index 00000000..cb777f9f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_assignable.h
@@ -0,0 +1,78 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <typename, typename _Tp>
+struct __select_2nd
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+#if defined(_CCCL_BUILTIN_IS_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_ASSIGNABLE_FALLBACK)
+
+template <class _T1, class _T2>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_assignable : public integral_constant<bool, _CCCL_BUILTIN_IS_ASSIGNABLE(_T1, _T2)>
+{};
+
+template <class _T1, class _T2>
+inline constexpr bool is_assignable_v = _CCCL_BUILTIN_IS_ASSIGNABLE(_T1, _T2);
+
+#else
+
+template <class _Tp, class _Arg>
+_CCCL_API inline
+  typename __select_2nd<decltype((::cuda::std::declval<_Tp>() = ::cuda::std::declval<_Arg>())), true_type>::type
+  __is_assignable_test(int);
+
+template <class, class>
+_CCCL_API inline false_type __is_assignable_test(...);
+
+template <class _Tp, class _Arg, bool = is_void_v<_Tp> || is_void_v<_Arg>>
+struct __is_assignable_imp : public decltype((::cuda::std::__is_assignable_test<_Tp, _Arg>(0)))
+{};
+
+template <class _Tp, class _Arg>
+struct __is_assignable_imp<_Tp, _Arg, true> : public false_type
+{};
+
+template <class _Tp, class _Arg>
+struct is_assignable : public __is_assignable_imp<_Tp, _Arg>
+{};
+
+template <class _Tp, class _Arg>
+inline constexpr bool is_assignable_v = is_assignable<_Tp, _Arg>::value;
+
+#endif // defined(_CCCL_BUILTIN_IS_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_ASSIGNABLE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_base_of.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_base_of.h
new file mode 100644
index 00000000..4ab74d92
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_base_of.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_BASE_OF_H
+#define _CUDA_STD___TYPE_TRAITS_IS_BASE_OF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_BASE_OF(...) __is_base_of(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Bp, class _Dp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_base_of : bool_constant<_CCCL_BUILTIN_IS_BASE_OF(_Bp, _Dp)>
+{};
+
+template <class _Bp, class _Dp>
+inline constexpr bool is_base_of_v = _CCCL_BUILTIN_IS_BASE_OF(_Bp, _Dp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_BASE_OF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_callable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_callable.h
new file mode 100644
index 00000000..3becceed
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_callable.h
@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_CALLABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_CALLABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_valid_expansion.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Func, class... _Args>
+using __call_result_t _CCCL_NODEBUG_ALIAS = // NOLINT(modernize-type-traits)
+  decltype(::cuda::std::declval<_Func>()(::cuda::std::declval<_Args>()...)); // NOLINT(modernize-type-traits)
+
+template <class _Func, class... _Args>
+struct __is_callable : _IsValidExpansion<__call_result_t, _Func, _Args...>
+{};
+
+template <class _Func, class... _Args>
+inline constexpr bool __is_callable_v = _IsValidExpansion<__call_result_t, _Func, _Args...>::value;
+
+namespace detail
+{
+template <class _Func, class... _Args>
+using __if_nothrow_callable_t _CCCL_NODEBUG_ALIAS =
+  ::cuda::std::enable_if_t<noexcept(::cuda::std::declval<_Func>()(::cuda::std::declval<_Args>()...))>;
+} // namespace detail
+
+template <class _Func, class... _Args>
+struct __is_nothrow_callable : _IsValidExpansion<detail::__if_nothrow_callable_t, _Func, _Args...>
+{};
+
+template <class _Func, class... _Args>
+inline constexpr bool __is_nothrow_callable_v =
+  _IsValidExpansion<detail::__if_nothrow_callable_t, _Func, _Args...>::value;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_CALLABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_class.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_class.h
new file mode 100644
index 00000000..336acae9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_class.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_CLASS_H
+#define _CUDA_STD___TYPE_TRAITS_IS_CLASS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_CLASS(...) __is_class(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_class : public integral_constant<bool, _CCCL_BUILTIN_IS_CLASS(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_class_v = _CCCL_BUILTIN_IS_CLASS(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_CLASS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_comparable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_comparable.h
new file mode 100644
index 00000000..fa15e545
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_comparable.h
@@ -0,0 +1,79 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_COMPARABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_COMPARABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__concepts/boolean_testable.h>
+#include <cuda/std/__type_traits/make_const_lvalue_ref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4389) // '==': signed/unsigned mismatch
+_CCCL_DIAG_SUPPRESS_MSVC(4018) // '<': signed/unsigned mismatch
+_CCCL_DIAG_SUPPRESS_MSVC(4244) // '==': conversion from 'const double' to 'int', possible loss of data
+
+template <class _Tp, class _Up, class = void>
+inline constexpr bool __is_cpp17_equality_comparable_v = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __is_cpp17_equality_comparable_v<
+  _Tp,
+  _Up,
+  void_t<decltype(::cuda::std::declval<__make_const_lvalue_ref<_Tp>>()
+                  == ::cuda::std::declval<__make_const_lvalue_ref<_Up>>())>> =
+  __boolean_testable<decltype(::cuda::std::declval<__make_const_lvalue_ref<_Tp>>()
+                              == ::cuda::std::declval<__make_const_lvalue_ref<_Up>>())>;
+
+template <class _Tp, class _Up, bool = __is_cpp17_equality_comparable_v<_Tp, _Up>>
+inline constexpr bool __is_cpp17_nothrow_equality_comparable_v = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __is_cpp17_nothrow_equality_comparable_v<_Tp, _Up, true> = noexcept(
+  ::cuda::std::declval<__make_const_lvalue_ref<_Tp>>() == ::cuda::std::declval<__make_const_lvalue_ref<_Up>>());
+
+template <class _Tp, class _Up, class = void>
+inline constexpr bool __is_cpp17_less_than_comparable_v = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __is_cpp17_less_than_comparable_v<
+  _Tp,
+  _Up,
+  void_t<decltype(::cuda::std::declval<__make_const_lvalue_ref<_Tp>>()
+                  < ::cuda::std::declval<__make_const_lvalue_ref<_Up>>())>> =
+  __boolean_testable<decltype(::cuda::std::declval<__make_const_lvalue_ref<_Tp>>()
+                              < ::cuda::std::declval<__make_const_lvalue_ref<_Up>>())>;
+
+template <class _Tp, class _Up, bool = __is_cpp17_less_than_comparable_v<_Tp, _Up>>
+inline constexpr bool __is_cpp17_nothrow_less_than_comparable_v = false;
+
+template <class _Tp, class _Up>
+inline constexpr bool __is_cpp17_nothrow_less_than_comparable_v<_Tp, _Up, true> =
+  noexcept(::cuda::std::declval<__make_const_lvalue_ref<_Tp>>() < ::cuda::std::declval<__make_const_lvalue_ref<_Up>>());
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_DIAG_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_COMPARABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_complete.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_complete.h
new file mode 100644
index 00000000..52946782
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_complete.h
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_COMPLETE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_COMPLETE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/void_t.h> // IWYU pragma: keep
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#ifdef _CCCL_BUILTIN_IS_COMPLETE_TYPE
+
+template <typename T>
+inline constexpr bool __is_complete_v = _CCCL_BUILTIN_IS_COMPLETE_TYPE(T);
+
+#else // ^^^  _CCCL_BUILTIN_IS_COMPLETE_TYPE ^^^ / vvv no builtin vvv
+// Must be a SFINAE trait instead of
+//
+// template <typename T>
+// inline constexpr __is_complete_v = sizeof(T) > 0;
+//
+// Because older NVCC doesn't even allow you to utter the phrase sizeof(T) if T is incomplete
+template <typename T, typename = void>
+inline constexpr bool __is_complete_v = false;
+
+template <typename T>
+inline constexpr bool __is_complete_v<T, void_t<decltype(sizeof(T))>> = (sizeof(T) > 0);
+#endif // ^^^ no builtin ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_COMPLETE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_const.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_const.h
new file mode 100644
index 00000000..ffdb142e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_const.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_CONST_H
+#define _CUDA_STD___TYPE_TRAITS_IS_CONST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_const)
+#  define _CCCL_BUILTIN_IS_CONST(...) __is_const(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_const)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_CONST)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_const : bool_constant<_CCCL_BUILTIN_IS_CONST(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_const_v = _CCCL_BUILTIN_IS_CONST(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_CONST ^^^ / vvv !_CCCL_BUILTIN_IS_CONST vvv
+
+template <class _Tp>
+inline constexpr bool is_const_v = false;
+
+template <class _Tp>
+inline constexpr bool is_const_v<const _Tp> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_const : public bool_constant<is_const_v<_Tp>>
+{};
+
+#endif // !_CCCL_BUILTIN_IS_CONST
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_CONST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_constructible.h
new file mode 100644
index 00000000..f8df56b4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_constructible.h
@@ -0,0 +1,174 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_IS_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_IS_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conjunction.h>
+#include <cuda/std/__type_traits/disjunction.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_destructible.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/negation.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+namespace __is_construct
+{
+struct __nat
+{};
+} // namespace __is_construct
+
+// FIXME: This logic isn't awesome.
+#if (!defined(_CCCL_BUILTIN_IS_CONSTRUCTIBLE) || defined(_LIBCUDACXX_TESTING_FALLBACK_IS_CONSTRUCTIBLE) \
+     || defined(_LIBCUDACXX_USE_IS_CONSTRUCTIBLE_FALLBACK))
+
+template <class _Tp, class... _Args>
+struct __cccl_is_constructible;
+
+template <class _To, class _From>
+struct __is_invalid_base_to_derived_cast
+{
+  static_assert(is_reference_v<_To>, "Wrong specialization");
+  using _RawFrom = remove_cvref_t<_From>;
+  using _RawTo   = remove_cvref_t<_To>;
+  static const bool value =
+    _And<_IsNotSame<_RawFrom, _RawTo>, is_base_of<_RawFrom, _RawTo>, _Not<__cccl_is_constructible<_RawTo, _From>>>::value;
+};
+
+template <class _To, class _From>
+struct __is_invalid_lvalue_to_rvalue_cast : false_type
+{
+  static_assert(is_reference_v<_To>, "Wrong specialization");
+};
+
+template <class _ToRef, class _FromRef>
+struct __is_invalid_lvalue_to_rvalue_cast<_ToRef&&, _FromRef&>
+{
+  using _RawFrom = remove_cvref_t<_FromRef>;
+  using _RawTo   = remove_cvref_t<_ToRef>;
+  static const bool value =
+    _And<_Not<is_function<_RawTo>>, _Or<_IsSame<_RawFrom, _RawTo>, is_base_of<_RawTo, _RawFrom>>>::value;
+};
+
+struct __is_constructible_helper
+{
+  template <class _To>
+  _CCCL_API inline static void __eat(_To);
+
+  // This overload is needed to work around a Clang bug that disallows
+  // static_cast<T&&>(e) for non-reference-compatible types.
+  // Example: static_cast<int&&>(declval<double>());
+  // NOTE: The static_cast implementation below is required to support
+  //  classes with explicit conversion operators.
+  template <class _To, class _From, class = decltype(__eat<_To>(::cuda::std::declval<_From>()))>
+  _CCCL_API inline static true_type __test_cast(int);
+
+  template <class _To, class _From, class = decltype(static_cast<_To>(::cuda::std::declval<_From>()))>
+  _CCCL_API inline static integral_constant<
+    bool,
+    !__is_invalid_base_to_derived_cast<_To, _From>::value && !__is_invalid_lvalue_to_rvalue_cast<_To, _From>::value>
+  __test_cast(long);
+
+  template <class, class>
+  _CCCL_API inline static false_type __test_cast(...);
+
+  template <class _Tp, class... _Args, class = decltype(_Tp(::cuda::std::declval<_Args>()...))>
+  _CCCL_API inline static true_type __test_nary(int);
+  template <class _Tp, class...>
+  _CCCL_API inline static false_type __test_nary(...);
+
+  template <class _Tp, class _A0, class = decltype(::new _Tp(::cuda::std::declval<_A0>()))>
+  _CCCL_API inline static is_destructible<_Tp> __test_unary(int);
+  template <class, class>
+  _CCCL_API inline static false_type __test_unary(...);
+};
+
+template <class _Tp, bool = is_void_v<_Tp>>
+struct __is_default_constructible : decltype(__is_constructible_helper::__test_nary<_Tp>(0))
+{};
+
+template <class _Tp>
+struct __is_default_constructible<_Tp, true> : false_type
+{};
+
+template <class _Tp>
+struct __is_default_constructible<_Tp[], false> : false_type
+{};
+
+template <class _Tp, size_t _Nx>
+struct __is_default_constructible<_Tp[_Nx], false> : __is_default_constructible<remove_all_extents_t<_Tp>>
+{};
+
+template <class _Tp, class... _Args>
+struct __cccl_is_constructible
+{
+  static_assert(sizeof...(_Args) > 1, "Wrong specialization");
+  using type = decltype(__is_constructible_helper::__test_nary<_Tp, _Args...>(0));
+};
+
+template <class _Tp>
+struct __cccl_is_constructible<_Tp> : __is_default_constructible<_Tp>
+{};
+
+template <class _Tp, class _A0>
+struct __cccl_is_constructible<_Tp, _A0> : public decltype(__is_constructible_helper::__test_unary<_Tp, _A0>(0))
+{};
+
+template <class _Tp, class _A0>
+struct __cccl_is_constructible<_Tp&, _A0> : public decltype(__is_constructible_helper::__test_cast<_Tp&, _A0>(0))
+{};
+
+template <class _Tp, class _A0>
+struct __cccl_is_constructible<_Tp&&, _A0> : public decltype(__is_constructible_helper::__test_cast<_Tp&&, _A0>(0))
+{};
+
+#endif
+
+#if defined(_CCCL_BUILTIN_IS_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_CONSTRUCTIBLE_FALLBACK)
+template <class _Tp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_constructible : public integral_constant<bool, _CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, _Args...)>
+{};
+
+template <class _Tp, class... _Args>
+inline constexpr bool is_constructible_v = _CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, _Args...);
+
+#else
+template <class _Tp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_constructible : public __cccl_is_constructible<_Tp, _Args...>::type
+{};
+
+template <class _Tp, class... _Args>
+inline constexpr bool is_constructible_v = is_constructible<_Tp, _Args...>::value;
+
+#endif // defined(_CCCL_BUILTIN_IS_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_CONSTRUCTIBLE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_IS_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_convertible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_convertible.h
new file mode 100644
index 00000000..3feead8e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_convertible.h
@@ -0,0 +1,211 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+// SPDX-FileCopyrightText: Copyright (c) Microsoft Corporation.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_CONVERTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_CONVERTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_function.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_CONVERTIBLE_TO) && !defined(_LIBCUDACXX_USE_IS_CONVERTIBLE_FALLBACK)
+
+template <class _T1, class _T2>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_convertible : public integral_constant<bool, _CCCL_BUILTIN_IS_CONVERTIBLE_TO(_T1, _T2)>
+{};
+
+template <class _T1, class _T2>
+inline constexpr bool is_convertible_v = _CCCL_BUILTIN_IS_CONVERTIBLE_TO(_T1, _T2);
+
+#  if _CCCL_COMPILER(MSVC) // Workaround for DevCom-1627396
+template <class _Ty>
+struct is_convertible<_Ty&, volatile _Ty&> : true_type
+{};
+
+template <class _Ty>
+struct is_convertible<volatile _Ty&, volatile _Ty&> : true_type
+{};
+
+template <class _Ty>
+struct is_convertible<_Ty&, const volatile _Ty&> : true_type
+{};
+
+template <class _Ty>
+struct is_convertible<volatile _Ty&, const volatile _Ty&> : true_type
+{};
+
+template <class _Ty>
+inline constexpr bool is_convertible_v<_Ty&, volatile _Ty&> = true;
+
+template <class _Ty>
+inline constexpr bool is_convertible_v<volatile _Ty&, volatile _Ty&> = true;
+
+template <class _Ty>
+inline constexpr bool is_convertible_v<_Ty&, const volatile _Ty&> = true;
+
+template <class _Ty>
+inline constexpr bool is_convertible_v<volatile _Ty&, const volatile _Ty&> = true;
+#  endif // _CCCL_COMPILER(MSVC)
+
+#else // ^^^ _CCCL_BUILTIN_IS_CONVERTIBLE_TO ^^^ / vvv !_CCCL_BUILTIN_IS_CONVERTIBLE_TO vvv
+
+namespace __is_convertible_imp
+{
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_NVHPC(volatile_func_param_deprecated)
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(volatile_func_param_deprecated)
+_CCCL_DIAG_SUPPRESS_CLANG("-Wdeprecated-volatile")
+
+template <class _Tp>
+_CCCL_API inline void __test_convert(_Tp);
+
+_CCCL_END_NV_DIAG_SUPPRESS()
+_CCCL_DIAG_POP
+
+template <class _From, class _To, class = void>
+struct __is_convertible_test : public false_type
+{};
+
+template <class _From, class _To>
+struct __is_convertible_test<
+  _From,
+  _To,
+  decltype(::cuda::std::__is_convertible_imp::__test_convert<_To>(::cuda::std::declval<_From>()))> : public true_type
+{};
+
+template <class _Tp, bool _IsArray = is_array_v<_Tp>, bool _IsFunction = is_function_v<_Tp>, bool _IsVoid = is_void_v<_Tp>>
+struct __is_array_function_or_void
+{
+  enum
+  {
+    value = 0
+  };
+};
+template <class _Tp>
+struct __is_array_function_or_void<_Tp, true, false, false>
+{
+  enum
+  {
+    value = 1
+  };
+};
+template <class _Tp>
+struct __is_array_function_or_void<_Tp, false, true, false>
+{
+  enum
+  {
+    value = 2
+  };
+};
+template <class _Tp>
+struct __is_array_function_or_void<_Tp, false, false, true>
+{
+  enum
+  {
+    value = 3
+  };
+};
+} // namespace __is_convertible_imp
+
+template <class _Tp, unsigned = __is_convertible_imp::__is_array_function_or_void<remove_reference_t<_Tp>>::value>
+struct __is_convertible_check
+{
+  static const size_t __v = 0;
+};
+
+template <class _Tp>
+struct __is_convertible_check<_Tp, 0>
+{
+  static const size_t __v = sizeof(_Tp);
+};
+
+template <class _T1,
+          class _T2,
+          unsigned _T1_is_array_function_or_void = __is_convertible_imp::__is_array_function_or_void<_T1>::value,
+          unsigned _T2_is_array_function_or_void = __is_convertible_imp::__is_array_function_or_void<_T2>::value>
+struct __is_convertible_fallback
+    : public integral_constant<bool, __is_convertible_imp::__is_convertible_test<_T1, _T2>::value>
+{};
+
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 0, 1> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 1, 1> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 2, 1> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 3, 1> : public false_type
+{};
+
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 0, 2> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 1, 2> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 2, 2> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 3, 2> : public false_type
+{};
+
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 0, 3> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 1, 3> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 2, 3> : public false_type
+{};
+template <class _T1, class _T2>
+struct __is_convertible_fallback<_T1, _T2, 3, 3> : public true_type
+{};
+
+template <class _T1, class _T2>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_convertible : public __is_convertible_fallback<_T1, _T2>
+{
+  static const size_t __complete_check1 = __is_convertible_check<_T1>::__v;
+  static const size_t __complete_check2 = __is_convertible_check<_T2>::__v;
+};
+
+template <class _From, class _To>
+inline constexpr bool is_convertible_v = is_convertible<_From, _To>::value;
+
+#endif // !_CCCL_BUILTIN_IS_CONVERTIBLE_TO
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_CONVERTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_assignable.h
new file mode 100644
index 00000000..498b470c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_assignable.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_COPY_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_COPY_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_const.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_ASSIGNABLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_copy_assignable
+    : public bool_constant<_CCCL_BUILTIN_IS_ASSIGNABLE(
+        add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<add_const_t<_Tp>>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_copy_assignable_v =
+  _CCCL_BUILTIN_IS_ASSIGNABLE(add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<add_const_t<_Tp>>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_copy_assignable : public is_assignable<add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<add_const_t<_Tp>>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_copy_assignable_v =
+  is_assignable<add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<add_const_t<_Tp>>>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_COPY_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_constructible.h
new file mode 100644
index 00000000..c382c9d6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_copy_constructible.h
@@ -0,0 +1,58 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_COPY_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_COPY_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_const.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_copy_constructible
+    : public bool_constant<_CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, add_lvalue_reference_t<add_const_t<_Tp>>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_copy_constructible_v =
+  _CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, add_lvalue_reference_t<add_const_t<_Tp>>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_copy_constructible : public is_constructible<_Tp, add_lvalue_reference_t<add_const_t<_Tp>>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_copy_constructible_v = is_constructible<_Tp, add_lvalue_reference_t<add_const_t<_Tp>>>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_COPY_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_core_convertible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_core_convertible.h
new file mode 100644
index 00000000..6027f70a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_core_convertible.h
@@ -0,0 +1,47 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_CORE_CONVERTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_CORE_CONVERTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// [conv.general]/3 says "E is convertible to T" whenever "T t=E;" is well-formed.
+// We can't test for that, but we can test implicit convertibility by passing it
+// to a function. Notice that __is_core_convertible<void,void> is false,
+// and __is_core_convertible<immovable-type,immovable-type> is true in C++17 and later.
+
+template <class _Tp, class _Up, class = void>
+struct __is_core_convertible : public false_type
+{};
+
+template <class _Tp, class _Up>
+struct __is_core_convertible<_Tp, _Up, decltype(static_cast<void (*)(_Up)>(0)(static_cast<_Tp (*)()>(0)()))>
+    : public true_type
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_CORE_CONVERTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_default_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_default_constructible.h
new file mode 100644
index 00000000..eb0ef039
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_default_constructible.h
@@ -0,0 +1,40 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_DEFAULT_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_DEFAULT_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_default_constructible : public is_constructible<_Tp>
+{};
+
+template <class _Tp>
+inline constexpr bool is_default_constructible_v = is_constructible_v<_Tp>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_DEFAULT_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_destructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_destructible.h
new file mode 100644
index 00000000..5d453185
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_destructible.h
@@ -0,0 +1,115 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_DESTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_DESTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_function.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_all_extents.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_DESTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_DESTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_destructible : public integral_constant<bool, _CCCL_BUILTIN_IS_DESTRUCTIBLE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_destructible_v = _CCCL_BUILTIN_IS_DESTRUCTIBLE(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_DESTRUCTIBLE ^^^ / vvv !_CCCL_BUILTIN_IS_DESTRUCTIBLE vvv
+
+//  if it's a reference, return true
+//  if it's a function, return false
+//  if it's   void,     return false
+//  if it's an array of unknown bound, return false
+//  Otherwise, return "declval<_Up&>().~_Up()" is well-formed
+//    where _Up is remove_all_extents<_Tp>::type
+
+template <class>
+struct __is_destructible_apply
+{
+  using type = int;
+};
+
+template <typename _Tp>
+struct __is_destructor_wellformed
+{
+  template <typename _Tp1>
+  _CCCL_API inline static true_type
+    __test(typename __is_destructible_apply<decltype(::cuda::std::declval<_Tp1&>().~_Tp1())>::type);
+
+  template <typename _Tp1>
+  _CCCL_API inline static false_type __test(...);
+
+  static const bool value = decltype(__test<_Tp>(12))::value;
+};
+
+template <class _Tp, bool>
+struct __destructible_imp;
+
+template <class _Tp>
+struct __destructible_imp<_Tp, false>
+    : public integral_constant<bool, __is_destructor_wellformed<remove_all_extents_t<_Tp>>::value>
+{};
+
+template <class _Tp>
+struct __destructible_imp<_Tp, true> : public true_type
+{};
+
+template <class _Tp, bool>
+struct __destructible_false;
+
+template <class _Tp>
+struct __destructible_false<_Tp, false> : public __destructible_imp<_Tp, is_reference<_Tp>::value>
+{};
+
+template <class _Tp>
+struct __destructible_false<_Tp, true> : public false_type
+{};
+
+template <class _Tp>
+struct is_destructible : public __destructible_false<_Tp, is_function<_Tp>::value>
+{};
+
+template <class _Tp>
+struct is_destructible<_Tp[]> : public false_type
+{};
+
+template <>
+struct is_destructible<void> : public false_type
+{};
+
+template <class _Tp>
+inline constexpr bool is_destructible_v = is_destructible<_Tp>::value;
+
+#endif // !_CCCL_BUILTIN_IS_DESTRUCTIBLE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_DESTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_empty.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_empty.h
new file mode 100644
index 00000000..70301a28
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_empty.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_EMPTY_H
+#define _CUDA_STD___TYPE_TRAITS_IS_EMPTY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_EMPTY(...) __is_empty(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_empty : bool_constant<_CCCL_BUILTIN_IS_EMPTY(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_empty_v = _CCCL_BUILTIN_IS_EMPTY(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_EMPTY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_enum.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_enum.h
new file mode 100644
index 00000000..3df0dce1
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_enum.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_ENUM_H
+#define _CUDA_STD___TYPE_TRAITS_IS_ENUM_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_ENUM(...) __is_enum(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_enum : bool_constant<_CCCL_BUILTIN_IS_ENUM(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_enum_v = _CCCL_BUILTIN_IS_ENUM(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_ENUM_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_extended_floating_point.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_extended_floating_point.h
new file mode 100644
index 00000000..a44367e5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_extended_floating_point.h
@@ -0,0 +1,84 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_EXTENDED_FLOATING_POINT_H
+#define _CUDA_STD___TYPE_TRAITS_IS_EXTENDED_FLOATING_POINT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __is_extended_floating_point_v = false;
+
+#if _CCCL_HAS_NVFP16()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__half> = true;
+#endif // _CCCL_HAS_NVFP16
+
+#if _CCCL_HAS_NVBF16()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_bfloat16> = true;
+#endif // _CCCL_HAS_NVBF16
+
+#if _CCCL_HAS_NVFP8_E4M3()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp8_e4m3> = true;
+#endif // _CCCL_HAS_NVFP8_E4M3()
+
+#if _CCCL_HAS_NVFP8_E5M2()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp8_e5m2> = true;
+#endif // _CCCL_HAS_NVFP8_E5M2()
+
+#if _CCCL_HAS_NVFP8_E8M0()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp8_e8m0> = true;
+#endif // _CCCL_HAS_NVFP8_E8M0()
+
+#if _CCCL_HAS_NVFP6_E2M3()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp6_e2m3> = true;
+#endif // _CCCL_HAS_NVFP6_E2M3()
+
+#if _CCCL_HAS_NVFP6_E3M2()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp6_e3m2> = true;
+#endif // _CCCL_HAS_NVFP6_E3M2()
+
+#if _CCCL_HAS_NVFP4_E2M1()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__nv_fp4_e2m1> = true;
+#endif // _CCCL_HAS_NVFP4_E2M1()
+
+#if _CCCL_HAS_FLOAT128()
+template <>
+inline constexpr bool __is_extended_floating_point_v<__float128> = true;
+#endif // _CCCL_HAS_FLOAT128()
+
+template <class _Tp>
+inline constexpr bool __is_cv_extended_floating_point_v = __is_extended_floating_point_v<::cuda::std::remove_cv_t<_Tp>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_EXTENDED_FLOATING_POINT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_floating_point.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_floating_point.h
new file mode 100644
index 00000000..656cb795
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_floating_point.h
@@ -0,0 +1,53 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_FLOATING_POINT_H
+#define _CUDA_STD___TYPE_TRAITS_IS_FLOATING_POINT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __cccl_is_floating_point_v = false;
+
+template <>
+inline constexpr bool __cccl_is_floating_point_v<float> = true;
+
+template <>
+inline constexpr bool __cccl_is_floating_point_v<double> = true;
+
+template <>
+inline constexpr bool __cccl_is_floating_point_v<long double> = true;
+
+template <class _Tp>
+inline constexpr bool is_floating_point_v = __cccl_is_floating_point_v<remove_cv_t<_Tp>>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_floating_point : bool_constant<is_floating_point_v<_Tp>>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_FLOATING_POINT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_function.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_function.h
new file mode 100644
index 00000000..60134354
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_function.h
@@ -0,0 +1,64 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_FUNCTIONAL_H
+#define _CUDA_STD___TYPE_TRAITS_IS_FUNCTIONAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_const.h>
+#include <cuda/std/__type_traits/is_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_function)
+#  define _CCCL_BUILTIN_IS_FUNCTION(...) __is_function(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_function)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_FUNCTION)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_function : bool_constant<_CCCL_BUILTIN_IS_FUNCTION(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_function_v = _CCCL_BUILTIN_IS_FUNCTION(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_FUNCTION ^^^ / vvv !_CCCL_BUILTIN_IS_FUNCTION vvv
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4180) // qualifier applied to function type has no meaning; ignored
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_function : bool_constant<!(is_reference_v<_Tp> || is_const_v<const _Tp>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_function_v = !(is_reference_v<_Tp> || is_const_v<const _Tp>);
+
+_CCCL_DIAG_POP
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_FUNCTION ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_FUNCTIONAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_implicitly_default_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_implicitly_default_constructible.h
new file mode 100644
index 00000000..639e10a9
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_implicitly_default_constructible.h
@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_IMPLICITLY_DEFAULT_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_IMPLICITLY_DEFAULT_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_default_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_SUPPRESS_DEPRECATED_PUSH
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// First of all, we can't implement this check in C++03 mode because the {}
+// default initialization syntax isn't valid.
+// Second, we implement the trait in a funny manner with two defaulted template
+// arguments to workaround Clang's PR43454.
+template <class _Tp>
+_CCCL_API inline void __test_implicit_default_constructible(_Tp);
+
+template <class _Tp, class = void, class = typename is_default_constructible<_Tp>::type>
+struct __is_implicitly_default_constructible : false_type
+{};
+
+template <class _Tp>
+struct __is_implicitly_default_constructible<_Tp,
+                                             decltype(__test_implicit_default_constructible<_Tp const&>({})),
+                                             true_type> : true_type
+{};
+
+template <class _Tp>
+struct __is_implicitly_default_constructible<_Tp,
+                                             decltype(__test_implicit_default_constructible<_Tp const&>({})),
+                                             false_type> : false_type
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+_CCCL_SUPPRESS_DEPRECATED_POP
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_IMPLICITLY_DEFAULT_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_integral.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_integral.h
new file mode 100644
index 00000000..c4779cf1
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_integral.h
@@ -0,0 +1,127 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_INTEGRAL_H
+#define _CUDA_STD___TYPE_TRAITS_IS_INTEGRAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_integral)
+#  define _CCCL_BUILTIN_IS_INTEGRAL(...) __is_integral(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_integral)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_INTEGRAL)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_integral : bool_constant<_CCCL_BUILTIN_IS_INTEGRAL(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_integral_v = _CCCL_BUILTIN_IS_INTEGRAL(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_INTEGRAL ^^^ / vvv !_CCCL_BUILTIN_IS_INTEGRAL vvv
+
+template <class _Tp>
+inline constexpr bool __cccl_is_integral_helper_v = false;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<bool> = true;
+
+// char types
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<char> = true;
+
+#  if _CCCL_HAS_CHAR8_T()
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<char8_t> = true;
+#  endif // _CCCL_HAS_CHAR8_T()
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<char16_t> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<char32_t> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<wchar_t> = true;
+
+// signed integer types
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<signed char> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<short> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<int> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<long> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<long long> = true;
+
+#  if _CCCL_HAS_INT128()
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<__int128_t> = true;
+#  endif // _CCCL_HAS_INT128()
+
+// unsigned integer types
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<unsigned char> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<unsigned short> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<unsigned int> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<unsigned long> = true;
+
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<unsigned long long> = true;
+
+#  if _CCCL_HAS_INT128()
+template <>
+inline constexpr bool __cccl_is_integral_helper_v<__uint128_t> = true;
+#  endif // _CCCL_HAS_INT128()
+
+template <class _Tp>
+inline constexpr bool is_integral_v = __cccl_is_integral_helper_v<remove_cv_t<_Tp>>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_integral : public bool_constant<is_integral_v<_Tp>>
+{};
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_INTEGRAL ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_INTEGRAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_function_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_function_pointer.h
new file mode 100644
index 00000000..553ca95e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_function_pointer.h
@@ -0,0 +1,79 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_MEMBER_FUNCTION_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_MEMBER_FUNCTION_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_function.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct __cccl_is_member_pointer
+{
+  enum
+  {
+    __is_member = false,
+    __is_func   = false,
+    __is_obj    = false
+  };
+};
+template <class _Tp, class _Up>
+struct __cccl_is_member_pointer<_Tp _Up::*>
+{
+  enum
+  {
+    __is_member = true,
+    __is_func   = is_function_v<_Tp>,
+    __is_obj    = !__is_func,
+  };
+};
+
+#if defined(_CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER) && !defined(_LIBCUDACXX_USE_IS_MEMBER_FUNCTION_POINTER_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_function_pointer : public integral_constant<bool, _CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_function_pointer_v = _CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER ^^^ / vvv !_CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER vvv
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_function_pointer : public integral_constant<bool, __cccl_is_member_pointer<remove_cv_t<_Tp>>::__is_func>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_function_pointer_v = is_member_function_pointer<_Tp>::value;
+
+#endif // !_CCCL_BUILTIN_IS_MEMBER_FUNCTION_POINTER
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_MEMBER_FUNCTION_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_object_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_object_pointer.h
new file mode 100644
index 00000000..f97c068b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_object_pointer.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_MEMBER_OBJECT_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_MEMBER_OBJECT_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_member_function_pointer.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER) && !defined(_LIBCUDACXX_USE_IS_MEMBER_OBJECT_POINTER_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_object_pointer : public integral_constant<bool, _CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_object_pointer_v = _CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER ^^^ / vvv !_CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER vvv
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_object_pointer : public integral_constant<bool, __cccl_is_member_pointer<remove_cv_t<_Tp>>::__is_obj>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_object_pointer_v = is_member_object_pointer<_Tp>::value;
+
+#endif // !_CCCL_BUILTIN_IS_MEMBER_OBJECT_POINTER
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_MEMBER_FUNCTION_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_pointer.h
new file mode 100644
index 00000000..cd9ea8ec
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_member_pointer.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_MEMBER_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_MEMBER_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_member_function_pointer.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_MEMBER_POINTER) && !defined(_LIBCUDACXX_USE_IS_MEMBER_POINTER_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_pointer : public integral_constant<bool, _CCCL_BUILTIN_IS_MEMBER_POINTER(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_pointer_v = _CCCL_BUILTIN_IS_MEMBER_POINTER(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_MEMBER_POINTER ^^^ / vvv !_CCCL_BUILTIN_IS_MEMBER_POINTER vvv
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_member_pointer : public integral_constant<bool, __cccl_is_member_pointer<remove_cv_t<_Tp>>::__is_member>
+{};
+
+template <class _Tp>
+inline constexpr bool is_member_pointer_v = is_member_pointer<_Tp>::value;
+
+#endif // !_CCCL_BUILTIN_IS_MEMBER_POINTER
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_MEMBER_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_assignable.h
new file mode 100644
index 00000000..8c793e36
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_assignable.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_MOVE_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_MOVE_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/add_rvalue_reference.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_ASSIGNABLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_move_assignable
+    : public bool_constant<_CCCL_BUILTIN_IS_ASSIGNABLE(add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_move_assignable_v =
+  _CCCL_BUILTIN_IS_ASSIGNABLE(add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_move_assignable : public is_assignable<add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_move_assignable_v =
+  is_assignable<add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_MOVE_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_constructible.h
new file mode 100644
index 00000000..8dd9310c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_move_constructible.h
@@ -0,0 +1,56 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_MOVE_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_MOVE_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/add_rvalue_reference.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_move_constructible : public bool_constant<_CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, add_rvalue_reference_t<_Tp>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_move_constructible_v = _CCCL_BUILTIN_IS_CONSTRUCTIBLE(_Tp, add_rvalue_reference_t<_Tp>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_move_constructible : public is_constructible<_Tp, add_rvalue_reference_t<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_move_constructible_v = is_constructible<_Tp, add_rvalue_reference_t<_Tp>>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_MOVE_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_assignable.h
new file mode 100644
index 00000000..58941512
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_assignable.h
@@ -0,0 +1,70 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+#include <cuda/std/__type_traits/is_scalar.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_ASSIGNABLE_FALLBACK)
+
+template <class _Tp, class _Arg>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_assignable : public integral_constant<bool, _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(_Tp, _Arg)>
+{};
+
+template <class _Tp, class _Arg>
+inline constexpr bool is_nothrow_assignable_v = _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(_Tp, _Arg);
+
+#else // ^^^ _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE ^^^ / vvv !_CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE vvv
+
+template <bool, class _Tp, class _Arg>
+struct __cccl_is_nothrow_assignable;
+
+template <class _Tp, class _Arg>
+struct __cccl_is_nothrow_assignable<false, _Tp, _Arg> : public false_type
+{};
+
+template <class _Tp, class _Arg>
+struct __cccl_is_nothrow_assignable<true, _Tp, _Arg>
+    : public integral_constant<bool, noexcept(::cuda::std::declval<_Tp>() = ::cuda::std::declval<_Arg>())>
+{};
+
+template <class _Tp, class _Arg>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_assignable : public __cccl_is_nothrow_assignable<is_assignable_v<_Tp, _Arg>, _Tp, _Arg>
+{};
+
+template <class _Tp, class _Arg>
+inline constexpr bool is_nothrow_assignable_v = is_nothrow_assignable<_Tp, _Arg>::value;
+
+#endif // !_CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_constructible.h
new file mode 100644
index 00000000..b785da44
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_constructible.h
@@ -0,0 +1,72 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/is_scalar.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_constructible : public integral_constant<bool, _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, _Args...)>
+{};
+
+template <class _Tp, class... _Args>
+inline constexpr bool is_nothrow_constructible_v = _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, _Args...);
+
+#else
+
+template <bool, class _Tp, class... _Args>
+inline constexpr bool __cccl_is_nothrow_constructible = false;
+
+template <class _Tp, class... _Args>
+inline constexpr bool __cccl_is_nothrow_constructible<true, _Tp, _Args...> =
+  noexcept(_Tp(::cuda::std::declval<_Args>()...));
+
+template <class _Tp, class _Arg>
+inline constexpr bool __cccl_is_nothrow_constructible<true, _Tp, _Arg> =
+  noexcept(static_cast<_Tp>(::cuda::std::declval<_Arg>()));
+
+template <class _Tp, size_t _Ns>
+inline constexpr bool __cccl_is_nothrow_constructible<true, _Tp[_Ns]> = __cccl_is_nothrow_constructible<true, _Tp>;
+
+template <class _Tp, class... _Args>
+inline constexpr bool is_nothrow_constructible_v =
+  __cccl_is_nothrow_constructible<is_constructible_v<_Tp, _Args...>, _Tp, _Args...>;
+
+template <class _Tp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_constructible : bool_constant<is_nothrow_constructible_v<_Tp, _Args...>>
+{};
+
+#endif // defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_convertible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_convertible.h
new file mode 100644
index 00000000..fdaead64
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_convertible.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONVERTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONVERTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conjunction.h>
+#include <cuda/std/__type_traits/disjunction.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/lazy.h>
+#include <cuda/std/__utility/declval.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <typename _Tp>
+_CCCL_HOST_DEVICE static void __test_noexcept(_Tp) noexcept;
+
+template <typename _Fm, typename _To>
+_CCCL_HOST_DEVICE static bool_constant<noexcept(::cuda::std::__test_noexcept<_To>(::cuda::std::declval<_Fm>()))>
+__is_nothrow_convertible_test();
+
+template <typename _Fm, typename _To>
+struct __is_nothrow_convertible_helper : decltype(__is_nothrow_convertible_test<_Fm, _To>())
+{};
+
+template <typename _Fm, typename _To>
+struct is_nothrow_convertible
+    : _Or<_And<is_void<_To>, is_void<_Fm>>,
+          _Lazy<_And, is_convertible<_Fm, _To>, __is_nothrow_convertible_helper<_Fm, _To>>>::type
+{};
+
+template <typename _Fm, typename _To>
+inline constexpr bool is_nothrow_convertible_v = is_nothrow_convertible<_Fm, _To>::value;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_CONVERTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_assignable.h
new file mode 100644
index 00000000..31f81764
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_assignable.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_const.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_ASSIGNABLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_copy_assignable
+    : public integral_constant<bool,
+                               _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(
+                                 add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<typename add_const<_Tp>::type>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_copy_assignable_v = _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(
+  add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<typename add_const<_Tp>::type>);
+
+#else
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_copy_assignable
+    : public is_nothrow_assignable<add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<typename add_const<_Tp>::type>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_copy_assignable_v = is_nothrow_copy_assignable<_Tp>::value;
+
+#endif
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_constructible.h
new file mode 100644
index 00000000..44669d09
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_copy_constructible.h
@@ -0,0 +1,58 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_const.h>
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_copy_constructible
+    : public bool_constant<_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, add_lvalue_reference_t<add_const_t<_Tp>>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_copy_constructible_v =
+  _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, add_lvalue_reference_t<add_const_t<_Tp>>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_copy_constructible
+    : public is_nothrow_constructible<_Tp, add_lvalue_reference_t<typename add_const<_Tp>::type>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_copy_constructible_v = is_nothrow_copy_constructible<_Tp>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_COPY_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_default_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_default_constructible.h
new file mode 100644
index 00000000..6113304c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_default_constructible.h
@@ -0,0 +1,54 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DEFAULT_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DEFAULT_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp, class... _Args>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_default_constructible : public integral_constant<bool, _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp)>
+{};
+
+template <class _Tp, class... _Args>
+inline constexpr bool is_nothrow_default_constructible_v = _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp);
+
+#else
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_default_constructible : public is_nothrow_constructible<_Tp>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_default_constructible_v = is_nothrow_constructible<_Tp>::value;
+
+#endif // defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DEFAULT_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_destructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_destructible.h
new file mode 100644
index 00000000..b1c08d55
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_destructible.h
@@ -0,0 +1,82 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DESTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DESTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_destructible.h>
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/is_scalar.h>
+#include <cuda/std/__type_traits/remove_all_extents.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// is_nothrow_destructible
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_DESTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct is_nothrow_destructible : public integral_constant<bool, _CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_destructible_v = _CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE ^^^ / vvv !_CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE vvv
+
+template <class _Tp, bool = is_destructible<_Tp>::value>
+struct __cccl_is_nothrow_destructible : false_type
+{};
+
+template <class _Tp>
+struct __cccl_is_nothrow_destructible<_Tp, true>
+    : public integral_constant<bool, noexcept(::cuda::std::declval<_Tp>().~_Tp())>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_destructible : public __cccl_is_nothrow_destructible<_Tp>
+{};
+
+template <class _Tp, size_t _Ns>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_destructible<_Tp[_Ns]> : public is_nothrow_destructible<_Tp>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_destructible<_Tp&> : public true_type
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_destructible<_Tp&&> : public true_type
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_destructible_v = is_nothrow_destructible<_Tp>::value;
+
+#endif // !_CCCL_BUILTIN_IS_NOTHROW_DESTRUCTIBLE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_DESTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_assignable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_assignable.h
new file mode 100644
index 00000000..a45fde09
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_assignable.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_ASSIGNABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_ASSIGNABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/add_rvalue_reference.h>
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_ASSIGNABLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_move_assignable
+    : public integral_constant<
+        bool,
+        _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_move_assignable_v =
+  _CCCL_BUILTIN_IS_NOTHROW_ASSIGNABLE(add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>);
+
+#else
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_move_assignable : public is_nothrow_assignable<add_lvalue_reference_t<_Tp>, add_rvalue_reference_t<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_move_assignable_v = is_nothrow_move_assignable<_Tp>::value;
+
+#endif
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_ASSIGNABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_constructible.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_constructible.h
new file mode 100644
index 00000000..91a8e70d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_nothrow_move_constructible.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_CONSTRUCTIBLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_CONSTRUCTIBLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_rvalue_reference.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE) && !defined(_LIBCUDACXX_USE_IS_NOTHROW_CONSTRUCTIBLE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_move_constructible
+    : public bool_constant<_CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, add_rvalue_reference_t<_Tp>)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_move_constructible_v =
+  _CCCL_BUILTIN_IS_NOTHROW_CONSTRUCTIBLE(_Tp, add_rvalue_reference_t<_Tp>);
+
+#else // ^^^ Use builtin ^^^ / vvv No builtin
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_move_constructible : public is_nothrow_constructible<_Tp, add_rvalue_reference_t<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_nothrow_move_constructible_v = is_nothrow_move_constructible<_Tp>::value;
+
+#endif // No builtin
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NOTHROW_MOVE_CONSTRUCTIBLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_null_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_null_pointer.h
new file mode 100644
index 00000000..377f21fb
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_null_pointer.h
@@ -0,0 +1,43 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_NULL_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_NULL_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool is_null_pointer_v = is_same_v<remove_cv_t<_Tp>, nullptr_t>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_null_pointer : public bool_constant<is_null_pointer_v<_Tp>>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_NULL_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_object.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_object.h
new file mode 100644
index 00000000..975ab617
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_object.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_OBJECT_H
+#define _CUDA_STD___TYPE_TRAITS_IS_OBJECT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_class.h>
+#include <cuda/std/__type_traits/is_scalar.h>
+#include <cuda/std/__type_traits/is_union.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_OBJECT) && !defined(_LIBCUDACXX_USE_IS_OBJECT_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_object : public bool_constant<_CCCL_BUILTIN_IS_OBJECT(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_object_v = _CCCL_BUILTIN_IS_OBJECT(_Tp);
+
+#else
+
+template <class _Tp>
+inline constexpr bool is_object_v = is_scalar_v<_Tp> || is_array_v<_Tp> || is_union_v<_Tp> || is_class_v<_Tp>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_object : public bool_constant<is_object_v<_Tp>>
+{};
+
+#endif // defined(_CCCL_BUILTIN_IS_OBJECT) && !defined(_LIBCUDACXX_USE_IS_OBJECT_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_OBJECT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_pointer.h
new file mode 100644
index 00000000..34d223d4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_pointer.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_POINTER) && !defined(_LIBCUDACXX_USE_IS_POINTER_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_pointer : public bool_constant<_CCCL_BUILTIN_IS_POINTER(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_pointer_v = _CCCL_BUILTIN_IS_POINTER(_Tp);
+
+#else
+
+template <class _Tp>
+inline constexpr bool __cccl_is_pointer_helper_v = false;
+
+template <class _Tp>
+inline constexpr bool __cccl_is_pointer_helper_v<_Tp*> = true;
+
+template <class _Tp>
+inline constexpr bool is_pointer_v = __cccl_is_pointer_helper_v<remove_cv_t<_Tp>>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_pointer : public bool_constant<is_pointer_v<_Tp>>
+{};
+
+#endif // defined(_CCCL_BUILTIN_IS_POINTER) && !defined(_LIBCUDACXX_USE_IS_POINTER_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_primary_template.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_primary_template.h
new file mode 100644
index 00000000..35945fac
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_primary_template.h
@@ -0,0 +1,121 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_PRIMARY_TEMPLATE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_PRIMARY_TEMPLATE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__fwd/iterator.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_pointer.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_valid_expansion.h>
+#include <cuda/std/__type_traits/void_t.h>
+
+#if _CCCL_HOSTED()
+#  include <iterator>
+#endif // _CCCL_HOSTED()
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_COMPILER(MSVC)
+
+template <class _Iter, class = void>
+struct __is_primary_cccl_template : false_type
+{};
+
+template <class _Iter>
+struct __is_primary_cccl_template<_Iter, void_t<typename iterator_traits<_Iter>::__cccl_primary_template>>
+    : public is_same<iterator_traits<_Iter>, typename iterator_traits<_Iter>::__cccl_primary_template>
+{};
+
+#else // ^^^ _CCCL_COMPILER(MSVC) ^^^ / vvv !_CCCL_COMPILER(MSVC) vvv
+
+template <class _Traits>
+using __test_for_primary_template = enable_if_t<_IsSame<_Traits, typename _Traits::__cccl_primary_template>::value>;
+template <class _Iter>
+using __is_primary_cccl_template = _IsValidExpansion<__test_for_primary_template, iterator_traits<_Iter>>;
+
+#endif // !_CCCL_COMPILER(MSVC)
+
+#if _CCCL_FREESTANDING()
+
+// No ::std::traits with freestanding implementations
+template <class _Iter>
+struct __is_primary_std_template : true_type
+{};
+
+template <class _Iter, class _OtherTraits>
+using __select_traits = conditional_t<__is_primary_cccl_template<_Iter>::value, _OtherTraits, iterator_traits<_Iter>>;
+
+#else // ^^^ _CCCL_FREESTANDING() ^^^ / vvv _CCCL_HOSTED() vvv
+
+// We also need to respect what the user is defining to std::iterator_traits
+#  if _CCCL_HOST_STD_LIB(LIBSTDCXX)
+// libstdc++ uses `is_base_of`
+template <class _Iter, bool>
+inline constexpr bool __is_primary_std_template_impl =
+  is_base_of_v<::std::__iterator_traits<_Iter>, ::std::iterator_traits<_Iter>>;
+template <class _Iter>
+inline constexpr bool __is_primary_std_template_impl<_Iter, true> = true;
+
+// This is needed because with a defaulted template argument subsumption fails for C++20 for concepts
+// that involve incrementable_traits
+template <class _Iter>
+struct __is_primary_std_template : bool_constant<__is_primary_std_template_impl<_Iter, is_pointer_v<_Iter>>>
+{};
+#  elif _CCCL_HOST_STD_LIB(LIBCXX)
+
+// libc++ uses the same mechanism than we do with __primary_template
+template <class _Traits>
+using __test_for_primary_std_template = enable_if_t<_IsSame<_Traits, typename _Traits::__primary_template>::value>;
+template <class _Iter>
+using __is_primary_std_template = _IsValidExpansion<__test_for_primary_std_template, ::std::iterator_traits<_Iter>>;
+
+#  elif _CCCL_HOST_STD_LIB(STL)
+// On MSVC we must check for the base class because `_From_primary` is only defined in C++20
+template <class _Iter, bool>
+inline constexpr bool __is_primary_std_template_impl =
+  is_base_of_v<::std::_Iterator_traits_base<_Iter>, ::std::iterator_traits<_Iter>>;
+template <class _Iter>
+inline constexpr bool __is_primary_std_template_impl<_Iter, true> = true;
+
+// This is needed because with a defaulted template argument subsumption fails for C++20 for concepts
+// that involve incrementable_traits
+template <class _Iter>
+struct __is_primary_std_template : bool_constant<__is_primary_std_template_impl<_Iter, is_pointer_v<_Iter>>>
+{};
+#  endif // ^^^ _CCCL_HOST_STD_LIB(STL) ^^^
+
+template <class _Iter, class _OtherTraits>
+using __select_traits =
+  conditional_t<__is_primary_std_template<_Iter>::value,
+                conditional_t<__is_primary_cccl_template<_Iter>::value, _OtherTraits, iterator_traits<_Iter>>,
+                ::std::iterator_traits<_Iter>>;
+
+#endif // _CCCL_HOSTED()
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_PRIMARY_TEMPLATE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_reference.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_reference.h
new file mode 100644
index 00000000..ac93f3e7
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_reference.h
@@ -0,0 +1,89 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_REFERENCE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_REFERENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_LVALUE_REFERENCE) && !defined(_LIBCUDACXX_USE_IS_LVALUE_REFERENCE_FALLBACK)  \
+  && defined(_CCCL_BUILTIN_IS_RVALUE_REFERENCE) && !defined(_LIBCUDACXX_USE_IS_RVALUE_REFERENCE_FALLBACK) \
+  && defined(_CCCL_BUILTIN_IS_REFERENCE) && !defined(_LIBCUDACXX_USE_IS_REFERENCE_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_lvalue_reference : bool_constant<_CCCL_BUILTIN_IS_LVALUE_REFERENCE(_Tp)>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_rvalue_reference : bool_constant<_CCCL_BUILTIN_IS_RVALUE_REFERENCE(_Tp)>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_reference : bool_constant<_CCCL_BUILTIN_IS_REFERENCE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_lvalue_reference_v = _CCCL_BUILTIN_IS_LVALUE_REFERENCE(_Tp);
+template <class _Tp>
+inline constexpr bool is_rvalue_reference_v = _CCCL_BUILTIN_IS_RVALUE_REFERENCE(_Tp);
+template <class _Tp>
+inline constexpr bool is_reference_v = _CCCL_BUILTIN_IS_REFERENCE(_Tp);
+
+#else
+
+template <class _Tp>
+inline constexpr bool is_lvalue_reference_v = false;
+template <class _Tp>
+inline constexpr bool is_lvalue_reference_v<_Tp&> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_lvalue_reference : bool_constant<is_lvalue_reference_v<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_rvalue_reference_v = false;
+template <class _Tp>
+inline constexpr bool is_rvalue_reference_v<_Tp&&> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_rvalue_reference : bool_constant<is_rvalue_reference_v<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_reference_v = false;
+template <class _Tp>
+inline constexpr bool is_reference_v<_Tp&> = true;
+template <class _Tp>
+inline constexpr bool is_reference_v<_Tp&&> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_reference : bool_constant<is_reference_v<_Tp>>
+{};
+
+#endif // !_CCCL_BUILTIN_IS_LVALUE_REFERENCE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_REFERENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_referenceable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_referenceable.h
new file mode 100644
index 00000000..d8949b8f
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_referenceable.h
@@ -0,0 +1,55 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_REFERENCEABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_REFERENCEABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_same.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_REFERENCEABLE) && !defined(_LIBCUDACXX_USE_IS_REFERENCEABLE_FALLBACK)
+
+template <class _Tp>
+struct __cccl_is_referenceable : public integral_constant<bool, _CCCL_BUILTIN_IS_REFERENCEABLE(_Tp)>
+{};
+
+#else
+struct __cccl_is_referenceable_impl
+{
+  template <class _Tp>
+  _CCCL_HOST_DEVICE static _Tp& __test(int);
+  template <class _Tp>
+  _CCCL_HOST_DEVICE static false_type __test(...);
+};
+
+template <class _Tp>
+struct __cccl_is_referenceable
+    : integral_constant<bool, _IsNotSame<decltype(__cccl_is_referenceable_impl::__test<_Tp>(0)), false_type>::value>
+{};
+#endif // defined(_CCCL_BUILTIN_IS_REFERENCEABLE) && !defined(_LIBCUDACXX_USE_IS_REFERENCEABLE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_REFERENCEABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_same.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_same.h
new file mode 100644
index 00000000..37d2beee
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_same.h
@@ -0,0 +1,88 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_SAME_H
+#define _CUDA_STD___TYPE_TRAITS_IS_SAME_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_HAS_BUILTIN(__is_same_as) || _CCCL_COMPILER(GCC)
+#  define _CCCL_BUILTIN_IS_SAME_AS(...) __is_same_as(__VA_ARGS__)
+#endif // _CCCL_HAS_BUILTIN(__is_same_as) || _CCCL_COMPILER(GCC)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_SAME_AS)
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_same : bool_constant<_CCCL_BUILTIN_IS_SAME_AS(_Tp, _Up)>
+{};
+
+template <class _Tp, class _Up>
+inline constexpr bool is_same_v = _CCCL_BUILTIN_IS_SAME_AS(_Tp, _Up);
+
+// _IsSame<T,U> has the same effect as is_same<T,U> but instantiates fewer types:
+// is_same<A,B> and is_same<C,D> are guaranteed to be different types, but
+// _IsSame<A,B> and _IsSame<C,D> are the same type (namely, false_type).
+// Neither GCC nor Clang can mangle the __is_same builtin, so _IsSame
+// mustn't be directly used anywhere that contributes to name-mangling
+// (such as in a dependent return type).
+
+template <class _Tp, class _Up>
+using _IsSame = bool_constant<_CCCL_BUILTIN_IS_SAME_AS(_Tp, _Up)>;
+
+template <class _Tp, class _Up>
+using _IsNotSame = bool_constant<!_CCCL_BUILTIN_IS_SAME_AS(_Tp, _Up)>;
+
+#else // ^^^ _CCCL_BUILTIN_IS_SAME_AS ^^^ / vvv !_CCCL_BUILTIN_IS_SAME_AS vvv
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_same : false_type
+{};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_same<_Tp, _Tp> : true_type
+{};
+
+template <class _Tp, class _Up>
+inline constexpr bool is_same_v = false;
+template <class _Tp>
+inline constexpr bool is_same_v<_Tp, _Tp> = true;
+
+// _IsSame<T,U> has the same effect as is_same<T,U> but instantiates fewer types:
+// is_same<A,B> and is_same<C,D> are guaranteed to be different types, but
+// _IsSame<A,B> and _IsSame<C,D> are the same type (namely, false_type).
+// Neither GCC nor Clang can mangle the __is_same builtin, so _IsSame
+// mustn't be directly used anywhere that contributes to name-mangling
+// (such as in a dependent return type).
+
+template <class _Tp, class _Up>
+using _IsSame = bool_constant<is_same_v<_Tp, _Up>>;
+
+template <class _Tp, class _Up>
+using _IsNotSame = bool_constant<!is_same_v<_Tp, _Up>>;
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_SAME_AS ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_SAME_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_scalar.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_scalar.h
new file mode 100644
index 00000000..ee77a231
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_scalar.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_SCALAR_H
+#define _CUDA_STD___TYPE_TRAITS_IS_SCALAR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_arithmetic.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_member_pointer.h>
+#include <cuda/std/__type_traits/is_null_pointer.h>
+#include <cuda/std/__type_traits/is_pointer.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_SCALAR) && !defined(_LIBCUDACXX_USE_IS_SCALAR_FALLBACK)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_scalar : public bool_constant<_CCCL_BUILTIN_IS_SCALAR(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_scalar_v = _CCCL_BUILTIN_IS_SCALAR(_Tp);
+
+#else
+
+template <class _Tp>
+inline constexpr bool is_scalar_v =
+  is_arithmetic_v<_Tp> || is_member_pointer_v<_Tp> || is_pointer_v<_Tp> || is_null_pointer_v<_Tp> || is_enum_v<_Tp>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_scalar : public bool_constant<is_scalar_v<_Tp>>
+{};
+
+#endif // defined(_CCCL_BUILTIN_IS_SCALAR) && !defined(_LIBCUDACXX_USE_IS_SCALAR_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_SCALAR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed.h
new file mode 100644
index 00000000..7f1d3951
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed.h
@@ -0,0 +1,76 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_SIGNED_H
+#define _CUDA_STD___TYPE_TRAITS_IS_SIGNED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_arithmetic.h>
+#include <cuda/std/__type_traits/is_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_signed)
+#  define _CCCL_BUILTIN_IS_SIGNED(...) __is_signed(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_signed)
+
+// if we put this trait to cuda::std::, it colides with libstdc++, putting it to to global namespace seems to work fine
+#if defined(_CCCL_BUILTIN_IS_SIGNED)
+template <class _Tp>
+inline constexpr bool __cccl_is_signed_v = _CCCL_BUILTIN_IS_SIGNED(_Tp);
+#endif // _CCCL_BUILTIN_IS_SIGNED
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_SIGNED)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_signed : bool_constant<::__cccl_is_signed_v<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_signed_v = ::__cccl_is_signed_v<_Tp>;
+
+#else // ^^^ _CCCL_BUILTIN_IS_SIGNED ^^^ / vvv !_CCCL_BUILTIN_IS_SIGNED vvv
+
+template <class _Tp, bool = is_integral_v<_Tp>>
+inline constexpr bool __cccl_is_signed_helper_v = true;
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4197) // top-level volatile in cast is ignored
+
+template <class _Tp>
+inline constexpr bool __cccl_is_signed_helper_v<_Tp, true> = _Tp(-1) < _Tp(0);
+
+_CCCL_DIAG_POP
+
+template <class _Tp>
+inline constexpr bool is_signed_v = is_arithmetic_v<_Tp> && __cccl_is_signed_helper_v<_Tp>;
+
+template <class _Tp>
+struct is_signed : bool_constant<is_signed_v<_Tp>>
+{};
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_SIGNED ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_SIGNED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed_integer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed_integer.h
new file mode 100644
index 00000000..336971ec
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_signed_integer.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_SIGNED_INTEGER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_SIGNED_INTEGER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __cccl_is_signed_integer_v = false;
+
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<signed char> = true;
+
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<signed short> = true;
+
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<signed int> = true;
+
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<signed long> = true;
+
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<signed long long> = true;
+
+#if _CCCL_HAS_INT128()
+template <>
+inline constexpr bool __cccl_is_signed_integer_v<__int128_t> = true;
+#endif // _CCCL_HAS_INT128()
+
+template <class _Tp>
+inline constexpr bool __cccl_is_cv_signed_integer_v = __cccl_is_signed_integer_v<remove_cv_t<_Tp>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_SIGNED_INTEGER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_swappable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_swappable.h
new file mode 100644
index 00000000..8e9a87c7
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_swappable.h
@@ -0,0 +1,211 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_SWAPPABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_SWAPPABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/add_lvalue_reference.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/disjunction.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_complete.h>
+#include <cuda/std/__type_traits/is_move_assignable.h>
+#include <cuda/std/__type_traits/is_move_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/is_referenceable.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_void.h>
+#include <cuda/std/__type_traits/nat.h>
+#include <cuda/std/__type_traits/type_identity.h>
+#include <cuda/std/__utility/declval.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// We need to detect whether there is already a free function swap that would end up being ambiguous.
+// This can happen when a type pulls in both namespace std and namespace cuda::std via ADL.
+// In that case we are always safe to just not do anything because that type must be host only.
+// However, we must be careful to ensure that we still create the overload if there is just a hidden friend swap
+namespace __detect_hidden_friend_swap
+{
+// This will intentionally create an ambiguity with std::swap if that is find-able by ADL. But it will not interfere
+// with hidden friend swap
+template <class _Tp>
+_CCCL_HOST_DEVICE void swap(_Tp&, _Tp&); // NOLINT(performance-noexcept-swap)
+
+struct __hidden_friend_swap_found
+{};
+
+template <class _Tp>
+_CCCL_API inline auto __swap(_Tp& __lhs, _Tp& __rhs) -> decltype(swap(__lhs, __rhs));
+_CCCL_API inline auto __swap(...) -> __hidden_friend_swap_found;
+template <class _Tp>
+struct __has_hidden_friend_swap
+    : is_same<decltype(__detect_hidden_friend_swap::__swap(::cuda::std::declval<_Tp&>(), ::cuda::std::declval<_Tp&>())),
+              void>
+{};
+} // namespace __detect_hidden_friend_swap
+
+namespace __detect_adl_swap
+{
+template <class _Tp>
+void swap(_Tp&, _Tp&) = delete;
+
+struct __no_adl_swap_found
+{};
+template <class _Tp>
+_CCCL_API inline auto __swap(_Tp& __lhs, _Tp& __rhs) -> decltype(swap(__lhs, __rhs));
+_CCCL_API inline auto __swap(...) -> __no_adl_swap_found;
+template <class _Tp>
+struct __has_no_adl_swap
+    : is_same<decltype(__detect_adl_swap::__swap(::cuda::std::declval<_Tp&>(), ::cuda::std::declval<_Tp&>())),
+              __no_adl_swap_found>
+{};
+template <class _Tp, size_t _Np>
+struct __has_no_adl_swap_array
+    : is_same<
+        decltype(__detect_adl_swap::__swap(::cuda::std::declval<_Tp (&)[_Np]>(), ::cuda::std::declval<_Tp (&)[_Np]>())),
+        __no_adl_swap_found>
+{};
+
+// We should only define swap if there is no ADL found function or it is a hidden friend
+template <class _Tp>
+struct __can_define_swap : _Or<__has_no_adl_swap<_Tp>, __detect_hidden_friend_swap::__has_hidden_friend_swap<_Tp>>
+{};
+} // namespace __detect_adl_swap
+
+template <class _Tp>
+struct __is_swappable;
+template <class _Tp>
+struct __is_nothrow_swappable;
+
+template <class _Tp>
+using __swap_result_t _CCCL_NODEBUG_ALIAS =
+  enable_if_t<__detect_adl_swap::__can_define_swap<_Tp>::value
+              && is_move_constructible_v<_Tp> && is_move_assignable_v<_Tp>>;
+
+// we use type_identity_t<_Tp> as second parameter, to avoid ambiguity with std::swap, which will thus be preferred by
+// overload resolution (which is ok since std::swap is only considered when explicitly called, or found by ADL for types
+// from std::)
+template <class _Tp>
+_CCCL_API constexpr __swap_result_t<_Tp> swap(_Tp& __x, type_identity_t<_Tp>& __y) noexcept(
+  is_nothrow_move_constructible_v<_Tp> && is_nothrow_move_assignable_v<_Tp>);
+
+template <class _Tp, size_t _Np>
+_CCCL_API constexpr enable_if_t<__detect_adl_swap::__has_no_adl_swap_array<_Tp, _Np>::value && __is_swappable<_Tp>::value>
+  swap(_Tp (&__a)[_Np], _Tp (&__b)[_Np]) noexcept(__is_nothrow_swappable<_Tp>::value);
+
+namespace __detail
+{
+// ALL generic swap overloads MUST already have a declaration available at this point.
+
+template <class _Tp, class _Up = _Tp, bool _NotVoid = !is_void_v<_Tp> && !is_void_v<_Up>>
+struct __swappable_with
+{
+  template <class _LHS, class _RHS>
+  _CCCL_API inline static decltype(swap(::cuda::std::declval<_LHS>(), ::cuda::std::declval<_RHS>())) __test_swap(int);
+  template <class, class>
+  _CCCL_API inline static __nat __test_swap(long);
+
+  // Extra parens are needed for the C++03 definition of decltype.
+  using __swap1 = decltype((__test_swap<_Tp, _Up>(0)));
+  using __swap2 = decltype((__test_swap<_Up, _Tp>(0)));
+
+  static const bool value = _IsNotSame<__swap1, __nat>::value && _IsNotSame<__swap2, __nat>::value;
+};
+
+template <class _Tp, class _Up>
+struct __swappable_with<_Tp, _Up, false> : false_type
+{};
+
+template <class _Tp, class _Up = _Tp, bool _Swappable = __swappable_with<_Tp, _Up>::value>
+struct __nothrow_swappable_with
+{
+  static const bool value = noexcept(swap(::cuda::std::declval<_Tp>(), ::cuda::std::declval<_Up>()))
+                         && noexcept(swap(::cuda::std::declval<_Up>(), ::cuda::std::declval<_Tp>()));
+};
+
+template <class _Tp, class _Up>
+struct __nothrow_swappable_with<_Tp, _Up, false> : false_type
+{};
+} // namespace __detail
+
+template <class _Tp>
+struct __is_swappable : public integral_constant<bool, __detail::__swappable_with<_Tp&>::value>
+{};
+
+template <class _Tp>
+struct __is_nothrow_swappable : public integral_constant<bool, __detail::__nothrow_swappable_with<_Tp&>::value>
+{};
+
+template <class _Tp, class _Up>
+inline constexpr bool is_swappable_with_v = __detail::__swappable_with<_Tp, _Up>::value;
+
+template <class _Tp, bool = __cccl_is_referenceable<_Tp>::value>
+inline constexpr bool is_swappable_v = false;
+
+template <class _Tp>
+inline constexpr bool is_swappable_v<_Tp, true> =
+  is_swappable_with_v<add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<_Tp>>;
+
+template <class _Tp, class _Up>
+inline constexpr bool is_nothrow_swappable_with_v = __detail::__nothrow_swappable_with<_Tp, _Up>::value;
+
+template <class _Tp, bool = __cccl_is_referenceable<_Tp>::value>
+inline constexpr bool is_nothrow_swappable_v = false;
+
+template <class _Tp>
+inline constexpr bool is_nothrow_swappable_v<_Tp, true> =
+  is_nothrow_swappable_with_v<add_lvalue_reference_t<_Tp>, add_lvalue_reference_t<_Tp>>;
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_swappable_with : public bool_constant<is_swappable_with_v<_Tp, _Up>>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_swappable : public bool_constant<is_swappable_v<_Tp>>
+{};
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+is_nothrow_swappable_with : public bool_constant<is_nothrow_swappable_with_v<_Tp, _Up>>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_nothrow_swappable : public bool_constant<is_nothrow_swappable_v<_Tp>>
+{};
+
+// Do not use this trait over is_nothrow_swappable unless necessary. It is only useful in rare
+// cases where a type may need to work with incomplete types and has friend functions that
+// greedily instantiate the nothrow_swappable functions.
+template <class _Tp, bool = __is_complete_v<_Tp>>
+inline constexpr bool __is_complete_and_nothrow_swappable_v = false;
+
+template <class _Tp>
+inline constexpr bool __is_complete_and_nothrow_swappable_v<_Tp, true> = is_nothrow_swappable_v<_Tp>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_SWAPPABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_trivially_copyable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_trivially_copyable.h
new file mode 100644
index 00000000..37518c14
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_trivially_copyable.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_TRIVIALLY_COPYABLE(...) __is_trivially_copyable(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_trivially_copyable : bool_constant<_CCCL_BUILTIN_IS_TRIVIALLY_COPYABLE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_trivially_copyable_v = _CCCL_BUILTIN_IS_TRIVIALLY_COPYABLE(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_TRIVIALLY_COPYABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unbounded_array.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unbounded_array.h
new file mode 100644
index 00000000..a4348001
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unbounded_array.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_UNBOUNDED_ARRAY_H
+#define _CUDA_STD___TYPE_TRAITS_IS_UNBOUNDED_ARRAY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_unbounded_array)
+#  define _CCCL_BUILTIN_IS_UNBOUNDED_ARRAY(...) __is_unbounded_array(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_unbounded_array)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_UNBOUNDED_ARRAY)
+
+template <class _Tp>
+inline constexpr bool is_unbounded_array_v = _CCCL_BUILTIN_IS_UNBOUNDED_ARRAY(_Tp);
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_unbounded_array : bool_constant<_CCCL_BUILTIN_IS_UNBOUNDED_ARRAY(_Tp)>
+{};
+
+#else // ^^^ _CCCL_BUILTIN_IS_UNBOUNDED_ARRAY ^^^ / vvv !_CCCL_BUILTIN_IS_UNBOUNDED_ARRAY vvv
+
+template <class _Tp>
+inline constexpr bool is_unbounded_array_v = false;
+
+template <class _Tp>
+inline constexpr bool is_unbounded_array_v<_Tp[]> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_unbounded_array : bool_constant<is_unbounded_array_v<_Tp>>
+{};
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_UNBOUNDED_ARRAY ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_UNBOUNDED_ARRAY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_union.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_union.h
new file mode 100644
index 00000000..b10acf1d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_union.h
@@ -0,0 +1,42 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_UNION_H
+#define _CUDA_STD___TYPE_TRAITS_IS_UNION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_IS_UNION(...) __is_union(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_union : bool_constant<_CCCL_BUILTIN_IS_UNION(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_union_v = _CCCL_BUILTIN_IS_UNION(_Tp);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_UNION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned.h
new file mode 100644
index 00000000..8dab2d72
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned.h
@@ -0,0 +1,76 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_H
+#define _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_arithmetic.h>
+#include <cuda/std/__type_traits/is_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_unsigned)
+#  define _CCCL_BUILTIN_IS_UNSIGNED(...) __is_unsigned(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_unsigned)
+
+// if we put this trait to cuda::std::, it colides with libstdc++, putting it to to global namespace seems to work fine
+#if defined(_CCCL_BUILTIN_IS_UNSIGNED)
+template <class _Tp>
+inline constexpr bool __cccl_is_unsigned_v = _CCCL_BUILTIN_IS_UNSIGNED(_Tp);
+#endif // _CCCL_BUILTIN_IS_UNSIGNED
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_UNSIGNED)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_unsigned : bool_constant<::__cccl_is_unsigned_v<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_unsigned_v = ::__cccl_is_unsigned_v<_Tp>;
+
+#else // ^^^ _CCCL_BUILTIN_IS_UNSIGNED ^^^ / vvv !_CCCL_BUILTIN_IS_UNSIGNED vvv
+
+template <class _Tp, bool = is_integral_v<_Tp>>
+inline constexpr bool __cccl_is_unsigned_helper_v = false;
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4197) //  top-level volatile in cast is ignored
+
+template <class _Tp>
+inline constexpr bool __cccl_is_unsigned_helper_v<_Tp, true> = _Tp(0) < _Tp(-1);
+
+_CCCL_DIAG_POP
+
+template <class _Tp>
+inline constexpr bool is_unsigned_v = is_arithmetic_v<_Tp> && __cccl_is_unsigned_helper_v<_Tp>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_unsigned : bool_constant<is_unsigned_v<_Tp>>
+{};
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_UNSIGNED ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned_integer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned_integer.h
new file mode 100644
index 00000000..032b8e16
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_unsigned_integer.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_INTEGER_H
+#define _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_INTEGER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/remove_cv.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+inline constexpr bool __cccl_is_unsigned_integer_v = false;
+
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<unsigned char> = true;
+
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<unsigned short> = true;
+
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<unsigned int> = true;
+
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<unsigned long> = true;
+
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<unsigned long long> = true;
+
+#if _CCCL_HAS_INT128()
+template <>
+inline constexpr bool __cccl_is_unsigned_integer_v<__uint128_t> = true;
+#endif // _CCCL_HAS_INT128()
+
+template <class _Tp>
+inline constexpr bool __cccl_is_cv_unsigned_integer_v = __cccl_is_unsigned_integer_v<remove_cv_t<_Tp>>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_UNSIGNED_INTEGER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_valid_expansion.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_valid_expansion.h
new file mode 100644
index 00000000..3e2c1140
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_valid_expansion.h
@@ -0,0 +1,41 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_VALID_EXPANSION_H
+#define _CUDA_STD___TYPE_TRAITS_IS_VALID_EXPANSION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <template <class...> class _Templ, class... _Args, class = _Templ<_Args...>>
+_CCCL_API inline true_type __sfinae_test_impl(int);
+template <template <class...> class, class...>
+_CCCL_API inline false_type __sfinae_test_impl(...);
+
+template <template <class...> class _Templ, class... _Args>
+using _IsValidExpansion _CCCL_NODEBUG_ALIAS = decltype(::cuda::std::__sfinae_test_impl<_Templ, _Args...>(0));
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_VALID_EXPANSION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_void.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_void.h
new file mode 100644
index 00000000..4aa23a31
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_void.h
@@ -0,0 +1,70 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_VOID_H
+#define _CUDA_STD___TYPE_TRAITS_IS_VOID_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/remove_cvref.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_void)
+#  define _CCCL_BUILTIN_IS_VOID(...) __is_void(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_void)
+
+// clang + nvcc fails with '_Tp does not refer to a value' or 'identifier __is_void is undefined'
+#if _CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC)
+#  undef _CCCL_BUILTIN_IS_VOID
+#endif // _CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC)
+
+// if we put this trait to cuda::std::, it colides with libstdc++, putting it to to global namespace seems to work fine
+#if defined(_CCCL_BUILTIN_IS_VOID)
+template <class _Tp>
+inline constexpr bool __cccl_is_void_v = _CCCL_BUILTIN_IS_VOID(_Tp);
+#endif // _CCCL_BUILTIN_IS_VOID
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_VOID)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_void : bool_constant<::__cccl_is_void_v<_Tp>>
+{};
+
+template <class _Tp>
+inline constexpr bool is_void_v = ::__cccl_is_void_v<_Tp>;
+
+#else // ^^^ _CCCL_BUILTIN_IS_VOID ^^^ / vvv !_CCCL_BUILTIN_IS_VOID vvv
+
+template <class _Tp>
+inline constexpr bool is_void_v = is_same_v<remove_cv_t<_Tp>, void>;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_void : public bool_constant<is_void_v<_Tp>>
+{};
+
+#endif // ^^^ !_CCCL_BUILTIN_IS_VOID ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_VOID_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_volatile.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_volatile.h
new file mode 100644
index 00000000..e17b6a62
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/is_volatile.h
@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_IS_VOLATILE_H
+#define _CUDA_STD___TYPE_TRAITS_IS_VOLATILE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(is_volatile)
+#  define _CCCL_BUILTIN_IS_VOLATILE(...) __is_volatile(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(is_volatile)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_IS_VOLATILE)
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_volatile : bool_constant<_CCCL_BUILTIN_IS_VOLATILE(_Tp)>
+{};
+
+template <class _Tp>
+inline constexpr bool is_volatile_v = _CCCL_BUILTIN_IS_VOLATILE(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_IS_VOLATILE ^^^ / vvv !_CCCL_BUILTIN_IS_VOLATILE vvv
+
+template <class _Tp>
+inline constexpr bool is_volatile_v = false;
+
+template <class _Tp>
+inline constexpr bool is_volatile_v<volatile _Tp> = true;
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT is_volatile : bool_constant<is_volatile_v<_Tp>>
+{};
+
+#endif // !_CCCL_BUILTIN_IS_VOLATILE
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_IS_VOLATILE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/lazy.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/lazy.h
new file mode 100644
index 00000000..74d0ef80
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/lazy.h
@@ -0,0 +1,35 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_LAZY_H
+#define _CUDA_STD___TYPE_TRAITS_LAZY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <template <class...> class _Func, class... _Args>
+struct _Lazy : _Func<_Args...>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_LAZY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_const_lvalue_ref.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_const_lvalue_ref.h
new file mode 100644
index 00000000..073b2119
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_const_lvalue_ref.h
@@ -0,0 +1,36 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_MAKE_CONST_LVALUE_REF_H
+#define _CUDA_STD___TYPE_TRAITS_MAKE_CONST_LVALUE_REF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/remove_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+using __make_const_lvalue_ref = const remove_reference_t<_Tp>&;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_MAKE_CONST_LVALUE_REF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_signed.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_signed.h
new file mode 100644
index 00000000..678a2f78
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_signed.h
@@ -0,0 +1,155 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_MAKE_SIGNED_H
+#define _CUDA_STD___TYPE_TRAITS_MAKE_SIGNED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/copy_cvref.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/nat.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/type_list.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(make_signed)
+#  define _CCCL_BUILTIN_MAKE_SIGNED(...) __make_signed(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(make_signed)
+
+// __make_signed doesn't work with clang < 20 or clang + nvcc
+#if _CCCL_COMPILER(CLANG, <, 20) || (_CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC))
+#  undef _CCCL_BUILTIN_MAKE_SIGNED
+#endif // _CCCL_COMPILER(CLANG, <, 20) || (_CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC))
+
+// nvrtc doesn't handle enums properly, see nvbug 5554624
+#if _CCCL_COMPILER(NVRTC, <, 13, 3)
+#  undef _CCCL_BUILTIN_MAKE_SIGNED
+#endif // _CCCL_COMPILER(NVRTC, <, 13, 3)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_MAKE_SIGNED)
+
+template <class _Tp>
+using make_signed_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_MAKE_SIGNED(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_MAKE_SIGNED ^^^ / vvv !_CCCL_BUILTIN_MAKE_SIGNED vvv
+
+using __signed_types =
+  __type_list<signed char,
+              signed short,
+              signed int,
+              signed long,
+              signed long long
+#  if _CCCL_HAS_INT128()
+              ,
+              __int128_t
+#  endif // _CCCL_HAS_INT128()
+              >;
+
+template <class _Tp, bool = is_integral_v<_Tp> || is_enum_v<_Tp>>
+struct __make_signed_impl
+{};
+
+template <class _Tp>
+struct __make_signed_impl<_Tp, true>
+{
+  struct __size_greater_equal_fn
+  {
+    template <class _Up>
+    using __call = bool_constant<(sizeof(_Tp) <= sizeof(_Up))>;
+  };
+
+  using type = __type_front<__type_find_if<__signed_types, __size_greater_equal_fn>>;
+};
+
+template <>
+struct __make_signed_impl<bool, true>
+{};
+template <>
+struct __make_signed_impl<signed short, true>
+{
+  using type = short;
+};
+template <>
+struct __make_signed_impl<unsigned short, true>
+{
+  using type = short;
+};
+template <>
+struct __make_signed_impl<signed int, true>
+{
+  using type = int;
+};
+template <>
+struct __make_signed_impl<unsigned int, true>
+{
+  using type = int;
+};
+template <>
+struct __make_signed_impl<signed long, true>
+{
+  using type = long;
+};
+template <>
+struct __make_signed_impl<unsigned long, true>
+{
+  using type = long;
+};
+template <>
+struct __make_signed_impl<signed long long, true>
+{
+  using type = long long;
+};
+template <>
+struct __make_signed_impl<unsigned long long, true>
+{
+  using type = long long;
+};
+#  if _CCCL_HAS_INT128()
+template <>
+struct __make_signed_impl<__int128_t, true>
+{
+  using type = __int128_t;
+};
+template <>
+struct __make_signed_impl<__uint128_t, true>
+{
+  using type = __int128_t;
+};
+#  endif // _CCCL_HAS_INT128()
+
+template <class _Tp>
+using make_signed_t _CCCL_NODEBUG_ALIAS = __copy_cvref_t<_Tp, typename __make_signed_impl<remove_cv_t<_Tp>>::type>;
+
+#endif // ^^^ !_CCCL_BUILTIN_MAKE_SIGNED !!!
+
+template <class _Tp>
+struct make_signed
+{
+  using type _CCCL_NODEBUG_ALIAS = make_signed_t<_Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_MAKE_SIGNED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_unsigned.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_unsigned.h
new file mode 100644
index 00000000..2bc41cba
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/make_unsigned.h
@@ -0,0 +1,166 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_MAKE_UNSIGNED_H
+#define _CUDA_STD___TYPE_TRAITS_MAKE_UNSIGNED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/copy_cvref.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_integral.h>
+#include <cuda/std/__type_traits/is_unsigned.h>
+#include <cuda/std/__type_traits/nat.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/type_list.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(make_unsigned)
+#  define _CCCL_BUILTIN_MAKE_UNSIGNED(...) __make_unsigned(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(make_unsigned)
+
+// __make_unsigned doesn't work with clang < 20 or clang + nvcc
+#if _CCCL_COMPILER(CLANG, <, 20) || (_CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC))
+#  undef _CCCL_BUILTIN_MAKE_UNSIGNED
+#endif // _CCCL_COMPILER(CLANG, <, 20) || (_CCCL_COMPILER(CLANG) && _CCCL_CUDA_COMPILER(NVCC))
+
+// nvrtc doesn't handle enums properly, see nvbug 5554624
+#if _CCCL_COMPILER(NVRTC, <, 13, 3)
+#  undef _CCCL_BUILTIN_MAKE_UNSIGNED
+#endif // _CCCL_COMPILER(NVRTC, <, 13, 3)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_MAKE_UNSIGNED)
+
+template <class _Tp>
+using make_unsigned_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_MAKE_UNSIGNED(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_MAKE_UNSIGNED ^^^ / vvv !_CCCL_BUILTIN_MAKE_UNSIGNED vvv
+
+using __unsigned_types =
+  __type_list<unsigned char,
+              unsigned short,
+              unsigned int,
+              unsigned long,
+              unsigned long long
+#  if _CCCL_HAS_INT128()
+              ,
+              __uint128_t
+#  endif // _CCCL_HAS_INT128()
+              >;
+
+template <class _Tp, bool = is_integral_v<_Tp> || is_enum_v<_Tp>>
+struct __make_unsigned_impl
+{};
+
+template <class _Tp>
+struct __make_unsigned_impl<_Tp, true>
+{
+  struct __size_greater_equal_fn
+  {
+    template <class _Up>
+    using __call = bool_constant<(sizeof(_Tp) <= sizeof(_Up))>;
+  };
+
+  using type = __type_front<__type_find_if<__unsigned_types, __size_greater_equal_fn>>;
+};
+
+template <>
+struct __make_unsigned_impl<bool, true>
+{};
+template <>
+struct __make_unsigned_impl<signed short, true>
+{
+  using type = unsigned short;
+};
+template <>
+struct __make_unsigned_impl<unsigned short, true>
+{
+  using type = unsigned short;
+};
+template <>
+struct __make_unsigned_impl<signed int, true>
+{
+  using type = unsigned int;
+};
+template <>
+struct __make_unsigned_impl<unsigned int, true>
+{
+  using type = unsigned int;
+};
+template <>
+struct __make_unsigned_impl<signed long, true>
+{
+  using type = unsigned long;
+};
+template <>
+struct __make_unsigned_impl<unsigned long, true>
+{
+  using type = unsigned long;
+};
+template <>
+struct __make_unsigned_impl<signed long long, true>
+{
+  using type = unsigned long long;
+};
+template <>
+struct __make_unsigned_impl<unsigned long long, true>
+{
+  using type = unsigned long long;
+};
+#  if _CCCL_HAS_INT128()
+template <>
+struct __make_unsigned_impl<__int128_t, true>
+{
+  using type = __uint128_t;
+};
+template <>
+struct __make_unsigned_impl<__uint128_t, true>
+{
+  using type = __uint128_t;
+};
+#  endif // _CCCL_HAS_INT128()
+
+template <class _Tp>
+using make_unsigned_t _CCCL_NODEBUG_ALIAS = __copy_cvref_t<_Tp, typename __make_unsigned_impl<remove_cv_t<_Tp>>::type>;
+
+#endif // ^^^ !_CCCL_BUILTIN_MAKE_UNSIGNED ^^^
+
+template <class _Tp>
+struct make_unsigned
+{
+  using type _CCCL_NODEBUG_ALIAS = make_unsigned_t<_Tp>;
+};
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr make_unsigned_t<_Tp> __to_unsigned_like(_Tp __x) noexcept
+{
+  return static_cast<make_unsigned_t<_Tp>>(__x);
+}
+
+template <class _Tp, class _Up>
+using __copy_unsigned_t _CCCL_NODEBUG_ALIAS = conditional_t<is_unsigned_v<_Tp>, make_unsigned_t<_Up>, _Up>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_MAKE_UNSIGNED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/nat.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/nat.h
new file mode 100644
index 00000000..5c5fda4d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/nat.h
@@ -0,0 +1,39 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_NAT_H
+#define _CUDA_STD___TYPE_TRAITS_NAT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct __nat
+{
+  __nat()                        = delete;
+  __nat(const __nat&)            = delete;
+  __nat& operator=(const __nat&) = delete;
+  ~__nat()                       = delete;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_NAT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/negation.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/negation.h
new file mode 100644
index 00000000..874f2437
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/negation.h
@@ -0,0 +1,43 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_NEGATION_H
+#define _CUDA_STD___TYPE_TRAITS_NEGATION_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct negation : bool_constant<!_Tp::value>
+{};
+
+template <class _Tp>
+inline constexpr bool negation_v = !_Tp::value;
+
+template <class _Tp>
+using _Not = negation<_Tp>;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_NEGATION_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/reference_constructs_from_temporary.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/reference_constructs_from_temporary.h
new file mode 100644
index 00000000..626c5bdd
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/reference_constructs_from_temporary.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REFERENCE_CONSTRUCTS_FROM_TEMPORARY_H
+#define _CUDA_STD___TYPE_TRAITS_REFERENCE_CONSTRUCTS_FROM_TEMPORARY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/always_false.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+template <class _Tp, class _Up>
+struct reference_constructs_from_temporary
+    : integral_constant<bool, _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY(_Tp, _Up)>
+{};
+
+template <class _Tp, class _Up>
+inline constexpr bool reference_constructs_from_temporary_v =
+  _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY(_Tp, _Up);
+
+#else
+
+template <class _Tp, class _Up>
+struct reference_constructs_from_temporary : integral_constant<bool, false>
+{
+  static_assert(__always_false_v<_Tp>, "The compiler does not support __reference_constructs_from_temporary");
+};
+
+template <class _Tp, class _Up>
+inline constexpr bool reference_constructs_from_temporary_v = reference_constructs_from_temporary<_Tp, _Up>::value;
+
+#endif // !_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REFERENCE_CONSTRUCTS_FROM_TEMPORARY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_all_extents.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_all_extents.h
new file mode 100644
index 00000000..0cbc8127
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_all_extents.h
@@ -0,0 +1,71 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_ALL_EXTENTS_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_ALL_EXTENTS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#if _CCCL_CHECK_BUILTIN(remove_all_extents)
+#  define _CCCL_BUILTIN_REMOVE_ALL_EXTENTS(...) __remove_all_extents(__VA_ARGS__)
+#endif // _CCCL_CHECK_BUILTIN(remove_all_extents)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_ALL_EXTENTS)
+
+template <class _Tp>
+struct remove_all_extents
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_ALL_EXTENTS(_Tp);
+};
+
+template <class _Tp>
+using remove_all_extents_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_ALL_EXTENTS(_Tp);
+
+#else // ^^^ _CCCL_BUILTIN_REMOVE_ALL_EXTENTS ^^^ / vvv !_CCCL_BUILTIN_REMOVE_ALL_EXTENTS vvv
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_all_extents
+{
+  using type = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_all_extents<_Tp[]>
+{
+  using type = typename remove_all_extents<_Tp>::type;
+};
+template <class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_all_extents<_Tp[_Np]>
+{
+  using type = typename remove_all_extents<_Tp>::type;
+};
+
+template <class _Tp>
+using remove_all_extents_t _CCCL_NODEBUG_ALIAS = typename remove_all_extents<_Tp>::type;
+
+#endif // ^^^ !_CCCL_BUILTIN_REMOVE_ALL_EXTENTS ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_ALL_EXTENTS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_const.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_const.h
new file mode 100644
index 00000000..cdd3d7d6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_const.h
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_CONST_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_CONST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_CONST) && !defined(_LIBCUDACXX_USE_REMOVE_CONST_FALLBACK)
+template <class _Tp>
+struct remove_const
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_CONST(_Tp);
+};
+
+template <class _Tp>
+using remove_const_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_CONST(_Tp);
+
+#else
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_const
+{
+  using type = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_const<const _Tp>
+{
+  using type = _Tp;
+};
+
+template <class _Tp>
+using remove_const_t _CCCL_NODEBUG_ALIAS = typename remove_const<_Tp>::type;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_CONST) && !defined(_LIBCUDACXX_USE_REMOVE_CONST_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_CONST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cv.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cv.h
new file mode 100644
index 00000000..85d1ef3c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cv.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_CV_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_CV_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/remove_const.h>
+#include <cuda/std/__type_traits/remove_volatile.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_CV) && !defined(_LIBCUDACXX_USE_REMOVE_CV_FALLBACK)
+template <class _Tp>
+struct remove_cv
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_CV(_Tp);
+};
+
+template <class _Tp>
+using remove_cv_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_CV(_Tp);
+
+#else
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_cv
+{
+  using type = remove_volatile_t<remove_const_t<_Tp>>;
+};
+
+template <class _Tp>
+using remove_cv_t _CCCL_NODEBUG_ALIAS = remove_volatile_t<remove_const_t<_Tp>>;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_CV) && !defined(_LIBCUDACXX_USE_REMOVE_CV_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_CV_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cvref.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cvref.h
new file mode 100644
index 00000000..4e5a9253
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_cvref.h
@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_CVREF_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_CVREF_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_CVREF) && !defined(_LIBCUDACXX_USE_REMOVE_CVREF_FALLBACK)
+
+template <class _Tp>
+using remove_cvref_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_CVREF(_Tp);
+
+#else
+
+template <class _Tp>
+using remove_cvref_t _CCCL_NODEBUG_ALIAS = remove_cv_t<remove_reference_t<_Tp>>;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_CVREF) && !defined(_LIBCUDACXX_USE_REMOVE_CVREF_FALLBACK)
+
+template <class _Tp, class _Up>
+struct __is_same_uncvref : _IsSame<remove_cvref_t<_Tp>, remove_cvref_t<_Up>>
+{};
+
+template <class _Tp>
+struct remove_cvref
+{
+  using type _CCCL_NODEBUG_ALIAS = remove_cvref_t<_Tp>;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_CVREF_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_extent.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_extent.h
new file mode 100644
index 00000000..856d5dac
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_extent.h
@@ -0,0 +1,65 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_EXTENT_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_EXTENT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_EXTENT) && !defined(_LIBCUDACXX_USE_REMOVE_EXTENT_FALLBACK)
+template <class _Tp>
+struct remove_extent
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_EXTENT(_Tp);
+};
+
+template <class _Tp>
+using remove_extent_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_EXTENT(_Tp);
+
+#else
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_extent
+{
+  using type = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_extent<_Tp[]>
+{
+  using type = _Tp;
+};
+template <class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_extent<_Tp[_Np]>
+{
+  using type = _Tp;
+};
+
+template <class _Tp>
+using remove_extent_t _CCCL_NODEBUG_ALIAS = typename remove_extent<_Tp>::type;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_EXTENT) && !defined(_LIBCUDACXX_USE_REMOVE_EXTENT_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_EXTENT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_pointer.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_pointer.h
new file mode 100644
index 00000000..5e1d54cc
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_pointer.h
@@ -0,0 +1,73 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_POINTER_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_POINTER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_POINTER) && !defined(_LIBCUDACXX_USE_REMOVE_POINTER_FALLBACK)
+template <class _Tp>
+struct remove_pointer
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_POINTER(_Tp);
+};
+
+template <class _Tp>
+using remove_pointer_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_POINTER(_Tp);
+
+#else
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_pointer
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_pointer<_Tp*>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_pointer<_Tp* const>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_pointer<_Tp* volatile>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_pointer<_Tp* const volatile>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+template <class _Tp>
+using remove_pointer_t _CCCL_NODEBUG_ALIAS = typename remove_pointer<_Tp>::type;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_POINTER) && !defined(_LIBCUDACXX_USE_REMOVE_POINTER_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_POINTER_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_reference.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_reference.h
new file mode 100644
index 00000000..8751b401
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_reference.h
@@ -0,0 +1,72 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_REFERENCE_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_REFERENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_REFERENCE_T) && !defined(_LIBCUDACXX_USE_REMOVE_REFERENCE_T_FALLBACK)
+template <class _Tp>
+struct remove_reference
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_REFERENCE_T(_Tp);
+};
+
+#  if _CCCL_COMPILER(GCC)
+// error: use of built-in trait in function signature; use library traits instead
+template <class _Tp>
+using remove_reference_t _CCCL_NODEBUG_ALIAS = typename remove_reference<_Tp>::type;
+#  else // ^^^ _CCCL_COMPILER(GCC) ^^^^/  vvv !_CCCL_COMPILER(GCC)
+template <class _Tp>
+using remove_reference_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_REFERENCE_T(_Tp);
+#  endif // !_CCCL_COMPILER(GCC)
+
+#else // ^^^ _CCCL_BUILTIN_REMOVE_REFERENCE_T ^^^ / vvv !_CCCL_BUILTIN_REMOVE_REFERENCE_T vvv
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_reference
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_reference<_Tp&>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_reference<_Tp&&>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+template <class _Tp>
+using remove_reference_t _CCCL_NODEBUG_ALIAS = typename remove_reference<_Tp>::type;
+
+#endif // !_CCCL_BUILTIN_REMOVE_REFERENCE_T
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_REFERENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_volatile.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_volatile.h
new file mode 100644
index 00000000..3ea53a95
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/remove_volatile.h
@@ -0,0 +1,58 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_REMOVE_VOLATILE_H
+#define _CUDA_STD___TYPE_TRAITS_REMOVE_VOLATILE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if defined(_CCCL_BUILTIN_REMOVE_VOLATILE) && !defined(_LIBCUDACXX_USE_REMOVE_VOLATILE_FALLBACK)
+template <class _Tp>
+struct remove_volatile
+{
+  using type _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_VOLATILE(_Tp);
+};
+
+template <class _Tp>
+using remove_volatile_t _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_REMOVE_VOLATILE(_Tp);
+
+#else
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_volatile
+{
+  using type = _Tp;
+};
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT remove_volatile<volatile _Tp>
+{
+  using type = _Tp;
+};
+
+template <class _Tp>
+using remove_volatile_t _CCCL_NODEBUG_ALIAS = typename remove_volatile<_Tp>::type;
+
+#endif // defined(_CCCL_BUILTIN_REMOVE_VOLATILE) && !defined(_LIBCUDACXX_USE_REMOVE_VOLATILE_FALLBACK)
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_REMOVE_VOLATILE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/sfinae_traits.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/sfinae_traits.h
new file mode 100644
index 00000000..dc5ba3a6
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/sfinae_traits.h
@@ -0,0 +1,63 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_SFINAE_TRAITS_H
+#define _CUDA_STD___TYPE_TRAITS_SFINAE_TRAITS_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+//! @brief Determines whether a constructor is valid and whether it is implicit or explicit
+enum class __select_constructor
+{
+  __invalid, //!< The constructor is not valid
+  __implicit, //!< The constructor is valid and implicit
+  __explicit, //!< The constructor is valid and explicit
+  __deleted, //!< The constructor is marked as deleted
+};
+
+template <__select_constructor _Trait>
+inline constexpr bool __can_construct_implicitly = _Trait == __select_constructor::__implicit;
+template <__select_constructor _Trait>
+inline constexpr bool __can_construct_explicitly = _Trait == __select_constructor::__explicit;
+template <__select_constructor _Trait>
+inline constexpr bool __can_construct =
+  (_Trait == __select_constructor::__implicit) || (_Trait == __select_constructor::__explicit);
+template <__select_constructor _Trait>
+inline constexpr bool __is_deleted = _Trait == __select_constructor::__deleted;
+
+//! @brief Determines whether an assignment is valid and also whether it does not throw
+enum class __select_assignment
+{
+  __invalid, //!< The assignment is invalid
+  __is_nothrow, //!< The assignment is valid and noexcept
+  __may_throw, //!< The assignment is valid but may throw
+};
+
+template <__select_assignment _Trait>
+inline constexpr bool __can_assign = _Trait != __select_assignment::__invalid;
+template <__select_assignment _Trait>
+inline constexpr bool __can_nothrow_assign = _Trait == __select_assignment::__is_nothrow;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_SFINAE_TRAITS_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_identity.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_identity.h
new file mode 100644
index 00000000..2d695747
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_identity.h
@@ -0,0 +1,40 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_TYPE_IDENTITY_H
+#define _CUDA_STD___TYPE_TRAITS_TYPE_IDENTITY_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+struct type_identity
+{
+  using type = _Tp;
+};
+
+template <class _Tp>
+using type_identity_t _CCCL_NODEBUG_ALIAS = typename type_identity<_Tp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_TYPE_IDENTITY_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_list.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_list.h
new file mode 100644
index 00000000..4b7c7383
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_list.h
@@ -0,0 +1,1098 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023-24 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_TYPE_LIST_H
+#define _CUDA_STD___TYPE_TRAITS_TYPE_LIST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/type_identity.h>
+#include <cuda/std/__type_traits/type_set.h>
+#include <cuda/std/__type_traits/void_t.h>
+#include <cuda/std/__utility/integer_sequence.h>
+#include <cuda/std/__utility/undefined.h>
+
+//! \file type_list.h
+//! This file defines a type-list type and some fundamental algorithms on type
+//! lists.
+//!
+//! It also defines a "meta-callable" protocol for parameterizing the type list
+//! algorithms, and some higher-order meta-callables. A meta-callable is a class
+//! type with a nested \c __call template alias.
+//!
+//! For the purpose of this file, a "trait type" is a class type with a nested
+//! type alias named \c type.
+
+#if !defined(_CCCL_META_UNROLL_LIMIT)
+#  if _CCCL_CUDA_COMPILER(NVCC) || _CCCL_COMPILER(NVHPC)
+#    define _CCCL_META_UNROLL_LIMIT 10
+#  else
+#    define _CCCL_META_UNROLL_LIMIT 16
+#  endif
+#endif
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#ifndef _CCCL_DOXYGEN_INVOKED // Do not document
+
+template <class... _Ts>
+struct __type_list;
+
+template <class _Ty>
+using __type = typename _Ty::type;
+
+//! \brief Evaluate a meta-callable with the given arguments
+template <class _Fn, class... _Ts>
+using __type_call = typename _Fn::template __call<_Ts...>;
+
+//! \brief Evaluate a unary meta-callable with the given argument
+template <class _Fn, class _Ty>
+using __type_call1 = typename _Fn::template __call<_Ty>;
+
+//! \brief Evaluate a binary meta-callable with the given arguments
+template <class _Fn, class _Ty, class _Uy>
+using __type_call2 = typename _Fn::template __call<_Ty, _Uy>;
+
+namespace __detail
+{
+template <size_t _DependentValue>
+struct __type_call_indirect_fn
+{
+  template <template <class...> class _Fn, class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Ts...>;
+};
+} // namespace __detail
+
+//! \brief Evaluate a meta-callable with the given arguments, with an indirection
+//! to avoid the dreaded "pack expansion argument for non-pack parameter" error.
+template <class _Fn, class... _Ts>
+using __type_call_indirect = //
+  typename __detail::__type_call_indirect_fn<sizeof(
+    __type_list<_Fn, _Ts...>*)>::template __call<_Fn::template __call, _Ts...>;
+
+//! \brief Turns a class or alias template into a meta-callable
+template <template <class...> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Ts...>;
+};
+
+//! \brief Turns a unary class or alias template into a meta-callable
+template <template <class> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote1
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Ty>;
+};
+
+//! \brief Turns a binary class or alias template into a meta-callable
+template <template <class, class> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote2
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Ty, _Uy>;
+};
+
+//! \brief Turns a trait class template \c _Fn into a meta-callable \c
+//! __type_quote_trait such that \c __type_quote_trait<_Fn>::__call<_Ts...> is
+//! equivalent to \c _Fn<_Ts...>::type.
+template <template <class...> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote_trait
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type<_Fn<_Ts...>>;
+};
+
+//! \brief Turns a unary trait class template \c _Fn into a meta-callable
+//! \c __type_quote_trait1 such that \c __type_quote_trait1<_Fn>::__call<_Ty> is
+//! equivalent to \c _Fn<_Ty>::type.
+template <template <class> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote_trait1
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = __type<_Fn<_Ty>>;
+};
+
+//! \brief Turns a binary trait class template \c _Fn into a meta-callable
+//! \c __type_quote_trait2 such that \c __type_quote_trait2<_Fn>::__call<_Ty,_Uy>
+//! is equivalent to \c _Fn<_Ty,_Uy>::type.
+template <template <class, class> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_quote_trait2
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = __type<_Fn<_Ty, _Uy>>;
+};
+
+//! \brief Adds an indirection to a meta-callable to avoid the dreaded "pack
+//! expansion argument for non-pack parameter" error.
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_indirect
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = typename __detail::__type_call_indirect_fn<sizeof(
+    __type_list<_Ts...>*)>::template __call<_Fn::template __call, _Ts...>;
+};
+
+//! \brief Turns a class or alias template into a meta-callable with an
+//! indirection to avoid the dreaded "pack expansion argument for non-pack
+//! parameter" error.
+template <template <class...> class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_indirect_quote
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS =
+    typename __detail::__type_call_indirect_fn<sizeof(__type_list<_Ts...>*)>::template __call<_Fn, _Ts...>;
+};
+
+//! \brief A meta-callable that composes two meta-callables
+template <class _Fn1, class _Fn2>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_compose
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call1<_Fn1, __type_call<_Fn2, _Ts...>>;
+};
+
+template <template <class...> class _Fn, class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_bind_front_quote
+{
+  template <class... _Us>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Ts..., _Us...>;
+};
+
+//! \brief A meta-callable that binds the front arguments to a meta-callable
+template <class _Fn, class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_bind_front : __type_bind_front_quote<_Fn::template __call, _Ts...>
+{};
+
+template <template <class...> class _Fn, class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_bind_back_quote
+{
+  template <class... _Us>
+  using __call _CCCL_NODEBUG_ALIAS = _Fn<_Us..., _Ts...>;
+};
+
+//! \brief A meta-callable that binds the back arguments to a meta-callable
+template <class _Fn, class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_bind_back : __type_bind_back_quote<_Fn::template __call, _Ts...>
+{};
+
+//! \brief A meta-callable that always evaluates to \c _Ty.
+template <class _Ty>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_always
+{
+  template <class...>
+  using __call _CCCL_NODEBUG_ALIAS = _Ty;
+};
+
+//! \brief A unary meta-callable that returns its argument unmodified.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_self
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = _Ty;
+};
+
+template <class _Ty>
+using __type_self_t _CCCL_NODEBUG_ALIAS = _Ty;
+
+//! \brief Perform a logical AND operation on a list of Boolean types.
+//!
+//! \note The AND operation is not short-circuiting.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_strict_and
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __fold_and<_Ts::value...>;
+};
+
+//! \brief Perform a logical OR operation on a list of Boolean types.
+//!
+//! \note The OR operation is not short-circuiting.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_strict_or
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __fold_or<_Ts::value...>;
+};
+
+//! \brief Perform a logical NOT operation on a Boolean type.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_not
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(!_Ty::value)>;
+};
+
+//! \brief Test whether two integral constants are equal.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_equal
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value == _Uy::value)>;
+};
+
+//! \brief Test whether two integral constants are not equal.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_not_equal
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value != _Uy::value)>;
+};
+
+//! \brief Test whether one integral constant is less than another.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_less
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value < _Uy::value)>;
+};
+
+//! \brief Test whether one integral constant is less than or equal to another.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_less_equal
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value <= _Uy::value)>;
+};
+
+//! \brief Test whether one integral constant is greater than another.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_greater
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value > _Uy::value)>;
+};
+
+//! \brief Test whether one integral constant is greater than or equal to another.
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_greater_equal
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<(_Ty::value >= _Uy::value)>;
+};
+
+//! \brief A functional adaptor that negates a unary predicate
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_negate1
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call1<__type_not, __type_call1<_Fn, _Ty>>;
+};
+
+//! \brief A functional adaptor that negates a binary predicate
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_negate2
+{
+  template <class _Ty, class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call1<__type_not, __type_call2<_Fn, _Ty, _Uy>>;
+};
+
+//! \brief A type list
+template <class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_list
+{
+  static constexpr size_t const __size = sizeof...(_Ts);
+
+  // A type_list behaves like a meta-callable
+  // that takes a meta-callable and applies the
+  // elements of the list to it.
+  template <class _Fn, class... _Us>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call<_Fn, _Ts..., _Us...>;
+};
+
+// Before the addition of inline variables, it was necessary to
+// provide a definition for constexpr class static data members.
+template <class... _Ts>
+constexpr size_t const __type_list<_Ts...>::__size;
+
+//! \brief A pointer to a type list, often used as a function argument.
+template <class... _Ts>
+using __type_list_ptr = __type_list<_Ts...>*;
+
+//! \brief Return the size of a type list.
+template <class _List>
+using __type_list_size = integral_constant<size_t, _List::__size>;
+
+//! \brief Given a type list and a list of types, append the types to the list.
+template <class _List, class... _Ts>
+using __type_push_back = __type_call<_List, __type_quote<__type_list>, _Ts...>;
+
+//! \brief Given a type list and a list of types, prepend the types to the list.
+template <class _List, class... _Ts>
+using __type_push_front = __type_call1<_List, __type_bind_front_quote<__type_list, _Ts...>>;
+
+#  if _CCCL_CUDA_COMPILER(NVCC, <, 12, 9)
+
+namespace __detail
+{
+template <template <class...> class _Fn, class... _Ts>
+_CCCL_API inline auto __as_type_list_fn(__undefined<_Fn<_Ts...>>*) //
+  -> __type_list<_Ts...>;
+
+template <template <class _Ty, _Ty...> class _Fn, class _Ty, _Ty... _Us>
+_CCCL_API inline auto __as_type_list_fn(__undefined<_Fn<_Ty, _Us...>>*) //
+  -> __type_list<integral_constant<_Ty, _Us>...>;
+
+template <class _Ret, class... _Args>
+_CCCL_API inline auto __as_type_list_fn(__undefined<_Ret(_Args...)>*) //
+  -> __type_list<_Ret, _Args...>;
+} // namespace __detail
+
+//! \brief Given a type that is can be interpreted as a type list, return its
+//! type list interpretation. Types that can be interpreted as a type
+//! list are of the following forms:
+//!
+//! \li `C<Ts...>`, for any class template `C` and types `Ts...`.
+//! \li `C<T, T... Vs>`, for any class template `C`, type `T` and values `Vs...`.
+//!     The resulting type is `_Fn<integral_constant<T, Vs>...>`.
+//! \li `R(As...)`, for any function type `R(As...)`. The resulting type is
+//!     `_Fn<R, As...>`.
+template <class _List>
+using __as_type_list = decltype(__detail::__as_type_list_fn(static_cast<__undefined<_List>*>(nullptr)));
+
+#  else // ^^^ if _CCCL_CUDA_COMPILER(NVCC, <, 12, 9) / vvv if _CCCL_CUDA_COMPILER(NVCC, >=, 12, 9) vvv
+
+namespace __detail
+{
+template <class _Ret, class... _Args>
+using __fn_ptr_t = _Ret (*)(_Args...);
+
+template <class _List>
+extern __undefined<_List> __as_type_list_v;
+
+template <template <class...> class _Cy, class... _Ts>
+extern __fn_ptr_t<__type_list<_Ts...>> __as_type_list_v<_Cy<_Ts...>>;
+
+template <template <class _Ty, _Ty...> class _Fn, class _Ty, _Ty... _Us>
+extern __fn_ptr_t<__type_list<std::integral_constant<_Ty, _Us>...>> __as_type_list_v<_Fn<_Ty, _Us...>>;
+
+template <class _Ret, class... _Args>
+extern __fn_ptr_t<__type_list<_Ret, _Args...>> __as_type_list_v<_Ret(_Args...)>;
+} // namespace __detail
+
+//! \brief Given a type that can be interpreted as a type list, return its
+//! type list interpretation. Types that can be interpreted as a type
+//! list are of the following forms:
+//!
+//! \li `C<Ts...>`, for any class template `C` and types `Ts...`.
+//! \li `C<T, T... Vs>`, for any class template `C`, type `T` and values `Vs...`.
+//!     The resulting type is `_Fn<integral_constant<T, Vs>...>`.
+//! \li `R(As...)`, for any function type `R(As...)`. The resulting type is
+//!     `_Fn<R, As...>`.
+template <class _List>
+using __as_type_list = decltype(__detail::__as_type_list_v<_List>());
+
+#  endif // if _CCCL_CUDA_COMPILER(NVCC, >=, 12, 9)
+
+//! \brief Given a type that can be interpreted as a type list and a
+//! meta-callable, invoke the meta-callable with the types in the list.
+//! \c __as_type_list is used to convert the \c _List type to a specialization
+//! of \c __type_list.
+//!
+//! \sa __as_type_list
+template <class _Fn, class _List>
+using __type_apply = __type_call<__as_type_list<_List>, _Fn>;
+
+//! \brief Given a type that can be interpreted as a type list and a class or
+//! alias template, instantiate the template with the types in the list.
+//! \c __as_type_list is used to convert the \c _List type to a specialization
+//! of \c __type_list.
+//!
+//! \sa __as_type_list
+template <template <class...> class _Fn, class _List>
+using __type_apply_q = __type_call<__as_type_list<_List>, __type_quote<_Fn>>;
+
+namespace __detail
+{
+// Only the following precise formulation works with nvcc < 12.2
+template <class _Fn, class... _Ts, template <class...> class _Fn2 = _Fn::template __call, class = _Fn2<_Ts...>>
+_CCCL_API inline auto __type_callable_fn(__type_list<_Fn, _Ts...>*) -> true_type;
+
+_CCCL_API inline auto __type_callable_fn(void*) -> false_type;
+} // namespace __detail
+
+//! \brief Test whether a meta-callable is callable with a given set of
+//! arguments.
+//!
+//! Given a meta-callable and a list of arguments, return \c true_type if
+//! the meta-callable can be called with the arguments, and \c false_type
+//! otherwise.
+template <class _Fn, class... _Ts>
+using __type_callable = decltype(__detail::__type_callable_fn(static_cast<__type_list<_Fn, _Ts...>*>(nullptr)));
+
+namespace __detail
+{
+template <bool _IsCallable>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_defer_fn
+{
+  template <class, class...>
+  using __call _CCCL_NODEBUG_ALIAS = __type_defer_fn;
+};
+
+template <>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_defer_fn<true>
+{
+  template <class _Fn, class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = type_identity<__type_call<_Fn, _Ts...>>;
+};
+} // namespace __detail
+
+//! \brief Defer the evaluation of a meta-callable with a list of arguments.
+//!
+//! Given a meta-callable and a list of arguments, return a trait type \c T
+//! where \c T::type is the result of evaluating the meta-callable with the
+//! arguments, or if the meta-callable is not callable with the arguments, a
+//! class type without a nested \c ::type type alias.
+template <class _Fn, class... _Ts>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT
+__type_defer : __type_call<__detail::__type_defer_fn<__type_callable<_Fn, _Ts...>::value>, _Fn, _Ts...>
+{};
+
+//! \brief Defer the instantiation of a template with a list of arguments.
+//!
+//! Given a variadic template and a list of arguments, return a trait type \c T
+//! where \c T::type is the result of instantiating the template with the
+//! arguments, or if the template cannot be instantiated with the arguments, a
+//! class type without a nested \c ::type type alias.
+template <template <class...> class _Fn, class... _Ts>
+using __type_defer_quote = __type_defer<__type_quote<_Fn>, _Ts...>;
+
+//! \brief A composition of two meta-callables that will attempt to call the
+//! first, and if that fails, call the second.
+//!
+//! Compose two meta-callables, \c _TryFn and \c _CatchFn, into a new
+//! meta-callable. Given arguments \c _Ts... , \c __type_try_catch will try to
+//! evaluate `__type_call<_TryFn, _Ts...>`. If this type is well-formed, it will
+//! be the result of the composition. Otherwise, the result will be
+//! `__type_call<_CatchFn, _Ts...>`.
+template <class _TryFn, class _CatchFn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_try_catch
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call<_If<__type_callable<_TryFn, _Ts...>::value, _TryFn, _CatchFn>, _Ts...>;
+};
+
+// Implementation for indexing into a list of types:
+#  if _CCCL_HAS_PACK_INDEXING()
+
+template <size_t _Ip, class... _Ts>
+using __type_index_c = _Ts...[_Ip];
+
+template <class _Ip, class... _Ts>
+using __type_index = _Ts...[_Ip::value];
+
+// Versions of nvcc prior to 12.0 have trouble with pack expansion into
+// __type_pack_element in an alias template, so we use the fall-back
+// implementation instead.
+#  elif defined(_CCCL_BUILTIN_TYPE_PACK_ELEMENT)
+
+namespace __detail
+{
+// On some versions of gcc, __type_pack_element cannot be mangled so
+// hide it behind a named template.
+template <size_t _Ip>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_index_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_TYPE_PACK_ELEMENT(_Ip, _Ts...);
+};
+} // namespace __detail
+
+template <size_t _Ip, class... _Ts>
+using __type_index_c = __type_call<__detail::__type_index_fn<_Ip>, _Ts...>;
+
+template <class _Ip, class... _Ts>
+using __type_index = __type_call<__detail::__type_index_fn<_Ip::value>, _Ts...>;
+
+#  else // ^^^ _CCCL_BUILTIN_TYPE_PACK_ELEMENT ^^^ / vvv !_CCCL_BUILTIN_TYPE_PACK_ELEMENT vvv
+
+// Fallback implementation for __type_index uses multiple inheritance. See:
+// https://ldionne.com/2015/11/29/efficient-parameter-pack-indexing/
+
+namespace __detail
+{
+template <size_t _Ip, class _Ty>
+struct __type_tuple_elem
+{
+  using type _CCCL_NODEBUG_ALIAS = _Ty;
+};
+
+template <class _Is, class... _Ts>
+struct __type_tupl;
+
+template <size_t... _Is, class... _Ts>
+struct __type_tupl<index_sequence<_Is...>, _Ts...> : __type_tuple_elem<_Is, _Ts>...
+{};
+
+template <class... _Ts>
+using __type_tuple = __type_tupl<make_index_sequence<sizeof...(_Ts)>, _Ts...>;
+
+template <size_t _Ip, class _Ty>
+_CCCL_API inline __type_tuple_elem<_Ip, _Ty> __type_tuple_get(__type_tuple_elem<_Ip, _Ty>);
+
+template <size_t _Ip, class... _Ts>
+using __type_tuple_element_t _CCCL_NODEBUG_ALIAS =
+  __type<decltype(__detail::__type_tuple_get<_Ip>(__type_tuple<_Ts...>{}))>;
+
+struct __type_index_large_size_fn
+{
+  template <class _Ip, class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_tuple_element_t<_Ip::value, _Ts...>;
+};
+
+template <size_t _Ip>
+struct __type_index_small_size_fn;
+
+#    define _M0(X) class,
+#    define _M1(_N)                                             \
+      template <>                                               \
+      struct __type_index_small_size_fn<_N>                     \
+      {                                                         \
+        template <_CCCL_PP_REPEAT(_N, _M0) class _Ty, class...> \
+        using __call _CCCL_NODEBUG_ALIAS = _Ty;                 \
+      };
+
+_CCCL_PP_REPEAT_REVERSE(_CCCL_META_UNROLL_LIMIT, _M1)
+
+#    undef _M0
+#    undef _M1
+
+template <bool _IsSmall>
+struct __type_index_select_fn : __type_index_large_size_fn // Default for larger indices
+{};
+
+template <>
+struct __type_index_select_fn<true> // Fast implementation for smaller indices
+{
+  template <class _Ip, class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call_indirect<__type_index_small_size_fn<_Ip::value>, _Ts...>;
+};
+} // namespace __detail
+
+#    if !_CCCL_COMPILER(MSVC)
+
+template <class _Ip, class... _Ts>
+using __type_index = __type_call<__detail::__type_index_select_fn<(_Ip::value < _CCCL_META_UNROLL_LIMIT)>, _Ip, _Ts...>;
+
+#    else // ^^^ !_CCCL_COMPILER(MSVC) ^^^ / vvv _CCCL_COMPILER(MSVC) vvv
+
+// Simplify the implementation for MSVC, which has trouble with the above
+template <class _Ip, class... _Ts>
+using __type_index = __detail::__type_index_large_size_fn::__call<_Ip, _Ts...>;
+
+#    endif // !_CCCL_COMPILER(MSVC)
+
+template <size_t _Ip, class... _Ts>
+using __type_index_c = __type_index<integral_constant<size_t, _Ip>, _Ts...>;
+
+#  endif // !_CCCL_BUILTIN_TYPE_PACK_ELEMENT
+
+namespace __detail
+{
+template <size_t _Ip>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_at_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_index_c<_Ip, _Ts...>;
+};
+} // namespace __detail
+
+//! \brief Given a type list and an index, return the type at that index.
+template <size_t _Ip, class _List>
+using __type_at_c = __type_call1<_List, __detail::__type_at_fn<_Ip>>;
+
+//! \brief Given a type list and an index, return the type at that index.
+template <class _Ip, class _List>
+using __type_at = __type_call1<_List, __detail::__type_at_fn<_Ip::value>>;
+
+//! \brief Given a type list return the type at the front of the list.
+template <class _List>
+using __type_front = __type_at_c<0, _List>;
+
+//! \brief Given a type list return the type at the back of the list.
+template <class _List>
+using __type_back = __type_at_c<_List::__size - 1, _List>;
+
+//! \brief A pair of types
+template <class _First, class _Second>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_pair
+{
+  using __first _CCCL_NODEBUG_ALIAS  = _First;
+  using __second _CCCL_NODEBUG_ALIAS = _Second;
+};
+
+//! \brief Retrieve the first of a pair of types
+//! \pre \c _Pair is a specialization of \c __type_pair
+template <class _Pair>
+using __type_pair_first _CCCL_NODEBUG_ALIAS = typename _Pair::__first;
+
+//! \brief Retrieve the second of a pair of types
+//! \pre \c _Pair is a specialization of \c __type_pair
+template <class _Pair>
+using __type_pair_second _CCCL_NODEBUG_ALIAS = typename _Pair::__second;
+
+//! \see __type_switch
+template <class _Value>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_default
+{
+  template <class>
+  using __rebind _CCCL_NODEBUG_ALIAS = __type_default;
+
+  using type _CCCL_NODEBUG_ALIAS = _Value;
+};
+
+#  if defined(_CCCL_DOXYGEN_INVOKED)
+
+//! \see __type_switch
+template <auto _Label, class _Value>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_case
+{
+  template <class _OtherInt>
+  using __rebind _CCCL_NODEBUG_ALIAS = __type_case<static_cast<_OtherInt>(_Label), _Value>;
+
+  using type = _Value;
+};
+
+#  else // ^^^ DOXYGEN ^^^ / vvv !DOXYGEN vvv
+
+template <class _Label, class _Value>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_case_
+{
+  template <class _OtherInt>
+  using __rebind _CCCL_NODEBUG_ALIAS = __type_case_<integral_constant<_OtherInt, _Label::value>, _Value>;
+
+  using type = _Value;
+};
+
+template <auto _Label, class _Value>
+using __type_case _CCCL_NODEBUG_ALIAS = __type_case_<integral_constant<decltype(_Label), _Label>, _Value>;
+
+#  endif // !DOXYGEN
+
+namespace __detail
+{
+template <auto _Label, class _Value>
+_CCCL_API inline auto __type_switch_fn(__type_case<_Label, _Value>*, int) -> __type_case<_Label, _Value>;
+
+template <auto _Label, class _Value>
+_CCCL_API inline auto __type_switch_fn(__type_default<_Value>*, long) -> __type_default<_Value>;
+} // namespace __detail
+
+//! \see __type_switch
+template <class _Type, class... _Cases>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT _CCCL_DECLSPEC_EMPTY_BASES __type_switch_fn : _Cases::template __rebind<_Type>...
+{
+  template <class _Label>
+  using __call _CCCL_NODEBUG_ALIAS =
+    __type<decltype(__detail::__type_switch_fn<_Label::value>(static_cast<__type_switch_fn*>(nullptr), 0))>;
+};
+
+//! \brief Given an integral constant \c _Label and a pack of "cases"
+//! consisting of one or more specializations of \c __type_case and zero or
+//! one specializations of \c __type_default, `__type_switch<_Label, _Cases...>`
+//! returns the value associated with the first case whose label matches the
+//! given label. If no such case exists, the value associated with the default
+//! case is returned. If no default case exists, the type is ill-formed.
+//!
+//! \p Example:
+//! \code
+//! using result = __type_switch<2,
+//!                              __type_case<1, char>,
+//!                              __type_case<2, double>,
+//!                              __type_default<float>>;
+//! static_assert(is_same_v<result, double>);
+//! \endcode
+template <auto _Label, class... _Cases>
+using __type_switch _CCCL_NODEBUG_ALIAS =
+  __type_call<__type_switch_fn<decltype(_Label), _Cases...>, integral_constant<decltype(_Label), _Label>>;
+
+namespace __detail
+{
+#  if _CCCL_COMPILER(MSVC, <, 19, 38)
+// A workaround for https://developercommunity.visualstudio.com/t/fatal-error-C1001:-Internal-compiler-err/10405847
+struct __type_concat_fn
+{
+  template <class... _Lists>
+  struct __trait
+  {};
+
+  template <class... _Ts, class... _Us, class... _Lists>
+  struct __trait<__type_list<_Ts...>, __type_list<_Us...>, _Lists...> : __trait<__type_list<_Ts..., _Us...>, _Lists...>
+  {};
+
+  template <class... _Ts>
+  struct __trait<__type_list<_Ts...>>
+  {
+    using type = __type_list<_Ts...>;
+  };
+
+  template <>
+  struct __trait<>
+  {
+    using type = __type_list<>;
+  };
+
+  template <class... _Lists>
+  using __call _CCCL_NODEBUG_ALIAS = __type<__trait<_Lists...>>;
+};
+#  else // ^^^ _CCCL_COMPILER(MSVC, <, 19, 38) ^^^ / vvv _CCCL_COMPILER(MSVC, >=, 19, 38) vvv
+template <size_t _Count>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_maybe_concat_fn
+{
+  using __next_cpo _CCCL_NODEBUG_ALIAS = __type_maybe_concat_fn<(_Count < 8 ? 0 : _Count - 8)>;
+
+  template <class... _Ts,
+            class... _As,
+            class... _Bs,
+            class... _Cs,
+            class... _Ds,
+            class... _Es,
+            class... _Fs,
+            class... _Gs,
+            class... _Hs,
+            class... _Tail>
+  _CCCL_API inline static auto __fn(
+    __type_list_ptr<_Ts...>, // state
+    __type_list_ptr<_As...>, // 1
+    __type_list_ptr<_Bs...>, // 2
+    __type_list_ptr<_Cs...>, // 3
+    __type_list_ptr<_Ds...>, // 4
+    __type_list_ptr<_Es...>, // 5
+    __type_list_ptr<_Fs...>, // 6
+    __type_list_ptr<_Gs...>, // 7
+    __type_list_ptr<_Hs...>, // 8
+    _Tail*... __tail) // rest
+    -> decltype(__next_cpo::__fn(
+      __type_list_ptr<_Ts..., _As..., _Bs..., _Cs..., _Ds..., _Es..., _Fs..., _Gs..., _Hs...>{nullptr},
+      __tail...,
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr},
+      __type_list_ptr<>{nullptr}));
+};
+
+template <>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_maybe_concat_fn<0>
+{
+  template <class... _Ts>
+  _CCCL_API inline static auto __fn(__type_list_ptr<_Ts...>, ...) -> __type_list<_Ts...>;
+};
+
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_concat_fn
+{
+  template <class... _Lists>
+  using __call _CCCL_NODEBUG_ALIAS = decltype(__type_maybe_concat_fn<sizeof...(_Lists)>::__fn(
+    __type_list_ptr<>{nullptr},
+    static_cast<_Lists*>(nullptr)...,
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr},
+    __type_list_ptr<>{nullptr}));
+};
+#  endif // _CCCL_COMPILER(MSVC, >=, 19, 38)
+} // namespace __detail
+
+//! \brief Concatenate a list of type lists into a single type list.
+//!
+//! When passed no type lists, \c __type_concat returns an empty type list.
+template <class... _Lists>
+using __type_concat = __type_call<__detail::__type_concat_fn, _Lists...>;
+
+//! \brief Given a list of type lists, concatenate all the lists into one.
+//!
+//! When passed an empty type list, \c __type_flatten returns an empty type list.
+template <class _ListOfLists>
+using __type_flatten = __type_call1<_ListOfLists, __type_quote<__type_concat>>;
+
+namespace __detail
+{
+template <bool _IsEmpty>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_maybe_find_if_fn;
+
+template <bool _Found>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_find_if_found
+{
+  template <class _Fn, class _Head, class... _Tail>
+  using __call _CCCL_NODEBUG_ALIAS = __type_list<_Head, _Tail...>;
+};
+
+template <>
+struct __type_find_if_found<false>
+{
+  template <class _Fn, class _Head, class... _Tail>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call<__type_maybe_find_if_fn<sizeof...(_Tail) == 0>, _Fn, _Tail...>;
+};
+
+template <bool _IsEmpty>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_maybe_find_if_fn // Type list is not empty
+{
+  template <class _Fn, class _Head, class... _Tail>
+  using __call _CCCL_NODEBUG_ALIAS =
+    __type_call<__type_find_if_found<__type_call1<_Fn, _Head>::value>, _Fn, _Head, _Tail...>;
+};
+
+template <>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_maybe_find_if_fn<true> // Type list is empty
+{
+  template <class, class... _None>
+  using __call _CCCL_NODEBUG_ALIAS = __type_list<>;
+};
+
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_find_if_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call<__type_maybe_find_if_fn<sizeof...(_Ts) == 0>, _Fn, _Ts...>;
+};
+
+template <class _Ty>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_same_as
+{
+  template <class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = bool_constant<is_same_v<_Ty, _Uy>>;
+};
+} // namespace __detail
+
+//! \brief Given a type list and a predicate, find the first type in the list
+//! that satisfies the predicate. It returns a type list containing the first
+//! type that satisfies the predicate and all the types after it.
+//!
+//! If no type in the list satisfies the predicate, \c __type_find_if returns an
+//! empty type list.
+template <class _List, class _Fn>
+using __type_find_if = __type_call1<_List, __detail::__type_find_if_fn<_Fn>>;
+
+//! \brief Given a type list and type, find the first occurrence of the type in
+//! the list. It returns a type list containing the type and all the types after
+//! it.
+//!
+//! If the type is not in the list, \c __type_find returns an empty type list.
+template <class _List, class _Ty>
+using __type_find = __type_find_if<_List, __detail::__type_same_as<_Ty>>;
+
+namespace __detail
+{
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_transform_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_list<__type_call1<_Fn, _Ts>...>;
+};
+} // namespace __detail
+
+//! \brief Given a type list and a unary meta-callable, apply the meta-callable
+//! to each type in the list. It returns a new type list containing the results.
+template <class _List, class _Fn>
+using __type_transform = __type_call1<_List, __detail::__type_transform_fn<_Fn>>;
+
+namespace __detail
+{
+template <size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_right_fn;
+
+template <size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_left_fn;
+
+#  define _M0(_N) , class _CCCL_PP_CAT(_, _N)
+#  define _M1(_N) typename _Fn::template __call <
+#  define _M2(_N) , _CCCL_PP_CAT(_, _N)
+#  define _M3(_N) _M2(_N) >
+#  define _M4(_N) _M2(_CCCL_PP_DEC(_N)) >
+
+#  define _LIBCUDACXX_TYPE_LIST_FOLD_RIGHT(_N)                                                     \
+    template <>                                                                                    \
+    struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_right_fn<_N>                                  \
+    {                                                                                              \
+      template <class _Fn, class _State _CCCL_PP_REPEAT(_N, _M0)>                                  \
+      using __call _CCCL_NODEBUG_ALIAS = _CCCL_PP_REPEAT(_N, _M1) _State _CCCL_PP_REPEAT(_N, _M3); \
+    };
+
+_CCCL_PP_REPEAT_REVERSE(_CCCL_PP_INC(_CCCL_META_UNROLL_LIMIT), _LIBCUDACXX_TYPE_LIST_FOLD_RIGHT)
+
+template <size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_right_fn
+{
+  template <class _Fn, class _State _CCCL_PP_REPEAT(_CCCL_META_UNROLL_LIMIT, _M0), class... _Rest>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call_indirect<
+    __type_fold_right_fn<_Np - _CCCL_META_UNROLL_LIMIT>,
+    _Fn,
+    __type_call<__type_fold_right_fn<_CCCL_META_UNROLL_LIMIT>, _Fn, _State _CCCL_PP_REPEAT(_CCCL_META_UNROLL_LIMIT, _M2)>,
+    _Rest...>;
+};
+
+template <class _Init, class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_right_select_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call_indirect<__type_fold_right_fn<sizeof...(_Ts)>, _Fn, _Init, _Ts...>;
+};
+
+#  define _LIBCUDACXX_TYPE_FOLD_LEFT(_N)                                                                             \
+    template <>                                                                                                      \
+    struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_left_fn<_N>                                                     \
+    {                                                                                                                \
+      template <class _Fn, class _State _CCCL_PP_REPEAT(_N, _M0)>                                                    \
+      using __call _CCCL_NODEBUG_ALIAS = _CCCL_PP_REPEAT(_N, _M1) _State _CCCL_PP_REPEAT(_N, _M4, _N, _CCCL_PP_DEC); \
+    };
+
+_CCCL_PP_REPEAT_REVERSE(_CCCL_PP_INC(_CCCL_META_UNROLL_LIMIT), _LIBCUDACXX_TYPE_FOLD_LEFT)
+
+template <size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_left_fn
+{
+  template <class _Fn, class _State _CCCL_PP_REPEAT(_CCCL_META_UNROLL_LIMIT, _M0), class... _Rest>
+  using __call _CCCL_NODEBUG_ALIAS =
+    __type_call<__type_fold_left_fn<_CCCL_META_UNROLL_LIMIT>,
+                _Fn,
+                __type_call_indirect<__type_fold_left_fn<_Np - _CCCL_META_UNROLL_LIMIT>, _Fn, _State, _Rest...> //
+                  _CCCL_PP_REPEAT(_CCCL_META_UNROLL_LIMIT, _M2, 0, _CCCL_PP_INC)>;
+};
+
+template <class _Init, class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_fold_left_select_fn
+{
+  template <class... _Ts>
+  using __call _CCCL_NODEBUG_ALIAS = __type_call_indirect<__type_fold_left_fn<sizeof...(_Ts)>, _Fn, _Init, _Ts...>;
+};
+
+#  undef _LIBCUDACXX_TYPE_FOLD_LEFT
+#  undef _LIBCUDACXX_TYPE_FOLD_RIGHT
+
+#  undef _M4
+#  undef _M3
+#  undef _M2
+#  undef _M1
+#  undef _M0
+} // namespace __detail
+
+//! \brief Fold a type list from the right with a binary meta-callable and an
+//! initial state.
+template <class _List, class _Init, class _Fn>
+using __type_fold_right = __type_call1<_List, __detail::__type_fold_right_select_fn<_Init, _Fn>>;
+
+//! \brief Fold a type list from the left with a binary meta-callable and an
+//! initial state.
+template <class _List, class _Init, class _Fn>
+using __type_fold_left = __type_call1<_List, __detail::__type_fold_left_select_fn<_Init, _Fn>>;
+
+namespace __detail
+{
+template <class _Ty>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_remove_fn
+{
+  template <class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = _If<is_same_v<_Ty, _Uy>, __type_list<>, __type_list<_Uy>>;
+};
+} // namespace __detail
+
+//! \brief Remove all occurrences of a type from a type list
+template <class _List, class _Ty>
+using __type_remove = __type_flatten<__type_transform<_List, __detail::__type_remove_fn<_Ty>>>;
+
+namespace __detail
+{
+template <class _Fn>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_remove_if_fn
+{
+  template <class _Uy>
+  using __call _CCCL_NODEBUG_ALIAS = _If<__type_call1<_Fn, _Uy>::value, __type_list<>, __type_list<_Uy>>;
+};
+} // namespace __detail
+
+//! \brief Remove all types satisfying a unary predicate from a type list
+template <class _List, class _Fn>
+using __type_remove_if = __type_flatten<__type_transform<_List, __detail::__type_remove_if_fn<_Fn>>>;
+
+//! \brief Remove all types not satisfying a unary predicate from a type list
+template <class _List, class _Fn>
+using __type_copy_if = __type_flatten<__type_transform<_List, __detail::__type_remove_if_fn<__type_negate1<_Fn>>>>;
+
+//! \brief Remove all duplicate types from a type list
+template <class _List>
+using __type_unique = __as_type_list<__type_call<_List, __type_quote<__make_type_set>>>;
+
+namespace __detail
+{
+// _State is the list of type lists being built by the __type_fold_left
+template <class _State, class _List>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_cartesian_product_fn
+{
+  template <class _Ty>
+  struct __lambda0
+  {
+    template <class _List2>
+    using __lambda1 _CCCL_NODEBUG_ALIAS = __type_list<__type_push_front<_List2, _Ty>>;
+
+    using type _CCCL_NODEBUG_ALIAS = __type_flatten<__type_transform<_State, __type_quote1<__lambda1>>>;
+  };
+
+  using type _CCCL_NODEBUG_ALIAS = __type_flatten<__type_transform<_List, __type_quote_trait1<__lambda0>>>;
+};
+} // namespace __detail
+/// \endcond
+
+//! \brief Given a list of lists \c _Lists, return a new list of lists that is
+//! the cartesian product.
+//!
+//! \par Complexity
+//! `O(N * M)`, where `N` is the size of the outer list, and
+//! `M` is the size of the inner lists.
+template <class... _Lists>
+using __type_cartesian_product =
+  __type_fold_left<__type_list<_Lists...>,
+                   __type_list<__type_list<>>,
+                   __type_quote_trait2<__detail::__type_cartesian_product_fn>>;
+
+// A unary meta-callable for converting a type to its size in bytes
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_sizeof
+{
+  template <class _Ty>
+  using __call _CCCL_NODEBUG_ALIAS = integral_constant<size_t, sizeof(_Ty)>;
+};
+
+//! \brief A list of compile-time values, and a meta-callable that accepts a
+//! meta-callable and evaluates it with the values, each value wrapped in an
+//! integral constant wrapper.
+template <class _Ty, _Ty... _Values>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT __type_value_list : __type_list<integral_constant<_Ty, _Values>...>
+{
+  using __type _CCCL_NODEBUG_ALIAS = _Ty;
+};
+
+namespace __detail
+{
+template <class _Ty, _Ty _Start, _Ty _Stride, _Ty... _Is>
+_CCCL_API inline auto __type_iota_fn(integer_sequence<_Ty, _Is...>*)
+  -> __type_value_list<_Ty, _Ty(_Start + (_Is * _Stride))...>;
+} // namespace __detail
+
+//! \brief Return an \c __type_value_list of size \c _Size starting at \c _Start
+//! and incrementing by \c _Stride.
+template <class _Ty, _Ty _Start, _Ty _Size, _Ty _Stride = _Ty(1)>
+using __type_iota =
+  decltype(__detail::__type_iota_fn<_Ty, _Start, _Stride>(static_cast<make_integer_sequence<_Ty, _Size>*>(nullptr)));
+
+#endif // _CCCL_DOXYGEN_INVOKED
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_TYPE_LIST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_set.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_set.h
new file mode 100644
index 00000000..d9e08dc5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/type_set.h
@@ -0,0 +1,131 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_TYPE_SET_H
+#define _CUDA_STD___TYPE_TRAITS_TYPE_SET_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/conjunction.h>
+#include <cuda/std/__type_traits/fold.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_base_of.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/type_identity.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+struct __type_list;
+
+template <class _Set, class... _Ty>
+struct __type_set_contains : __fold_and<is_base_of_v<type_identity<_Ty>, _Set>...>
+{};
+
+template <class _Set, class... _Ty>
+inline constexpr bool __type_set_contains_v = __fold_and_v<is_base_of_v<type_identity<_Ty>, _Set>...>;
+
+namespace __set
+{
+template <class... _Ts>
+struct __tupl;
+
+template <>
+struct __tupl<>
+{
+  template <class _Ty>
+  using __maybe_insert _CCCL_NODEBUG_ALIAS = __tupl<_Ty>;
+
+  _CCCL_API static constexpr size_t __size() noexcept
+  {
+    return 0;
+  }
+};
+
+template <class _Ty, class... _Ts>
+struct __tupl<_Ty, _Ts...>
+    : type_identity<_Ty>
+    , __tupl<_Ts...>
+{
+  template <class _Uy>
+  using __maybe_insert _CCCL_NODEBUG_ALIAS = _If<__type_set_contains_v<__tupl, _Uy>, __tupl, __tupl<_Uy, _Ty, _Ts...>>;
+
+  _CCCL_API static constexpr size_t __size() noexcept
+  {
+    return sizeof...(_Ts) + 1;
+  }
+};
+
+template <bool _Empty>
+struct __bulk_insert
+{
+  template <class _Set, class...>
+  using __call _CCCL_NODEBUG_ALIAS = _Set;
+};
+
+template <>
+struct __bulk_insert<false>
+{
+#if _CCCL_COMPILER(MSVC, <, 19, 20)
+  template <class _Set, class _Ty, class... _Us>
+  _CCCL_API inline static auto __insert_fn(__type_list<_Ty, _Us...>*) ->
+    typename __bulk_insert<sizeof...(_Us) == 0>::template __call<typename _Set::template __maybe_insert<_Ty>, _Us...>;
+
+  template <class _Set, class... _Us>
+  using __call _CCCL_NODEBUG_ALIAS = decltype(__insert_fn<_Set>(static_cast<__type_list<_Us...>*>(nullptr)));
+#else
+  template <class _Set, class _Ty, class... _Us>
+  using __call _CCCL_NODEBUG_ALIAS =
+    typename __bulk_insert<sizeof...(_Us) == 0>::template __call<typename _Set::template __maybe_insert<_Ty>, _Us...>;
+#endif
+};
+} // namespace __set
+
+// When comparing sets for equality, use conjunction<> to short-circuit the set
+// comparison if the sizes are different.
+template <class _ExpectedSet, class... _Ts>
+using __type_set_eq =
+  conjunction<bool_constant<sizeof...(_Ts) == _ExpectedSet::__size()>, __type_set_contains<_ExpectedSet, _Ts...>>;
+
+template <class _ExpectedSet, class... _Ts>
+inline constexpr bool __type_set_eq_v = __type_set_eq<_ExpectedSet, _Ts...>::value;
+
+template <class... _Ts>
+using __type_set = __set::__tupl<_Ts...>;
+
+template <class _Set, class... _Ts>
+using __type_set_insert = typename __set::__bulk_insert<sizeof...(_Ts) == 0>::template __call<_Set, _Ts...>;
+
+template <class... _Ts>
+using __make_type_set = __type_set_insert<__type_set<>, _Ts...>;
+
+template <class _Ty, class... _Ts>
+struct __is_included_in : __fold_or<is_same_v<_Ty, _Ts>...>
+{};
+
+template <class _Ty, class... _Ts>
+inline constexpr bool __is_included_in_v = (is_same_v<_Ty, _Ts> || ...);
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_TYPE_SET_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/underlying_type.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/underlying_type.h
new file mode 100644
index 00000000..340a8fe4
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/underlying_type.h
@@ -0,0 +1,52 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_UNDERLYING_TYPE_H
+#define _CUDA_STD___TYPE_TRAITS_UNDERLYING_TYPE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_enum.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+#define _CCCL_BUILTIN_UNDERLYING_TYPE(...) __underlying_type(__VA_ARGS__)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, bool = is_enum_v<_Tp>>
+struct __cccl_underlying_type_impl
+{
+  using type = _CCCL_BUILTIN_UNDERLYING_TYPE(_Tp);
+};
+
+template <class _Tp>
+struct __cccl_underlying_type_impl<_Tp, false>
+{};
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT underlying_type : __cccl_underlying_type_impl<_Tp>
+{};
+
+template <class _Tp>
+using underlying_type_t _CCCL_NODEBUG_ALIAS = typename underlying_type<_Tp>::type;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_UNDERLYING_TYPE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/void_t.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/void_t.h
new file mode 100644
index 00000000..bc83c944
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__type_traits/void_t.h
@@ -0,0 +1,34 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___TYPE_TRAITS_VOID_T_H
+#define _CUDA_STD___TYPE_TRAITS_VOID_T_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+using void_t = void;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___TYPE_TRAITS_VOID_T_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/auto_cast.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/auto_cast.h
new file mode 100644
index 00000000..5ade10fc
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/auto_cast.h
@@ -0,0 +1,34 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_AUTO_CAST_H
+#define _CUDA_STD___UTILITY_AUTO_CAST_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/decay.h>
+
+#if _CCCL_STD_VER >= 2023 && __cpp_auto_cast >= 202110L && !_CCCL_CUDA_COMPILER(NVCC)
+#  define _LIBCUDACXX_AUTO_CAST(expr) auto(expr)
+#elif _CCCL_STD_VER < 2020 && _CCCL_COMPILER(MSVC)
+#  define _LIBCUDACXX_AUTO_CAST(expr) (::cuda::std::decay_t<decltype((expr))>) (expr)
+#else
+#  define _LIBCUDACXX_AUTO_CAST(expr) static_cast<::cuda::std::decay_t<decltype((expr))>>(expr)
+#endif
+
+#endif // _CUDA_STD___UTILITY_AUTO_CAST_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/convert_to_integral.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/convert_to_integral.h
new file mode 100644
index 00000000..d3e7514b
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/convert_to_integral.h
@@ -0,0 +1,101 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_CONVERT_TO_INTEGRAL_H
+#define _CUDA_STD___UTILITY_CONVERT_TO_INTEGRAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_enum.h>
+#include <cuda/std/__type_traits/is_floating_point.h>
+#include <cuda/std/__type_traits/underlying_type.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_API constexpr int __convert_to_integral(int __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr unsigned __convert_to_integral(unsigned __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr long __convert_to_integral(long __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr unsigned long __convert_to_integral(unsigned long __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr long long __convert_to_integral(long long __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr unsigned long long __convert_to_integral(unsigned long long __val)
+{
+  return __val;
+}
+
+template <typename _Fp>
+_CCCL_API constexpr enable_if_t<is_floating_point_v<_Fp>, long long> __convert_to_integral(_Fp __val)
+{
+  return __val;
+}
+
+#if _CCCL_HAS_INT128()
+_CCCL_API constexpr __int128_t __convert_to_integral(__int128_t __val)
+{
+  return __val;
+}
+
+_CCCL_API constexpr __uint128_t __convert_to_integral(__uint128_t __val)
+{
+  return __val;
+}
+#endif // _CCCL_HAS_INT128()
+
+template <class _Tp, bool = is_enum_v<_Tp>>
+struct __sfinae_underlying_type
+{
+  using type            = typename underlying_type<_Tp>::type;
+  using __promoted_type = decltype(((type) 1) + 0);
+};
+
+template <class _Tp>
+struct __sfinae_underlying_type<_Tp, false>
+{};
+
+template <class _Tp>
+_CCCL_API constexpr typename __sfinae_underlying_type<_Tp>::__promoted_type __convert_to_integral(_Tp __val)
+{
+  return __val;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_CONVERT_TO_INTEGRAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/declval.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/declval.h
new file mode 100644
index 00000000..92802640
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/declval.h
@@ -0,0 +1,77 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_DECLVAL_H
+#define _CUDA_STD___UTILITY_DECLVAL_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/void_t.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#ifndef _CCCL_DOXYGEN_INVOKED
+// Handy in concept definitions for turning:
+//   requires (T&& t) { ... std::forward<T>(t) ...; }
+// into:
+//   requires (__declfn_t<T> t) { ... t() ...; }
+//
+// NOTE: Unlike std::declval, the return type of __declfn_t is not
+// rvalue-reference-qualified. In most cases this does not matter, except when _Tp is an
+// incomplete class type. In places where that matters, use __declfn_t<_Tp&&> instead of
+// __declfn_t<_Tp>.
+template <class _Tp, bool _Noexcept = true>
+using __declfn_t _CCCL_NODEBUG_ALIAS = _Tp (*)() noexcept(_Noexcept);
+#endif // _CCCL_DOXYGEN_INVOKED
+
+// When variable templates and noexcept function types are available, a faster
+// implementation of declval is available. It compiles approximately 2x faster
+// than the fallback. NOTE: this implementation causes nvcc < 12.4 to ICE and
+// MSVC < 19.39 to miscompile so we use the fallback instead. The use of the
+// `__declfn_t` alias is help MSVC parse the declaration correctly.
+#if !_CCCL_CUDA_COMPILER(NVCC, <, 12, 4) && !_CCCL_COMPILER(MSVC, <, 19, 39) && !_CCCL_TILE_COMPILATION()
+
+template <class _Tp, class = void>
+extern __declfn_t<void> declval;
+
+template <class _Tp>
+extern __declfn_t<_Tp&&> declval<_Tp, void_t<_Tp&&>>;
+
+#else // ^^^ fast declval ^^^ / vvv default impl vvv
+
+// Suppress deprecation notice for volatile-qualified return type resulting
+// from volatile-qualified types _Tp.
+_CCCL_SUPPRESS_DEPRECATED_PUSH
+_CCCL_SUPPRESS_DEPRECATED_NVRTC_DIAG
+template <class _Tp>
+_CCCL_API inline _Tp&& __declval(int);
+template <class _Tp>
+_CCCL_API inline _Tp __declval(long);
+_CCCL_SUPPRESS_DEPRECATED_POP
+
+template <class _Tp>
+_CCCL_API inline decltype(::cuda::std::__declval<_Tp>(0)) declval() noexcept;
+
+#endif // default impl
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_DECLVAL_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/exchange.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/exchange.h
new file mode 100644
index 00000000..806543b2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/exchange.h
@@ -0,0 +1,46 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_EXCHANGE_H
+#define _CUDA_STD___UTILITY_EXCHANGE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2 = _T1>
+_CCCL_API constexpr _T1 exchange(_T1& __obj, _T2&& __new_value) noexcept(
+  is_nothrow_move_constructible_v<_T1> && is_nothrow_assignable_v<_T1&, _T2>)
+{
+  _T1 __old_value = ::cuda::std::move(__obj);
+  __obj           = ::cuda::std::forward<_T2>(__new_value);
+  return __old_value;
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_EXCHANGE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/forward.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/forward.h
new file mode 100644
index 00000000..1b91ca4e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/forward.h
@@ -0,0 +1,90 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_FORWARD_H
+#define _CUDA_STD___UTILITY_FORWARD_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/is_reference.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+#include <cuda/std/cstddef>
+
+#if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 12) \
+  || (_CCCL_COMPILER(NVRTC) && defined(__NV_BUILTIN_MOVE_FORWARD))
+#  define _CCCL_HAS_BUILTIN_STD_FORWARD() 1
+#else // ^^^ has builtin std::forward ^^^ / vvv no builtin std::forward vvv
+#  define _CCCL_HAS_BUILTIN_STD_FORWARD() 0
+#endif // ^^^ no builtin std::forward ^^^
+
+// nvcc always supports std::forward in device code.
+#if _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+#  undef _CCCL_HAS_BUILTIN_STD_FORWARD
+#  define _CCCL_HAS_BUILTIN_STD_FORWARD() 1
+#endif // _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+
+// We cannot use host features if we are building in freestanding
+// However, NVRTC handles it specially
+#if _CCCL_FREESTANDING() && !_CCCL_COMPILER(NVRTC)
+#  undef _CCCL_HAS_BUILTIN_STD_FORWARD
+#  define _CCCL_HAS_BUILTIN_STD_FORWARD() 0
+#endif // _CCCL_ENABLE_FREESTANDING
+
+// include minimal std:: headers, nvcc in device mode doesn't need the std:: header
+#if _CCCL_HAS_BUILTIN_STD_FORWARD() && !(_CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION())
+#  if _CCCL_HOST_STD_LIB(LIBSTDCXX) && __has_include(<bits/move.h>)
+#    include <bits/move.h>
+#  elif _CCCL_HOST_STD_LIB(LIBCXX) && __has_include(<__utility/forward.h>)
+#    include <__utility/forward.h>
+#  elif _CCCL_HOSTED()
+#    include <utility>
+#  endif // _CCCL_HOSTED()
+#endif // _CCCL_HAS_BUILTIN_STD_FORWARD() && !(_CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION())
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+#if _CCCL_HAS_BUILTIN_STD_FORWARD()
+
+// The compiler treats ::std::forward as a builtin function so it does not need to be instantiated and will be compiled
+// away even at -O0.
+using ::std::forward;
+
+#else // ^^^ _CCCL_HAS_BUILTIN_STD_FORWARD() ^^^ / vvv !_CCCL_HAS_BUILTIN_STD_FORWARD() vvv
+
+template <class _Tp>
+[[nodiscard]] _CCCL_INTRINSIC _CCCL_API constexpr _Tp&& forward(remove_reference_t<_Tp>& __t) noexcept
+{
+  return static_cast<_Tp&&>(__t);
+}
+
+template <class _Tp>
+[[nodiscard]] _CCCL_INTRINSIC _CCCL_API constexpr _Tp&& forward(remove_reference_t<_Tp>&& __t) noexcept
+{
+  static_assert(!is_lvalue_reference_v<_Tp>, "cannot forward an rvalue as an lvalue");
+  return static_cast<_Tp&&>(__t);
+}
+
+#endif // ^^^ !_CCCL_HAS_BUILTIN_STD_FORWARD() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_FORWARD_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/integer_sequence.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/integer_sequence.h
new file mode 100644
index 00000000..af4db54e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/integer_sequence.h
@@ -0,0 +1,308 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_INTEGER_SEQUENCE_H
+#define _CUDA_STD___UTILITY_INTEGER_SEQUENCE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_integral.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <size_t...>
+struct __tuple_indices
+{};
+
+template <class _IdxType, _IdxType... _Values>
+struct __integer_sequence
+{
+  template <template <class _OIdxType, _OIdxType...> class _ToIndexSeq, class _ToIndexType>
+  using __convert _CCCL_NODEBUG_ALIAS = _ToIndexSeq<_ToIndexType, _Values...>;
+
+  template <size_t _Sp>
+  using __to_tuple_indices _CCCL_NODEBUG_ALIAS = __tuple_indices<(_Values + _Sp)...>;
+};
+
+#if defined(_CCCL_BUILTIN_MAKE_INTEGER_SEQ)
+
+template <size_t _Ep, size_t _Sp>
+using __make_indices_imp _CCCL_NODEBUG_ALIAS =
+  typename _CCCL_BUILTIN_MAKE_INTEGER_SEQ(__integer_sequence, size_t, _Ep - _Sp)::template __to_tuple_indices<_Sp>;
+
+#elif defined(_CCCL_BUILTIN_INTEGER_PACK)
+
+template <size_t _Ep, size_t _Sp>
+using __make_indices_imp _CCCL_NODEBUG_ALIAS =
+  typename __integer_sequence<size_t, _CCCL_BUILTIN_INTEGER_PACK(_Ep - _Sp)...>::template __to_tuple_indices<_Sp>;
+
+#else // ^^^ _CCCL_BUILTIN_INTEGER_PACK ^^^ / vvv !_CCCL_BUILTIN_INTEGER_PACK vvv
+
+namespace __detail
+{
+template <typename _Tp, size_t... _Extra>
+struct __repeat;
+template <typename _Tp, _Tp... _Np, size_t... _Extra>
+struct __repeat<__integer_sequence<_Tp, _Np...>, _Extra...>
+{
+  using type _CCCL_NODEBUG_ALIAS = __integer_sequence<
+    _Tp,
+    _Np...,
+    sizeof...(_Np) + _Np...,
+    2 * sizeof...(_Np) + _Np...,
+    3 * sizeof...(_Np) + _Np...,
+    4 * sizeof...(_Np) + _Np...,
+    5 * sizeof...(_Np) + _Np...,
+    6 * sizeof...(_Np) + _Np...,
+    7 * sizeof...(_Np) + _Np...,
+    _Extra...>;
+};
+
+template <size_t _Np>
+struct __parity;
+
+template <size_t _Np>
+struct __make : __parity<_Np % 8>::template __pmake<_Np>
+{};
+
+template <>
+struct __make<0>
+{
+  using type = __integer_sequence<size_t>;
+};
+template <>
+struct __make<1>
+{
+  using type = __integer_sequence<size_t, 0>;
+};
+template <>
+struct __make<2>
+{
+  using type = __integer_sequence<size_t, 0, 1>;
+};
+template <>
+struct __make<3>
+{
+  using type = __integer_sequence<size_t, 0, 1, 2>;
+};
+template <>
+struct __make<4>
+{
+  using type = __integer_sequence<size_t, 0, 1, 2, 3>;
+};
+template <>
+struct __make<5>
+{
+  using type = __integer_sequence<size_t, 0, 1, 2, 3, 4>;
+};
+template <>
+struct __make<6>
+{
+  using type = __integer_sequence<size_t, 0, 1, 2, 3, 4, 5>;
+};
+template <>
+struct __make<7>
+{
+  using type = __integer_sequence<size_t, 0, 1, 2, 3, 4, 5, 6>;
+};
+
+template <>
+struct __parity<0>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type>
+  {};
+};
+template <>
+struct __parity<1>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 1>
+  {};
+};
+template <>
+struct __parity<2>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 2, _Np - 1>
+  {};
+};
+template <>
+struct __parity<3>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 3, _Np - 2, _Np - 1>
+  {};
+};
+template <>
+struct __parity<4>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 4, _Np - 3, _Np - 2, _Np - 1>
+  {};
+};
+template <>
+struct __parity<5>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 5, _Np - 4, _Np - 3, _Np - 2, _Np - 1>
+  {};
+};
+template <>
+struct __parity<6>
+{
+  template <size_t _Np>
+  struct __pmake : __repeat<typename __make<_Np / 8>::type, _Np - 6, _Np - 5, _Np - 4, _Np - 3, _Np - 2, _Np - 1>
+  {};
+};
+template <>
+struct __parity<7>
+{
+  template <size_t _Np>
+  struct __pmake
+      : __repeat<typename __make<_Np / 8>::type, _Np - 7, _Np - 6, _Np - 5, _Np - 4, _Np - 3, _Np - 2, _Np - 1>
+  {};
+};
+} // namespace __detail
+
+template <size_t _Ep, size_t _Sp>
+using __make_indices_imp _CCCL_NODEBUG_ALIAS =
+  typename __detail::__make<_Ep - _Sp>::type::template __to_tuple_indices<_Sp>;
+
+#endif // !_CCCL_BUILTIN_INTEGER_PACK
+
+template <class _Tp, _Tp... _Ip>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT integer_sequence
+{
+  using value_type = _Tp;
+  static_assert(is_integral_v<_Tp>, "std::integer_sequence can only be instantiated with an integral type");
+  static _CCCL_API constexpr size_t size() noexcept
+  {
+    return sizeof...(_Ip);
+  }
+};
+
+template <size_t... _Ip>
+using index_sequence = integer_sequence<size_t, _Ip...>;
+
+#if defined(_CCCL_BUILTIN_MAKE_INTEGER_SEQ)
+
+template <class _Tp, _Tp _Ep>
+using __make_integer_sequence _CCCL_NODEBUG_ALIAS = _CCCL_BUILTIN_MAKE_INTEGER_SEQ(integer_sequence, _Tp, _Ep);
+
+#elif defined(_CCCL_BUILTIN_INTEGER_PACK)
+
+template <class _Tp, _Tp _Ep>
+using __make_integer_sequence _CCCL_NODEBUG_ALIAS = integer_sequence<_Tp, __integer_pack(_Ep)...>;
+
+#else // ^^^ _CCCL_BUILTIN_INTEGER_PACK ^^^ / vvv !_CCCL_BUILTIN_INTEGER_PACK vvv
+
+template <typename _Tp, _Tp _Np>
+using __make_integer_sequence_unchecked _CCCL_NODEBUG_ALIAS =
+  typename __detail::__make<_Np>::type::template __convert<integer_sequence, _Tp>;
+
+template <class _Tp, _Tp _Ep>
+struct __make_integer_sequence_checked
+{
+  static_assert(is_integral_v<_Tp>, "std::make_integer_sequence can only be instantiated with an integral type");
+  static_assert(0 <= _Ep, "std::make_integer_sequence must have a non-negative sequence length");
+  // Workaround GCC bug by preventing bad installations when 0 <= _Ep
+  // https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68929
+  using type _CCCL_NODEBUG_ALIAS = __make_integer_sequence_unchecked<_Tp, 0 <= _Ep ? _Ep : 0>;
+};
+
+template <class _Tp, _Tp _Ep>
+using __make_integer_sequence _CCCL_NODEBUG_ALIAS = typename __make_integer_sequence_checked<_Tp, _Ep>::type;
+
+#endif // !_CCCL_BUILTIN_INTEGER_PACK
+
+template <class _Tp, _Tp _Np>
+using make_integer_sequence = __make_integer_sequence<_Tp, _Np>;
+
+template <size_t _Np>
+using make_index_sequence = make_integer_sequence<size_t, _Np>;
+
+template <class... _Tp>
+using index_sequence_for = make_index_sequence<sizeof...(_Tp)>;
+
+// specialize cuda::std::tuple_size and cuda::std::tuple_element for cuda::std::integer_sequence
+
+template <class _Tp, _Tp... _Indices>
+struct tuple_size<integer_sequence<_Tp, _Indices...>> : integral_constant<size_t, sizeof...(_Indices)>
+{};
+
+template <size_t _Ip, class _Tp, _Tp... _Indices>
+struct tuple_element<_Ip, integer_sequence<_Tp, _Indices...>>
+{
+  static_assert(_Ip < sizeof...(_Indices),
+                "Index out of bounds in cuda::std::tuple_element<> (cuda::std::integer_sequence)");
+  using type = _Tp;
+};
+
+template <size_t _Ip, class _Tp, _Tp... _Indices>
+struct tuple_element<_Ip, const integer_sequence<_Tp, _Indices...>>
+{
+  static_assert(_Ip < sizeof...(_Indices),
+                "Index out of bounds in cuda::std::tuple_element<> (const cuda::std::integer_sequence)");
+  using type = _Tp;
+};
+
+template <size_t _Ip, class _Tp, _Tp... _Indices>
+[[nodiscard]] _CCCL_API constexpr _Tp get(integer_sequence<_Tp, _Indices...>) noexcept
+{
+  static_assert(_Ip < sizeof...(_Indices), "Index out of bounds in cuda::std::get<> (cuda::std::integer_sequence)");
+  constexpr _Tp __indices[]{_Indices...};
+  return __indices[_Ip];
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// tuple protocol for cuda::std::integer_sequence
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _Tp, _Tp... _Indices>
+struct tuple_size<::cuda::std::integer_sequence<_Tp, _Indices...>>
+    : ::cuda::std::integral_constant<::cuda::std::size_t, sizeof...(_Indices)>
+{};
+
+template <::cuda::std::size_t _Ip, class _Tp, _Tp... _Indices>
+struct tuple_element<_Ip, ::cuda::std::integer_sequence<_Tp, _Indices...>>
+{
+  static_assert(_Ip < sizeof...(_Indices), "Index out of bounds in std::tuple_element<> (cuda::std::integer_sequence)");
+  using type = _Tp;
+};
+
+template <::cuda::std::size_t _Ip, class _Tp, _Tp... _Indices>
+struct tuple_element<_Ip, const ::cuda::std::integer_sequence<_Tp, _Indices...>>
+{
+  static_assert(_Ip < sizeof...(_Indices),
+                "Index out of bounds in std::tuple_element<> (const cuda::std::integer_sequence)");
+  using type = _Tp;
+};
+
+_CCCL_END_NAMESPACE_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_INTEGER_SEQUENCE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/move.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/move.h
new file mode 100644
index 00000000..defff1bf
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/move.h
@@ -0,0 +1,140 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_MOVE_H
+#define _CUDA_STD___UTILITY_MOVE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/is_copy_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/remove_reference.h>
+
+#if _CCCL_COMPILER(CLANG, >=, 15) || _CCCL_COMPILER(GCC, >=, 12) \
+  || (_CCCL_COMPILER(NVRTC) && defined(__NV_BUILTIN_MOVE_FORWARD))
+#  define _CCCL_HAS_BUILTIN_STD_MOVE() 1
+#else // ^^^ has builtin std::move ^^^ / vvv no builtin std::move vvv
+#  define _CCCL_HAS_BUILTIN_STD_MOVE() 0
+#endif // ^^^ no builtin std::move ^^^
+
+// nvcc always supports std::move in device code.
+#if _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+#  undef _CCCL_HAS_BUILTIN_STD_MOVE
+#  define _CCCL_HAS_BUILTIN_STD_MOVE() 1
+#endif // _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+
+// We cannot use host features if we are building in freestanding
+// However, NVRTC handles it specially
+#if _CCCL_FREESTANDING() && !_CCCL_COMPILER(NVRTC)
+#  undef _CCCL_HAS_BUILTIN_STD_MOVE
+#  define _CCCL_HAS_BUILTIN_STD_MOVE() 0
+#endif // _CCCL_ENABLE_FREESTANDING
+
+#if _CCCL_COMPILER(CLANG, >=, 15)
+#  define _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() 1
+#else // ^^^ has builtin std::move_if_noexcept ^^^ / vvv no builtin std::move_if_noexcept vvv
+#  define _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() 0
+#endif // ^^^ no builtin std::move_if_noexcept ^^^
+
+// nvcc warns about host only std::move_if_noexcept being used in device code
+#if _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+#  undef _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT
+#  define _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() 0
+#endif // _CCCL_CUDA_COMPILER(NVCC) && _CCCL_DEVICE_COMPILATION()
+
+// We cannot use host features if we are building in freestanding
+#if _CCCL_FREESTANDING()
+#  undef _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT
+#  define _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() 0
+#endif // _CCCL_FREESTANDING()
+
+// include minimal std:: headers, nvcc in device mode doesn't need the std:: header
+#if _CCCL_HAS_BUILTIN_STD_MOVE() || _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT()
+#  if _CCCL_HOST_STD_LIB(LIBSTDCXX) && __has_include(<bits/move.h>)
+#    include <bits/move.h>
+#  elif _CCCL_HOST_STD_LIB(LIBCXX) && __has_include(<__utility/move.h>)
+#    include <__utility/move.h> // includes std::move_if_noexcept, too
+#  elif _CCCL_HOSTED()
+#    include <utility>
+#  endif // _CCCL_HOSTED()
+#endif // _CCCL_HAS_BUILTIN_STD_MOVE() || _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT()
+
+// _CCCL_MOVE macro always expands to std::move builtin, if available, and fallbacks to cuda::std::move otherwise.
+#if _CCCL_HAS_BUILTIN_STD_MOVE()
+#  define _CCCL_MOVE(...) ::std::move(__VA_ARGS__)
+#else // ^^^ _CCCL_HAS_BUILTIN_STD_MOVE() ^^^ / vvv !_CCCL_HAS_BUILTIN_STD_MOVE() vvv
+#  define _CCCL_MOVE(...) ::cuda::std::move(__VA_ARGS__)
+#endif // ^^^ !_CCCL_HAS_BUILTIN_STD_MOVE() ^^^
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// We cannot simply use ::std::move here because it would drag in the move algorithm. This can work only if either/
+// compiling with nvrtc or decorating the move algorithm with `requires true` which overrides the imported move
+// algorithm. But this solution requires concepts.
+#if _CCCL_HAS_BUILTIN_STD_MOVE() && (_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVRTC))
+
+// Forward declare move algorithm decorated with `requires true` to override the imported move algorithm.
+#  if _CCCL_HAS_CONCEPTS()
+template <class _InputIterator, class _OutputIterator>
+  requires true
+_CCCL_API constexpr _OutputIterator move(_InputIterator __first, _InputIterator __last, _OutputIterator __result);
+#  endif // _CCCL_HAS_CONCEPTS()
+
+// The compiler treats ::std::move as a builtin function so it does not need to be instantiated and will be compiled
+// away even at -O0.
+using ::std::move;
+
+#else // ^^^ _CCCL_HAS_BUILTIN_STD_MOVE() && (_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVRTC)) ^^^ /
+      // vvv !_CCCL_HAS_BUILTIN_STD_MOVE() || !(_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVRTC)) vvv
+
+template <class _Tp>
+[[nodiscard]] _CCCL_INTRINSIC _CCCL_API constexpr remove_reference_t<_Tp>&& move(_Tp&& __t) noexcept
+{
+  return static_cast<remove_reference_t<_Tp>&&>(__t);
+}
+
+#endif // ^^^ !_CCCL_HAS_BUILTIN_STD_MOVE() || !(_CCCL_HAS_CONCEPTS() || _CCCL_COMPILER(NVRTC)) ^^^
+
+#if _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT()
+
+// The compiler treats ::std::move_if_noexcept as a builtin function so it does not need to be
+// instantiated and will be compiled away even at -O0.
+using ::std::move_if_noexcept;
+
+#else // ^^^ _CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() ^^^ / vvv !_CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() vvv
+
+template <class _Tp>
+using __move_if_noexcept_result_t =
+  conditional_t<!is_nothrow_move_constructible_v<_Tp> && is_copy_constructible_v<_Tp>, const _Tp&, _Tp&&>;
+
+template <class _Tp>
+[[nodiscard]] _CCCL_INTRINSIC _CCCL_API constexpr __move_if_noexcept_result_t<_Tp> move_if_noexcept(_Tp& __x) noexcept
+{
+  return ::cuda::std::move(__x);
+}
+
+#endif // ^^^ !_CCCL_HAS_BUILTIN_STD_MOVE_IF_NOEXCEPT() ^^^
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_MOVE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/pair.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/pair.h
new file mode 100644
index 00000000..5dd2138d
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/pair.h
@@ -0,0 +1,943 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_PAIR_H
+#define _CUDA_STD___UTILITY_PAIR_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+#  include <cuda/std/__compare/common_comparison_category.h>
+#  include <cuda/std/__compare/synth_three_way.h>
+#endif // _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/__functional/unwrap_ref.h>
+#include <cuda/std/__fwd/get.h>
+#include <cuda/std/__fwd/pair.h>
+#include <cuda/std/__fwd/subrange.h>
+#include <cuda/std/__fwd/tuple.h>
+#include <cuda/std/__tuple_dir/sfinae_helpers.h>
+#include <cuda/std/__tuple_dir/tuple_constraints.h>
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_indices.h>
+#include <cuda/std/__tuple_dir/tuple_like.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+#include <cuda/std/__type_traits/common_reference.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/decay.h>
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/integral_constant.h>
+#include <cuda/std/__type_traits/is_assignable.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/is_convertible.h>
+#include <cuda/std/__type_traits/is_copy_assignable.h>
+#include <cuda/std/__type_traits/is_default_constructible.h>
+#include <cuda/std/__type_traits/is_implicitly_default_constructible.h>
+#include <cuda/std/__type_traits/is_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_copy_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_default_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_swappable.h>
+#include <cuda/std/__type_traits/make_const_lvalue_ref.h>
+#include <cuda/std/__type_traits/reference_constructs_from_temporary.h>
+#include <cuda/std/__type_traits/sfinae_traits.h>
+#include <cuda/std/__utility/forward.h>
+#include <cuda/std/__utility/move.h>
+#include <cuda/std/__utility/piecewise_construct.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+// On Windows we are getting a warning when compiling the const assignment operators with a reference type
+_CCCL_BEGIN_NV_DIAG_SUPPRESS(1770) // type qualifiers are ignored (underlying type is a reference)
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _UPair, class _T1, class _T2>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_constructor __pair_select_pair_like_constructible() noexcept
+{
+  using ::cuda::std::get;
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (__is_cuda_std_ranges_subrange_v<remove_cvref_t<_UPair>>)
+  { // [pairs#pair]-15.1: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!__pair_like<_UPair>)
+  {
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!is_constructible_v<_T1, decltype(get<0>(::cuda::std::declval<_UPair>()))>)
+  { // [pairs#pair]-15.2: is_constructible<T1, decltype(get<0>(std::forward<UTuple>(u)))>... is true
+    return __select_constructor::__invalid;
+  }
+  else if constexpr (!is_constructible_v<_T2, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+  { // [pairs#pair]-15.3: is_constructible<T2, decltype(get<1>(std::forward<UTuple>(u)))>... is true
+    return __select_constructor::__invalid;
+  }
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  else if constexpr (reference_constructs_from_temporary_v<_T1, decltype(get<0>(::cuda::std::declval<_UPair>()))>
+                     || reference_constructs_from_temporary_v<_T2, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+  { // [pairs#pair]-17: This constructor is defined as deleted if
+    // reference_constructs_from_temporary_v<T1, decltype(get<0>(FWD(u)))> or
+    // reference_constructs_from_temporary_v<T2, decltype(get<1>(FWD(u)))> is true
+    return __select_constructor::__deleted;
+  }
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  else if constexpr (is_convertible_v<decltype(get<0>(::cuda::std::declval<_UPair>())), _T1>
+                     && is_convertible_v<decltype(get<1>(::cuda::std::declval<_UPair>())), _T2>)
+  { // [pairs#pair]-17 !is_convertible_v<decltype(get<0>(FWD(u))), T1> &&
+    //                 !is_convertible_v<decltype(get<1>(FWD(u))), T2>
+    return __select_constructor::__implicit;
+  }
+  else
+  {
+    return __select_constructor::__explicit;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <class _UPair, class _T1, class _T2>
+inline constexpr __select_constructor __pair_select_pair_like_constructible_v =
+  ::cuda::std::__pair_select_pair_like_constructible<_UPair, _T1, _T2>();
+
+_CCCL_EXEC_CHECK_DISABLE
+template <bool _IsConst, class _UPair, class _T1, class _T2>
+[[nodiscard]] _CCCL_API _CCCL_CONSTEVAL __select_assignment __pair_select_pair_like_assignable() noexcept
+{
+  using ::cuda::std::get;
+  // NOLINTBEGIN(bugprone-branch-clone)
+  if constexpr (is_same_v<remove_cvref_t<_UPair>, pair<_T1, _T2>>)
+  { // [pairs.pair]-42.1: different-from<UPair, pair>
+    // [pairs.pair]-45.1: different-from<UPair, pair>
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (__is_cuda_std_ranges_subrange_v<remove_cvref_t<_UPair>>)
+  { // [pairs.pair]-42.2: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    // [pairs.pair]-45.2: remove_cvref_t<UTuple> is not a specialization of ranges::subrange,
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (!__pair_like<_UPair>)
+  { // [pairs.pair]-42: pair-like P
+    // [pairs.pair]-45: pair-like P
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (_IsConst)
+  {
+    if constexpr (!is_assignable_v<const _T1&, decltype(get<0>(::cuda::std::declval<_UPair>()))>)
+    { // [pairs.pair]-45.3: is_assignable_v<const T1&, decltype(get<0>(std::forward<P>(p)))> is true
+      return __select_assignment::__invalid;
+    }
+    else if constexpr (!is_assignable_v<const _T2&, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+    { // [pairs.pair]-45.4: is_assignable_v<const T2&, decltype(get<1>(std::forward<P>(p)))> is true
+      return __select_assignment::__invalid;
+    }
+    else if constexpr (is_nothrow_assignable_v<const _T1&, decltype(get<0>(::cuda::std::declval<_UPair>()))>
+                       && is_nothrow_assignable_v<const _T2&, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+    {
+      return __select_assignment::__is_nothrow;
+    }
+    else
+    {
+      return __select_assignment::__may_throw;
+    }
+  }
+  else if constexpr (!is_assignable_v<_T1&, decltype(get<0>(::cuda::std::declval<_UPair>()))>)
+  { // [pairs.pair]-42.3: is_assignable_v<const T1&, decltype(get<0>(std::forward<P>(p)))> is true
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (!is_assignable_v<_T2&, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+  { // [pairs.pair]-42.4: is_assignable_v<const T2&, decltype(get<1>(std::forward<P>(p)))> is true
+    return __select_assignment::__invalid;
+  }
+  else if constexpr (is_nothrow_assignable_v<_T1&, decltype(get<0>(::cuda::std::declval<_UPair>()))>
+                     && is_nothrow_assignable_v<_T2&, decltype(get<1>(::cuda::std::declval<_UPair>()))>)
+  {
+    return __select_assignment::__is_nothrow;
+  }
+  else
+  {
+    return __select_assignment::__may_throw;
+  }
+  // NOLINTEND(bugprone-branch-clone)
+}
+
+template <bool _IsConst, class _UPair, class _T1, class _T2>
+inline constexpr __select_assignment __pair_select_pair_like_assignable_v =
+  ::cuda::std::__pair_select_pair_like_assignable<_IsConst, _UPair, _T1, _T2>();
+
+// base class to ensure `is_trivially_copyable` when possible
+template <class _T1, class _T2, bool = __must_synthesize_assignment_v<_T1> || __must_synthesize_assignment_v<_T2>>
+struct __pair_base
+{
+  _T1 first;
+  _T2 second;
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr __pair_base() noexcept(is_nothrow_default_constructible_v<_T1>
+                                             && is_nothrow_default_constructible_v<_T2>)
+      : first()
+      , second()
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr __pair_base() noexcept(
+    is_nothrow_default_constructible_v<_T1> && is_nothrow_default_constructible_v<_T2>)
+      : first()
+      , second()
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _U1, class _U2>
+  _CCCL_API constexpr __pair_base(_U1&& __t1, _U2&& __t2) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : first(::cuda::std::forward<_U1>(__t1))
+      , second(::cuda::std::forward<_U2>(__t2))
+  {}
+
+protected:
+  template <class... _Args1, class... _Args2, size_t... _I1, size_t... _I2>
+  _CCCL_API constexpr __pair_base(
+    piecewise_construct_t,
+    tuple<_Args1...>& __first_args,
+    tuple<_Args2...>& __second_args,
+    __tuple_indices<_I1...>,
+    __tuple_indices<_I2...>);
+};
+
+template <class _T1, class _T2>
+struct __pair_base<_T1, _T2, true>
+{
+  _T1 first;
+  _T2 second;
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr __pair_base() noexcept(is_nothrow_default_constructible_v<_T1>
+                                             && is_nothrow_default_constructible_v<_T2>)
+      : first()
+      , second()
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr __pair_base() noexcept(
+    is_nothrow_default_constructible_v<_T1> && is_nothrow_default_constructible_v<_T2>)
+      : first()
+      , second()
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_HIDE_FROM_ABI constexpr __pair_base(const __pair_base&) = default;
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_HIDE_FROM_ABI constexpr __pair_base(__pair_base&&) = default;
+
+  // We need to ensure that a reference type, which would inhibit the implicit copy assignment still works
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr __pair_base&
+  operator=(const conditional_t<is_copy_assignable_v<_T1> && is_copy_assignable_v<_T2>, __pair_base, __nat>&
+              __p) noexcept(is_nothrow_copy_assignable_v<_T1> && is_nothrow_copy_assignable_v<_T2>)
+  {
+    first  = __p.first;
+    second = __p.second;
+    return *this;
+  }
+
+  // We need to ensure that a reference type, which would inhibit the implicit move assignment still works
+  _CCCL_EXEC_CHECK_DISABLE
+  _CCCL_API constexpr __pair_base&
+  operator=(conditional_t<is_move_assignable_v<_T1> && is_move_assignable_v<_T2>, __pair_base, __nat>&& __p) noexcept(
+    is_nothrow_move_assignable_v<_T1> && is_nothrow_move_assignable_v<_T2>)
+  {
+    first  = ::cuda::std::move(__p.first);
+    second = ::cuda::std::move(__p.second);
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _U1, class _U2>
+  _CCCL_API constexpr __pair_base(_U1&& __t1, _U2&& __t2) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : first(::cuda::std::forward<_U1>(__t1))
+      , second(::cuda::std::forward<_U2>(__t2))
+  {}
+
+protected:
+  template <class... _Args1, class... _Args2, size_t... _I1, size_t... _I2>
+  _CCCL_API constexpr __pair_base(
+    piecewise_construct_t,
+    tuple<_Args1...>& __first_args,
+    tuple<_Args2...>& __second_args,
+    __tuple_indices<_I1...>,
+    __tuple_indices<_I2...>);
+};
+
+template <class _T1, class _T2>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT pair : public __pair_base<_T1, _T2>
+{
+  using __base = __pair_base<_T1, _T2>;
+
+  using first_type  = _T1;
+  using second_type = _T2;
+
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair() noexcept(is_nothrow_default_constructible_v<_T1> && is_nothrow_default_constructible_v<_T2>)
+      : __base()
+  {}
+
+  template <__select_constructor _Trait = ::cuda::std::__tuple_select_default_constructible(__tuple_types<_T1, _T2>{}),
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair() noexcept(is_nothrow_default_constructible_v<_T1>
+                                               && is_nothrow_default_constructible_v<_T2>)
+      : __base()
+  {}
+
+  // copy and move constructors
+  _CCCL_HIDE_FROM_ABI pair(const pair&) = default;
+  _CCCL_HIDE_FROM_ABI pair(pair&&)      = default;
+
+  // element wise constructors
+  template <__select_constructor _Trait = __tuple_select_variadic_copy_constructible_v<__tuple_types<_T1, _T2>>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(const _T1& __t1, const _T2& __t2) noexcept(
+    is_nothrow_copy_constructible_v<_T1> && is_nothrow_copy_constructible_v<_T2>)
+      : __base(__t1, __t2)
+  {}
+  template <__select_constructor _Trait = __tuple_select_variadic_copy_constructible_v<__tuple_types<_T1, _T2>>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(const _T1& __t1, const _T2& __t2) noexcept(
+    is_nothrow_copy_constructible_v<_T1> && is_nothrow_copy_constructible_v<_T2>)
+      : __base(__t1, __t2)
+  {}
+
+  template <class _U1 = _T1,
+            class _U2 = _T2,
+            __select_constructor _Trait =
+              __tuple_select_variadic_constructible_v<__tuple_types<_T1, _T2>, __tuple_types<_U1, _U2>>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(_U1&& __u1, _U2&& __u2) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : __base(::cuda::std::forward<_U1>(__u1), ::cuda::std::forward<_U2>(__u2))
+  {}
+
+  template <class _U1 = _T1,
+            class _U2 = _T2,
+            __select_constructor _Trait =
+              __tuple_select_variadic_constructible_v<__tuple_types<_T1, _T2>, __tuple_types<_U1, _U2>>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(_U1&& __u1, _U2&& __u2) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : __base(::cuda::std::forward<_U1>(__u1), ::cuda::std::forward<_U2>(__u2))
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _U1 = _T1,
+            class _U2 = _T2,
+            __select_constructor _Trait =
+              __tuple_select_variadic_constructible_v<__tuple_types<_T1, _T2>, __tuple_types<_U1, _U2>>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  constexpr pair(_U1&&, _U2&&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+  // converting constructors
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(pair<_U1, _U2>& __p) noexcept(
+    is_nothrow_constructible_v<_T1, _U1&> && is_nothrow_constructible_v<_T2, _U2&>)
+      : __base(__p.first, __p.second)
+  {}
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(pair<_U1, _U2>& __p) noexcept(
+    is_nothrow_constructible_v<_T1, _U1&> && is_nothrow_constructible_v<_T2, _U2&>)
+      : __base(__p.first, __p.second)
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait            = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  constexpr pair(pair<_U1, _U2>&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(const pair<_U1, _U2>& __p) noexcept(
+    is_nothrow_constructible_v<_T1, const _U1&> && is_nothrow_constructible_v<_T2, const _U2&>)
+      : __base(__p.first, __p.second)
+  {}
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(const pair<_U1, _U2>& __p) noexcept(
+    is_nothrow_constructible_v<_T1, const _U1&> && is_nothrow_constructible_v<_T2, const _U2&>)
+      : __base(__p.first, __p.second)
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&, _T1, _T2>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  constexpr pair(const pair<_U1, _U2>&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(pair<_U1, _U2>&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : __base(::cuda::std::move(__p.first), ::cuda::std::move(__p.second))
+  {}
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(pair<_U1, _U2>&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, _U1> && is_nothrow_constructible_v<_T2, _U2>)
+      : __base(::cuda::std::move(__p.first), ::cuda::std::move(__p.second))
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  constexpr pair(pair<_U1, _U2>&&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(const pair<_U1, _U2>&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, const _U1> && is_nothrow_constructible_v<_T2, const _U2>)
+      : __base(::cuda::std::move(__p.first), ::cuda::std::move(__p.second))
+  {}
+
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(const pair<_U1, _U2>&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, const _U1> && is_nothrow_constructible_v<_T2, const _U2>)
+      : __base(::cuda::std::move(__p.first), ::cuda::std::move(__p.second))
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _U1,
+            class _U2,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<const pair<_U1, _U2>&&, _T1, _T2>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  constexpr pair(const pair<_U1, _U2>&&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+
+  // NOLINTBEGIN(bugprone-forwarding-reference-overload)
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _UPair,
+            enable_if_t<!is_same_v<remove_cvref_t<_UPair>, pair>, int> = 0,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<_UPair, _T1, _T2>,
+            enable_if_t<__can_construct_implicitly<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(_UPair&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, decltype(::cuda::std::__adl_get<0>(::cuda::std::forward<_UPair>(__p)))>
+    && is_nothrow_constructible_v<_T2, decltype(::cuda::std::__adl_get<1>(::cuda::std::forward<_UPair>(__p)))>)
+      : __base(
+          // __adl_get() specifically will only move the sub-object, it's therefore OK to
+          // "move" the outer pair twice
+          // NOLINTBEGIN(bugprone-use-after-move)
+          ::cuda::std::__adl_get<0>(::cuda::std::forward<_UPair>(__p)),
+          ::cuda::std::__adl_get<1>(::cuda::std::forward<_UPair>(__p))
+          // NOLINTEND(bugprone-use-after-move)
+        )
+  {}
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _UPair,
+            enable_if_t<!is_same_v<remove_cvref_t<_UPair>, pair>, int> = 0,
+            __select_constructor _Trait = __pair_select_pair_like_constructible_v<_UPair, _T1, _T2>,
+            enable_if_t<__can_construct_explicitly<_Trait>, int> = 0>
+  _CCCL_API explicit constexpr pair(_UPair&& __p) noexcept(
+    is_nothrow_constructible_v<_T1, decltype(::cuda::std::__adl_get<0>(::cuda::std::forward<_UPair>(__p)))>
+    && is_nothrow_constructible_v<_T2, decltype(::cuda::std::__adl_get<1>(::cuda::std::forward<_UPair>(__p)))>)
+      : __base(::cuda::std::__adl_get<0>(::cuda::std::forward<_UPair>(__p)),
+               ::cuda::std::__adl_get<1>(::cuda::std::forward<_UPair>(__p)))
+  {}
+
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY)
+  template <class _UPair,
+            enable_if_t<!is_same_v<remove_cvref_t<_UPair>, pair>, int> = 0,
+            __select_constructor _Trait            = __pair_select_pair_like_constructible_v<_UPair, _T1, _T2>,
+            enable_if_t<__is_deleted<_Trait>, int> = 0>
+  _CCCL_API constexpr pair(_UPair&&) = delete;
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY
+  // NOLINTEND(bugprone-forwarding-reference-overload)
+
+  template <class... _Args1, class... _Args2>
+  _CCCL_API constexpr pair(piecewise_construct_t __pc,
+                           tuple<_Args1...> __first_args,
+                           tuple<_Args2...> __second_args) noexcept((is_nothrow_constructible_v<_T1, _Args1...>
+                                                                     && is_nothrow_constructible_v<_T2, _Args2...>) )
+      : __base(__pc,
+               __first_args,
+               __second_args,
+               __make_tuple_indices_t<sizeof...(_Args1)>(),
+               __make_tuple_indices_t<sizeof...(_Args2)>())
+  {}
+
+  // assignments
+  _CCCL_HIDE_FROM_ABI pair& operator=(const pair&) = default;
+  _CCCL_HIDE_FROM_ABI pair& operator=(pair&&)      = default;
+
+  template <class _U1,
+            class _U2,
+            enable_if_t<is_assignable_v<_T1&, const _U1&>, int> = 0,
+            enable_if_t<is_assignable_v<_T2&, const _U2&>, int> = 0>
+  _CCCL_API constexpr pair& operator=(const pair<_U1, _U2>& __p) noexcept(
+    is_nothrow_assignable_v<_T1&, const _U1&> && is_nothrow_assignable_v<_T2&, const _U2&>)
+  {
+    this->first  = __p.first;
+    this->second = __p.second;
+    return *this;
+  }
+
+  template <class _U1,
+            class _U2,
+            enable_if_t<is_assignable_v<const _T1&, const _U1&>, int> = 0,
+            enable_if_t<is_assignable_v<const _T2&, const _U2&>, int> = 0>
+  _CCCL_API constexpr const pair& operator=(const pair<_U1, _U2>& __p) const
+    noexcept(is_nothrow_assignable_v<const _T1&, const _U1&> && is_nothrow_assignable_v<const _T2&, const _U2&>)
+  {
+    this->first  = __p.first;
+    this->second = __p.second;
+    return *this;
+  }
+
+  template <class _U1,
+            class _U2,
+            enable_if_t<is_assignable_v<_T1&, _U1>, int> = 0,
+            enable_if_t<is_assignable_v<_T2&, _U2>, int> = 0>
+  _CCCL_API constexpr pair&
+  operator=(pair<_U1, _U2>&& __p) noexcept(is_nothrow_assignable_v<_T1&, _U1> && is_nothrow_assignable_v<_T2&, _U2>)
+  {
+    this->first  = ::cuda::std::forward<_U1>(__p.first);
+    this->second = ::cuda::std::forward<_U2>(__p.second);
+    return *this;
+  }
+
+  template <class _U1,
+            class _U2,
+            enable_if_t<is_assignable_v<const _T1&, _U1>, int> = 0,
+            enable_if_t<is_assignable_v<const _T2&, _U2>, int> = 0>
+  _CCCL_API constexpr const pair& operator=(pair<_U1, _U2>&& __p) const
+    noexcept(is_nothrow_assignable_v<const _T1&, _U1> && is_nothrow_assignable_v<const _T2&, _U2>)
+  {
+    this->first  = ::cuda::std::forward<_U1>(__p.first);
+    this->second = ::cuda::std::forward<_U2>(__p.second);
+    return *this;
+  }
+
+  // ``get`` will specifically only move the sub-object, it's therefore OK to
+  // "move" the outer pair twice
+  // NOLINTBEGIN(bugprone-use-after-move)
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _UPair,
+            __select_assignment _Trait = __pair_select_pair_like_assignable_v</*__is_const=*/false, _UPair, _T1, _T2>,
+            enable_if_t<__can_assign<_Trait>, int> = 0>
+  _CCCL_API constexpr pair& operator=(_UPair&& __p) noexcept(__can_nothrow_assign<_Trait>)
+  {
+    using ::cuda::std::get;
+    this->first  = get<0>(::cuda::std::forward<_UPair>(__p));
+    this->second = get<1>(::cuda::std::forward<_UPair>(__p));
+    return *this;
+  }
+
+  _CCCL_EXEC_CHECK_DISABLE
+  template <class _UPair,
+            __select_assignment _Trait = __pair_select_pair_like_assignable_v</*__is_const=*/true, _UPair, _T1, _T2>,
+            enable_if_t<__can_assign<_Trait>, int> = 0>
+  _CCCL_API constexpr const pair& operator=(_UPair&& __p) const noexcept(__can_nothrow_assign<_Trait>)
+  {
+    using ::cuda::std::get;
+    this->first  = get<0>(::cuda::std::forward<_UPair>(__p));
+    this->second = get<1>(::cuda::std::forward<_UPair>(__p));
+    return *this;
+  }
+  // NOLINTEND(bugprone-use-after-move)
+
+  _CCCL_API constexpr void swap(pair& __p) noexcept(is_nothrow_swappable_v<_T1> && is_nothrow_swappable_v<_T2>)
+  {
+    using ::cuda::std::swap;
+    swap(this->first, __p.first);
+    swap(this->second, __p.second);
+  }
+
+#if _CCCL_STD_VER >= 2023
+  _CCCL_API constexpr void swap(const pair& __p) const
+    noexcept(is_nothrow_swappable_v<const _T1> && is_nothrow_swappable_v<const _T2>)
+  {
+    using ::cuda::std::swap;
+    swap(this->first, __p.first);
+    swap(this->second, __p.second);
+  }
+#endif // _CCCL_STD_VER >= 2023
+
+#if _CCCL_HOSTED()
+  _CCCL_HOST_API constexpr operator ::std::pair<_T1, _T2>() const
+  {
+    return {this->first, this->second};
+  }
+#endif // _CCCL_HOSTED()
+};
+
+template <class _T1, class _T2>
+_CCCL_DEDUCTION_GUIDE_ATTRIBUTES pair(_T1, _T2) -> pair<_T1, _T2>;
+
+// [pairs.spec], specialized algorithms
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator==(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return __x.first == __y.first && __x.second == __y.second;
+}
+
+#if _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr common_comparison_category_t<__synth_three_way_result<_T1>, __synth_three_way_result<_T2>>
+operator<=>(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  if (auto __c = ::cuda::std::__synth_three_way(__x.first, __y.first); __c != 0)
+  {
+    return __c;
+  }
+  return ::cuda::std::__synth_three_way(__x.second, __y.second);
+}
+
+#else // ^^^ _LIBCUDACXX_HAS_SPACESHIP_OPERATOR() ^^^ / vvv !_LIBCUDACXX_HAS_SPACESHIP_OPERATOR() vvv
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator!=(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return !(__x == __y);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator<(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return __x.first < __y.first || (!(__y.first < __x.first) && __x.second < __y.second);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator>(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return __y < __x;
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator>=(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return !(__x < __y);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _T1, class _T2>
+_CCCL_API constexpr bool operator<=(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y)
+{
+  return !(__y < __x);
+}
+
+#endif // !_LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+
+#if _CCCL_STD_VER >= 2023
+template <class _T1, class _T2, class _U1, class _U2, template <class> class _TQual, template <class> class _UQual>
+  requires requires {
+    typename pair<common_reference_t<_TQual<_T1>, _UQual<_U1>>, common_reference_t<_TQual<_T2>, _UQual<_U2>>>;
+  }
+struct basic_common_reference<pair<_T1, _T2>, pair<_U1, _U2>, _TQual, _UQual>
+{
+  using type = pair<common_reference_t<_TQual<_T1>, _UQual<_U1>>, common_reference_t<_TQual<_T2>, _UQual<_U2>>>;
+};
+
+template <class _T1, class _T2, class _U1, class _U2>
+  requires requires { typename pair<common_type_t<_T1, _U1>, common_type_t<_T2, _U2>>; }
+struct common_type<pair<_T1, _T2>, pair<_U1, _U2>>
+{
+  using type = pair<common_type_t<_T1, _U1>, common_type_t<_T2, _U2>>;
+};
+#endif // _CCCL_STD_VER >= 2023
+
+_CCCL_TEMPLATE(class _T1, class _T2)
+_CCCL_REQUIRES(is_swappable_v<_T1> _CCCL_AND is_swappable_v<_T2>)
+_CCCL_API constexpr void
+swap(pair<_T1, _T2>& __x, pair<_T1, _T2>& __y) noexcept((is_nothrow_swappable_v<_T1> && is_nothrow_swappable_v<_T2>) )
+{
+  __x.swap(__y);
+}
+
+_CCCL_TEMPLATE(class _T1, class _T2)
+_CCCL_REQUIRES(is_swappable_v<const _T1> _CCCL_AND is_swappable_v<const _T2>)
+_CCCL_API constexpr void swap(const pair<_T1, _T2>& __x, const pair<_T1, _T2>& __y) noexcept(
+  (is_nothrow_swappable_v<const _T1> && is_nothrow_swappable_v<const _T2>) )
+{
+  __x.swap(__y);
+}
+
+template <class _T1, class _T2>
+_CCCL_API constexpr pair<unwrap_ref_decay_t<_T1>, unwrap_ref_decay_t<_T2>> make_pair(_T1&& __t1, _T2&& __t2)
+{
+  return pair<unwrap_ref_decay_t<_T1>, unwrap_ref_decay_t<_T2>>(
+    ::cuda::std::forward<_T1>(__t1), ::cuda::std::forward<_T2>(__t2));
+}
+
+template <size_t _Ip>
+struct __get_pair;
+
+template <>
+struct __get_pair<0>
+{
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr _T1& get(pair<_T1, _T2>& __p) noexcept
+  {
+    return __p.first;
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr const _T1& get(const pair<_T1, _T2>& __p) noexcept
+  {
+    return __p.first;
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr _T1&& get(pair<_T1, _T2>&& __p) noexcept
+  {
+    return ::cuda::std::forward<_T1>(__p.first);
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr const _T1&& get(const pair<_T1, _T2>&& __p) noexcept
+  {
+    return ::cuda::std::forward<const _T1>(__p.first);
+  }
+};
+
+template <>
+struct __get_pair<1>
+{
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr _T2& get(pair<_T1, _T2>& __p) noexcept
+  {
+    return __p.second;
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr const _T2& get(const pair<_T1, _T2>& __p) noexcept
+  {
+    return __p.second;
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr _T2&& get(pair<_T1, _T2>&& __p) noexcept
+  {
+    return ::cuda::std::forward<_T2>(__p.second);
+  }
+
+  template <class _T1, class _T2>
+  [[nodiscard]] _CCCL_API static constexpr const _T2&& get(const pair<_T1, _T2>&& __p) noexcept
+  {
+    return ::cuda::std::forward<const _T2>(__p.second);
+  }
+};
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, pair<_T1, _T2>>& get(pair<_T1, _T2>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<_Ip>::get(__p);
+}
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, pair<_T1, _T2>>& get(const pair<_T1, _T2>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<_Ip>::get(__p);
+}
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr tuple_element_t<_Ip, pair<_T1, _T2>>&& get(pair<_T1, _T2>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<_Ip>::get(::cuda::std::move(__p));
+}
+
+template <size_t _Ip, class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const tuple_element_t<_Ip, pair<_T1, _T2>>&& get(const pair<_T1, _T2>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<_Ip>::get(::cuda::std::move(__p));
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr _T1& get(pair<_T1, _T2>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<0>::get(__p);
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const _T1& get(const pair<_T1, _T2>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<0>::get(__p);
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr _T1&& get(pair<_T1, _T2>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<0>::get(::cuda::std::move(__p));
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const _T1&& get(const pair<_T1, _T2>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<0>::get(::cuda::std::move(__p));
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr _T1& get(pair<_T2, _T1>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<1>::get(__p);
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const _T1& get(const pair<_T2, _T1>& __p) noexcept
+{
+  return ::cuda::std::__get_pair<1>::get(__p);
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr _T1&& get(pair<_T2, _T1>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<1>::get(::cuda::std::move(__p));
+}
+
+template <class _T1, class _T2>
+[[nodiscard]] _CCCL_API constexpr const _T1&& get(const pair<_T2, _T1>&& __p) noexcept
+{
+  return ::cuda::std::__get_pair<1>::get(::cuda::std::move(__p));
+}
+
+// specialize cuda::std::tuple_size and cuda::std::tuple_element for std::pair and cuda::std::pair
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<::std::pair<_Tp, _Up>> : integral_constant<size_t, 2>
+{};
+
+template <size_t _Ip, class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, ::std::pair<_Tp, _Up>>
+{
+  static_assert(_Ip < 2, "Index out of bounds in cuda::std::tuple_element<std::pair<_Tp, _Up>>");
+};
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<0, ::std::pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<1, ::std::pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Up;
+};
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<pair<_Tp, _Up>> : integral_constant<size_t, 2>
+{};
+
+template <size_t _Ip, class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, pair<_Tp, _Up>>
+{
+  static_assert(_Ip < 2, "Index out of bounds in cuda::std::tuple_element<cuda::std::pair<_Tp, _Up>>");
+};
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<0, pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp, class _Up>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<1, pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Up;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// tuple protocol for cuda::std::pair
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _Tp, class _Up>
+struct tuple_size<::cuda::std::pair<_Tp, _Up>> : ::cuda::std::integral_constant<::cuda::std::size_t, 2>
+{};
+
+template <::cuda::std::size_t _Ip, class _Tp, class _Up>
+struct tuple_element<_Ip, ::cuda::std::pair<_Tp, _Up>>
+{
+  static_assert(_Ip < 2, "Index out of bounds in std::tuple_element<cuda::std::pair<_Tp, _Up>>");
+};
+template <class _Tp, class _Up>
+struct tuple_element<0, ::cuda::std::pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+template <class _Tp, class _Up>
+struct tuple_element<1, ::cuda::std::pair<_Tp, _Up>>
+{
+  using type _CCCL_NODEBUG_ALIAS = _Up;
+};
+
+_CCCL_END_NAMESPACE_STD
+
+_CCCL_END_NV_DIAG_SUPPRESS()
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_PAIR_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/piecewise_construct.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/piecewise_construct.h
new file mode 100644
index 00000000..a8d634b8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/piecewise_construct.h
@@ -0,0 +1,37 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_PIECEWISE_CONSTRUCT_H
+#define _CUDA_STD___UTILITY_PIECEWISE_CONSTRUCT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+struct _CCCL_TYPE_VISIBILITY_DEFAULT piecewise_construct_t
+{
+  _CCCL_HIDE_FROM_ABI explicit piecewise_construct_t() = default;
+};
+_CCCL_GLOBAL_CONSTANT piecewise_construct_t piecewise_construct = piecewise_construct_t();
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_PIECEWISE_CONSTRUCT_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/priority_tag.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/priority_tag.h
new file mode 100644
index 00000000..caed8a61
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/priority_tag.h
@@ -0,0 +1,40 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_PRIORITY_TAG_H
+#define _CUDA_STD___UTILITY_PRIORITY_TAG_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <size_t _Ip>
+struct __priority_tag : __priority_tag<_Ip - 1>
+{};
+template <>
+struct __priority_tag<0>
+{};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_PRIORITY_TAG_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/swap.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/swap.h
new file mode 100644
index 00000000..67a21fdd
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/swap.h
@@ -0,0 +1,64 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_SWAP_H
+#define _CUDA_STD___UTILITY_SWAP_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/enable_if.h>
+#include <cuda/std/__type_traits/is_move_assignable.h>
+#include <cuda/std/__type_traits/is_move_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_assignable.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/is_swappable.h>
+#include <cuda/std/__utility/move.h>
+#include <cuda/std/cstddef>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+// we use type_identity_t<_Tp> as second parameter, to avoid ambiguity with std::swap, which will thus be preferred by
+// overload resolution (which is ok since std::swap is only considered when explicitly called, or found by ADL for types
+// from std::)
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp>
+_CCCL_API constexpr __swap_result_t<_Tp> swap(_Tp& __x, type_identity_t<_Tp>& __y) noexcept(
+  is_nothrow_move_constructible_v<_Tp> && is_nothrow_move_assignable_v<_Tp>)
+{
+  _Tp __t(::cuda::std::move(__x));
+  __x = ::cuda::std::move(__y);
+  __y = ::cuda::std::move(__t);
+}
+
+_CCCL_EXEC_CHECK_DISABLE
+template <class _Tp, size_t _Np>
+_CCCL_API constexpr enable_if_t<__detect_adl_swap::__has_no_adl_swap_array<_Tp, _Np>::value && __is_swappable<_Tp>::value>
+swap(_Tp (&__a)[_Np], _Tp (&__b)[_Np]) noexcept(__is_nothrow_swappable<_Tp>::value)
+{
+  for (size_t __i = 0; __i != _Np; ++__i)
+  {
+    swap(__a[__i], __b[__i]);
+  }
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_SWAP_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/to_underlying.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/to_underlying.h
new file mode 100644
index 00000000..bad0df87
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/to_underlying.h
@@ -0,0 +1,40 @@
+// -*- C++ -*-
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_TO_UNDERLYING_H
+#define _CUDA_STD___UTILITY_TO_UNDERLYING_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__type_traits/underlying_type.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp>
+[[nodiscard]] _CCCL_API constexpr underlying_type_t<_Tp> to_underlying(_Tp __val) noexcept
+{
+  return static_cast<underlying_type_t<_Tp>>(__val);
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_TO_UNDERLYING_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/undefined.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/undefined.h
new file mode 100644
index 00000000..05edaec2
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/undefined.h
@@ -0,0 +1,34 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_UNDEFINED_H
+#define _CUDA_STD___UTILITY_UNDEFINED_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class...>
+struct __undefined; // leave this undefined
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_UNDEFINED_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/unreachable.h b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/unreachable.h
new file mode 100644
index 00000000..f69ed410
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/__utility/unreachable.h
@@ -0,0 +1,37 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD___UTILITY_UNREACHABLE_H
+#define _CUDA_STD___UTILITY_UNREACHABLE_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+[[noreturn]] _CCCL_API inline void unreachable()
+{
+  _CCCL_ASSERT(false, "unreachable");
+  _CCCL_UNREACHABLE();
+}
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD___UTILITY_UNREACHABLE_H
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/array b/qwen3_6_scripts/cccl_preload/include/cuda/std/array
new file mode 100644
index 00000000..bccd99f8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/array
@@ -0,0 +1,553 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the CUDA Toolkit, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_ARRAY
+#define _CUDA_STD_ARRAY
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__algorithm/equal.h>
+#include <cuda/std/__algorithm/fill_n.h>
+#include <cuda/std/__algorithm/lexicographical_compare.h>
+#include <cuda/std/__algorithm/swap_ranges.h>
+#include <cuda/std/__concepts/concept_macros.h>
+#include <cuda/std/__exception/exception_macros.h>
+#include <cuda/std/__fwd/array.h>
+#include <cuda/std/__host_stdlib/stdexcept>
+#include <cuda/std/__iterator/reverse_iterator.h>
+#include <cuda/std/__tuple_dir/sfinae_helpers.h>
+#include <cuda/std/__type_traits/conditional.h>
+#include <cuda/std/__type_traits/is_array.h>
+#include <cuda/std/__type_traits/is_const.h>
+#include <cuda/std/__type_traits/is_constructible.h>
+#include <cuda/std/__type_traits/is_move_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_constructible.h>
+#include <cuda/std/__type_traits/is_nothrow_move_constructible.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__type_traits/is_swappable.h>
+#include <cuda/std/__type_traits/remove_cv.h>
+#include <cuda/std/__utility/integer_sequence.h>
+#include <cuda/std/__utility/move.h>
+#include <cuda/std/__utility/unreachable.h>
+#include <cuda/std/cstdint>
+#include <cuda/std/limits>
+
+// standard-mandated includes
+#include <cuda/std/version>
+
+// [iterator.range]
+#include <cuda/std/__iterator/access.h>
+#include <cuda/std/__iterator/data.h>
+#include <cuda/std/__iterator/empty.h>
+#include <cuda/std/__iterator/reverse_access.h>
+#include <cuda/std/__iterator/size.h>
+
+// [array.syn]
+#if _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+#  include <compare>
+#endif // _LIBCUDACXX_HAS_SPACESHIP_OPERATOR()
+#include <cuda/std/initializer_list>
+
+// [tuple.helper]
+#include <cuda/std/__tuple_dir/tuple_element.h>
+#include <cuda/std/__tuple_dir/tuple_size.h>
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+template <class _Tp, size_t _Size>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT array
+{
+  // types:
+  using __self                 = array;
+  using value_type             = _Tp;
+  using reference              = value_type&;
+  using const_reference        = const value_type&;
+  using iterator               = value_type*;
+  using const_iterator         = const value_type*;
+  using pointer                = value_type*;
+  using const_pointer          = const value_type*;
+  using size_type              = size_t;
+  using difference_type        = ptrdiff_t;
+  using reverse_iterator       = ::cuda::std::reverse_iterator<iterator>;
+  using const_reverse_iterator = ::cuda::std::reverse_iterator<const_iterator>;
+
+  _Tp __elems_[_Size];
+
+  // No explicit construct/copy/destroy for aggregate type
+  _CCCL_API constexpr void fill(const value_type& __u)
+  {
+    ::cuda::std::fill_n(__elems_, _Size, __u);
+  }
+
+  _CCCL_API constexpr void swap(array& __a) noexcept(is_nothrow_swappable_v<_Tp>)
+  {
+    ::cuda::std::swap_ranges(__elems_, __elems_ + _Size, __a.data());
+  }
+
+  // iterators:
+  [[nodiscard]] _CCCL_API constexpr iterator begin() noexcept
+  {
+    return iterator(__elems_);
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator begin() const noexcept
+  {
+    return const_iterator(__elems_);
+  }
+  [[nodiscard]] _CCCL_API constexpr iterator end() noexcept
+  {
+    return iterator(__elems_ + _Size);
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator end() const noexcept
+  {
+    return const_iterator(__elems_ + _Size);
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator rbegin() noexcept
+  {
+    return reverse_iterator(end());
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator rbegin() const noexcept
+  {
+    return const_reverse_iterator(end());
+  }
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator rend() noexcept
+  {
+    return reverse_iterator(begin());
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator rend() const noexcept
+  {
+    return const_reverse_iterator(begin());
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_iterator cbegin() const noexcept
+  {
+    return begin();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator cend() const noexcept
+  {
+    return end();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator crbegin() const noexcept
+  {
+    return rbegin();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator crend() const noexcept
+  {
+    return rend();
+  }
+
+  // capacity:
+  [[nodiscard]] _CCCL_API constexpr size_type size() const noexcept
+  {
+    return _Size;
+  }
+  [[nodiscard]] _CCCL_API constexpr size_type max_size() const noexcept
+  {
+    return _Size;
+  }
+  [[nodiscard]] _CCCL_API constexpr bool empty() const noexcept
+  {
+    return _Size == 0;
+  }
+
+  // element access:
+  [[nodiscard]] _CCCL_API constexpr reference operator[](size_type __n) noexcept
+  {
+    _CCCL_ASSERT(__n < _Size, "out-of-bounds access in std::array<T, N>");
+    return __elems_[__n];
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reference operator[](size_type __n) const noexcept
+  {
+    _CCCL_ASSERT(__n < _Size, "out-of-bounds access in std::array<T, N>");
+    return __elems_[__n];
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reference at(size_type __n)
+  {
+    if (__n >= _Size)
+    {
+      _CCCL_THROW(::std::out_of_range, "array::at");
+    }
+    return __elems_[__n];
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_reference at(size_type __n) const
+  {
+    if (__n >= _Size)
+    {
+      _CCCL_THROW(::std::out_of_range, "array::at");
+    }
+    return __elems_[__n];
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reference front() noexcept
+  {
+    return (*this)[0];
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reference front() const noexcept
+  {
+    return (*this)[0];
+  }
+  [[nodiscard]] _CCCL_API constexpr reference back() noexcept
+  {
+    return (*this)[_Size - 1];
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reference back() const noexcept
+  {
+    return (*this)[_Size - 1];
+  }
+
+  [[nodiscard]] _CCCL_API constexpr value_type* data() noexcept
+  {
+    return __elems_;
+  }
+  [[nodiscard]] _CCCL_API constexpr const value_type* data() const noexcept
+  {
+    return __elems_;
+  }
+};
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_MSVC(4702) // Unreachable code
+
+template <class _Tp>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT array<_Tp, 0>
+{
+  // types:
+  using __self                 = array;
+  using value_type             = _Tp;
+  using reference              = value_type&;
+  using const_reference        = const value_type&;
+  using iterator               = value_type*;
+  using const_iterator         = const value_type*;
+  using pointer                = value_type*;
+  using const_pointer          = const value_type*;
+  using size_type              = size_t;
+  using difference_type        = ptrdiff_t;
+  using reverse_iterator       = ::cuda::std::reverse_iterator<iterator>;
+  using const_reverse_iterator = ::cuda::std::reverse_iterator<const_iterator>;
+
+  using _CharType = conditional_t<is_const_v<_Tp>, const char, char>;
+
+  struct _ArrayInStructT
+  {
+    _Tp __data_[1];
+  };
+  _CCCL_ALIGNAS_TYPE(_ArrayInStructT) _CharType __elems_[sizeof(_ArrayInStructT)];
+
+  [[nodiscard]] _CCCL_API constexpr value_type* data() noexcept
+  {
+    return nullptr;
+  }
+  [[nodiscard]] _CCCL_API constexpr const value_type* data() const noexcept
+  {
+    return nullptr;
+  }
+
+  // No explicit construct/copy/destroy for aggregate type
+  _CCCL_API constexpr void fill(const value_type&)
+  {
+    static_assert(!is_const_v<_Tp>, "cannot fill zero-sized array of type 'const T'");
+  }
+
+  _CCCL_API constexpr void swap(array&) noexcept
+  {
+    static_assert(!is_const_v<_Tp>, "cannot swap zero-sized array of type 'const T'");
+  }
+
+  // iterators:
+  [[nodiscard]] _CCCL_API constexpr iterator begin() noexcept
+  {
+    return iterator(nullptr);
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator begin() const noexcept
+  {
+    return const_iterator(nullptr);
+  }
+  [[nodiscard]] _CCCL_API constexpr iterator end() noexcept
+  {
+    return iterator(nullptr);
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator end() const noexcept
+  {
+    return const_iterator(nullptr);
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator rbegin() noexcept
+  {
+    return reverse_iterator(end());
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator rbegin() const noexcept
+  {
+    return const_reverse_iterator(end());
+  }
+  [[nodiscard]] _CCCL_API constexpr reverse_iterator rend() noexcept
+  {
+    return reverse_iterator(begin());
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator rend() const noexcept
+  {
+    return const_reverse_iterator(begin());
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_iterator cbegin() const noexcept
+  {
+    return begin();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_iterator cend() const noexcept
+  {
+    return end();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator crbegin() const noexcept
+  {
+    return rbegin();
+  }
+  [[nodiscard]] _CCCL_API constexpr const_reverse_iterator crend() const noexcept
+  {
+    return rend();
+  }
+
+  // capacity:
+  [[nodiscard]] _CCCL_API constexpr size_type size() const noexcept
+  {
+    return 0;
+  }
+  [[nodiscard]] _CCCL_API constexpr size_type max_size() const noexcept
+  {
+    return 0;
+  }
+  [[nodiscard]] [[nodiscard]] _CCCL_API constexpr bool empty() const noexcept
+  {
+    return true;
+  }
+
+  // element access:
+  [[nodiscard]] _CCCL_API constexpr reference operator[](size_type) noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::operator[] on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_reference operator[](size_type) const noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::operator[] on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reference at(size_type)
+  {
+    _CCCL_THROW(::std::out_of_range, "array<T, 0>::at");
+    _CCCL_UNREACHABLE();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_reference at(size_type) const
+  {
+    _CCCL_THROW(::std::out_of_range, "array<T, 0>::at");
+    _CCCL_UNREACHABLE();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reference front() noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::front() on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_reference front() const noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::front() on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr reference back() noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::back() on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+
+  [[nodiscard]] _CCCL_API constexpr const_reference back() const noexcept
+  {
+    _CCCL_ASSERT(false, "cannot call array<T, 0>::back() on a zero-sized array");
+    _CCCL_UNREACHABLE();
+    return *data();
+  }
+};
+
+_CCCL_DIAG_POP
+
+_CCCL_TEMPLATE(class _Tp, class... _Args)
+_CCCL_REQUIRES((is_same_v<_Tp, _Args> && ...))
+_CCCL_DEDUCTION_GUIDE_ATTRIBUTES array(_Tp, _Args...) -> array<_Tp, 1 + sizeof...(_Args)>;
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator==(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return ::cuda::std::equal(__x.begin(), __x.end(), __y.begin());
+}
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator!=(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return !(__x == __y);
+}
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator<(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return ::cuda::std::lexicographical_compare(__x.begin(), __x.end(), __y.begin(), __y.end());
+}
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator>(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return __y < __x;
+}
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator<=(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return !(__y < __x);
+}
+
+template <class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr bool operator>=(const array<_Tp, _Size>& __x, const array<_Tp, _Size>& __y)
+{
+  return !(__x < __y);
+}
+
+_CCCL_TEMPLATE(class _Tp, size_t _Size)
+_CCCL_REQUIRES((_Size == 0) || is_nothrow_swappable_v<_Tp>)
+_CCCL_API constexpr void swap(array<_Tp, _Size>& __x, array<_Tp, _Size>& __y) noexcept(noexcept(__x.swap(__y)))
+{
+  __x.swap(__y);
+}
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr _Tp& get(array<_Tp, _Size>& __a) noexcept
+{
+  static_assert(_Ip < _Size, "Index out of bounds in std::get<> (std::array)");
+  return __a.__elems_[_Ip];
+}
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr const _Tp& get(const array<_Tp, _Size>& __a) noexcept
+{
+  static_assert(_Ip < _Size, "Index out of bounds in std::get<> (const std::array)");
+  return __a.__elems_[_Ip];
+}
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr _Tp&& get(array<_Tp, _Size>&& __a) noexcept
+{
+  static_assert(_Ip < _Size, "Index out of bounds in std::get<> (std::array &&)");
+  return ::cuda::std::move(__a.__elems_[_Ip]);
+}
+
+template <size_t _Ip, class _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr const _Tp&& get(const array<_Tp, _Size>&& __a) noexcept
+{
+  static_assert(_Ip < _Size, "Index out of bounds in std::get<> (const std::array &&)");
+  return ::cuda::std::move(__a.__elems_[_Ip]);
+}
+
+template <typename _Tp, size_t _Size, size_t... _Index>
+[[nodiscard]] _CCCL_API constexpr array<remove_cv_t<_Tp>, _Size>
+__to_array_lvalue_impl(_Tp (&__arr)[_Size], index_sequence<_Index...>)
+{
+  return {{__arr[_Index]...}};
+}
+
+template <typename _Tp, size_t _Size, size_t... _Index>
+[[nodiscard]] _CCCL_API constexpr array<remove_cv_t<_Tp>, _Size>
+__to_array_rvalue_impl(_Tp (&&__arr)[_Size], index_sequence<_Index...>)
+{
+  return {{::cuda::std::move(__arr[_Index])...}};
+}
+
+template <typename _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr array<remove_cv_t<_Tp>, _Size>
+to_array(_Tp (&__arr)[_Size]) noexcept(is_nothrow_constructible_v<_Tp, _Tp&>)
+{
+  static_assert(!is_array_v<_Tp>, "[array.creation]/1: to_array does not accept multidimensional arrays.");
+  static_assert(is_constructible_v<_Tp, _Tp&>, "[array.creation]/1: to_array requires copy constructible elements.");
+  return ::cuda::std::__to_array_lvalue_impl(__arr, make_index_sequence<_Size>());
+}
+
+template <typename _Tp, size_t _Size>
+[[nodiscard]] _CCCL_API constexpr array<remove_cv_t<_Tp>, _Size>
+to_array(_Tp (&&__arr)[_Size]) noexcept(is_nothrow_move_constructible_v<_Tp>)
+{
+  static_assert(!is_array_v<_Tp>, "[array.creation]/4: to_array does not accept multidimensional arrays.");
+  static_assert(is_move_constructible_v<_Tp>, "[array.creation]/4: to_array requires move constructible elements.");
+  return ::cuda::std::__to_array_rvalue_impl(::cuda::std::move(__arr), make_index_sequence<_Size>());
+}
+
+// specialize cuda::std::tuple_size and cuda::std::tuple_element for both std::array and cuda::std::array
+
+#if _CCCL_HAS_HOST_STD_LIB()
+template <class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<::std::array<_Tp, _Np>> : integral_constant<size_t, _Np>
+{};
+
+template <size_t _Ip, class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, ::std::array<_Tp, _Np>>
+{
+  static_assert(_Ip < _Np, "Index out of bounds in cuda::std::tuple_element<> (std::array)");
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+#endif // _CCCL_HAS_HOST_STD_LIB()
+
+template <class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<array<_Tp, _Np>> : integral_constant<size_t, _Np>
+{};
+
+template <size_t _Ip, class _Tp, size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, array<_Tp, _Np>>
+{
+  static_assert(_Ip < _Np, "Index out of bounds in cuda::std::tuple_element<> (cuda::std::array)");
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+// tuple protocol for cuda::std::array
+
+_CCCL_BEGIN_NAMESPACE_STD
+
+template <class _Tp, ::cuda::std::size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_size<::cuda::std::array<_Tp, _Np>>
+    : ::cuda::std::integral_constant<::cuda::std::size_t, _Np>
+{};
+
+template <::cuda::std::size_t _Ip, class _Tp, ::cuda::std::size_t _Np>
+struct _CCCL_TYPE_VISIBILITY_DEFAULT tuple_element<_Ip, ::cuda::std::array<_Tp, _Np>>
+{
+  static_assert(_Ip < _Np, "Index out of bounds in std::tuple_element<> (cuda::std::array)");
+  using type _CCCL_NODEBUG_ALIAS = _Tp;
+};
+
+_CCCL_END_NAMESPACE_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD_ARRAY
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/cassert b/qwen3_6_scripts/cccl_preload/include/cuda/std/cassert
new file mode 100644
index 00000000..ace1d2a8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/cassert
@@ -0,0 +1,27 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_CASSERT
+#define _CUDA_STD_CASSERT
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <cassert>
+#endif // _CCCL_HOSTED()
+#endif // _CUDA_STD_CASSERT
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/cfloat b/qwen3_6_scripts/cccl_preload/include/cuda/std/cfloat
new file mode 100644
index 00000000..2bd0431a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/cfloat
@@ -0,0 +1,59 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_CFLOAT
+#define _CUDA_STD_CFLOAT
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <cfloat>
+#else // ^^^ _CCCL_HOSTED() ^^^ / vvv _CCCL_FREESTANDING() vvv
+#  define FLT_RADIX       2
+#  define DECIMAL_DIG     DBL_DECIMAL_DIG
+#  define FLT_ROUNDS      1
+#  define FLT_EVAL_METHOD 0
+
+#  define FLT_DECIMAL_DIG 9
+#  define FLT_MIN         1.17549435082228750796873653722224568e-38f
+#  define FLT_TRUE_MIN    1.40129846432481707092372958328991613e-45f
+#  define FLT_MAX         3.40282346638528859811704183484516925e+38f
+#  define FLT_EPSILON     1.19209289550781250000000000000000000e-7f
+#  define FLT_DIG         6
+#  define FLT_MANT_DIG    24
+#  define FLT_MIN_EXP     (-125)
+#  define FLT_MIN_10_EXP  (-37)
+#  define FLT_MAX_EXP     128
+#  define FLT_MAX_10_EXP  38
+#  define FLT_HAS_SUBNORM 1
+
+#  define DBL_DECIMAL_DIG 17
+#  define DBL_MIN         2.22507385850720138309023271733240406e-308
+#  define DBL_TRUE_MIN    4.94065645841246544176568792868221372e-324
+#  define DBL_MAX         1.79769313486231570814527423731704357e+308
+#  define DBL_EPSILON     2.22044604925031308084726333618164062e-16
+#  define DBL_DIG         15
+#  define DBL_MANT_DIG    53
+#  define DBL_MIN_EXP     (-1021)
+#  define DBL_MIN_10_EXP  (-307)
+#  define DBL_MAX_EXP     1024
+#  define DBL_MAX_10_EXP  308
+#  define DBL_HAS_SUBNORM 1
+#endif // ^^^ _CCCL_FREESTANDING() ^^^
+
+#endif // _CUDA_STD_CFLOAT
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/climits b/qwen3_6_scripts/cccl_preload/include/cuda/std/climits
new file mode 100644
index 00000000..c4135118
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/climits
@@ -0,0 +1,61 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_CLIMITS
+#define _CUDA_STD_CLIMITS
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HOSTED()
+#  include <climits>
+#else // ^^^ _CCCL_HOSTED() ^^^ / vvv _CCCL_FREESTANDING() vvv
+#  define _CCCL_CHAR_IS_UNSIGNED() ('\xff' > 0) // CURSED
+
+#  define CHAR_BIT 8
+
+#  define SCHAR_MIN (-128)
+#  define SCHAR_MAX 127
+#  define UCHAR_MAX 255
+
+#  if _CCCL_CHAR_IS_UNSIGNED()
+#    define CHAR_MIN 0
+#    define CHAR_MAX UCHAR_MAX
+#  else
+#    define CHAR_MIN SCHAR_MIN
+#    define CHAR_MAX SCHAR_MAX
+#  endif
+#  define SHRT_MIN  (-SHRT_MAX - 1)
+#  define SHRT_MAX  0x7fff
+#  define USHRT_MAX 0xffff
+#  define INT_MIN   (-INT_MAX - 1)
+#  define INT_MAX   0x7fffffff
+#  define UINT_MAX  0xffffffff
+#  define LONG_MIN  (-LONG_MAX - 1)
+#  ifdef __LP64__
+#    define LONG_MAX  LLONG_MAX
+#    define ULONG_MAX ULLONG_MAX
+#  else
+#    define LONG_MAX  INT_MAX
+#    define ULONG_MAX UINT_MAX
+#  endif
+#  define LLONG_MIN  (-LLONG_MAX - 1)
+#  define LLONG_MAX  0x7fffffffffffffffLL
+#  define ULLONG_MAX 0xffffffffffffffffUL
+#endif // _CCCL_FREESTANDING()
+
+#endif // _CUDA_STD_CLIMITS
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/cstddef b/qwen3_6_scripts/cccl_preload/include/cuda/std/cstddef
new file mode 100644
index 00000000..d02201f8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/cstddef
@@ -0,0 +1,28 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_CSTDDEF
+#define _CUDA_STD_CSTDDEF
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__cstddef/byte.h>
+#include <cuda/std/__cstddef/types.h>
+#include <cuda/std/version>
+
+#endif // _CUDA_STD_CSTDDEF
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/cstdint b/qwen3_6_scripts/cccl_preload/include/cuda/std/cstdint
new file mode 100644
index 00000000..66416195
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/cstdint
@@ -0,0 +1,204 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_CSTDINT
+#define _CUDA_STD_CSTDINT
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/version>
+
+#if _CCCL_HOSTED()
+#  include <cstdint>
+#elif _CCCL_COMPILER(NVRTC) || (_CCCL_HOSTJIT() && _CCCL_OS(WINDOWS))
+#  include <cuda/std/climits>
+
+using int8_t   = signed char;
+using int16_t  = signed short;
+using int32_t  = signed int;
+using int64_t  = signed long long;
+using uint8_t  = unsigned char;
+using uint16_t = unsigned short;
+using uint32_t = unsigned int;
+using uint64_t = unsigned long long;
+
+#  if _CCCL_OS(WINDOWS) // See https://learn.microsoft.com/en-us/cpp/c-runtime-library/standard-types?view=msvc-170
+using int_fast8_t   = signed char;
+using int_fast16_t  = int;
+using int_fast32_t  = int;
+using int_fast64_t  = long long;
+using uint_fast8_t  = unsigned char;
+using uint_fast16_t = unsigned int;
+using uint_fast32_t = unsigned int;
+using uint_fast64_t = unsigned long long;
+#  else // ^^^ _CCCL_OS(WINDOWS) ^^^ / vvv !_CCCL_OS(WINDOWS) vvv
+using int_fast8_t   = int8_t;
+using int_fast16_t  = int16_t;
+using int_fast32_t  = int32_t;
+using int_fast64_t  = int64_t;
+using uint_fast8_t  = uint8_t;
+using uint_fast16_t = uint16_t;
+using uint_fast32_t = uint32_t;
+using uint_fast64_t = uint64_t;
+#  endif // !_CCCL_OS(WINDOWS)
+
+using int_least8_t   = int8_t;
+using int_least16_t  = int16_t;
+using int_least32_t  = int32_t;
+using int_least64_t  = int64_t;
+using uint_least8_t  = uint8_t;
+using uint_least16_t = uint16_t;
+using uint_least32_t = uint32_t;
+using uint_least64_t = uint64_t;
+
+using intptr_t  = int64_t;
+using uintptr_t = uint64_t;
+
+using intmax_t  = int64_t;
+using uintmax_t = uint64_t;
+
+#  define INT8_MIN   SCHAR_MIN
+#  define INT16_MIN  SHRT_MIN
+#  define INT32_MIN  INT_MIN
+#  define INT64_MIN  LLONG_MIN
+#  define INT8_MAX   SCHAR_MAX
+#  define INT16_MAX  SHRT_MAX
+#  define INT32_MAX  INT_MAX
+#  define INT64_MAX  LLONG_MAX
+#  define UINT8_MAX  UCHAR_MAX
+#  define UINT16_MAX USHRT_MAX
+#  define UINT32_MAX UINT_MAX
+#  define UINT64_MAX ULLONG_MAX
+
+#  if _CCCL_OS(WINDOWS) // See https://learn.microsoft.com/en-us/cpp/c-runtime-library/standard-types?view=msvc-170
+#    define INT_FAST8_MIN   INT8_MIN
+#    define INT_FAST16_MIN  INT32_MIN
+#    define INT_FAST32_MIN  INT32_MIN
+#    define INT_FAST64_MIN  INT64_MIN
+#    define INT_FAST8_MAX   INT8_MAX
+#    define INT_FAST16_MAX  INT32_MAX
+#    define INT_FAST32_MAX  INT32_MAX
+#    define INT_FAST64_MAX  INT64_MAX
+#    define UINT_FAST8_MAX  UINT8_MAX
+#    define UINT_FAST16_MAX UINT32_MAX
+#    define UINT_FAST32_MAX UINT32_MAX
+#    define UINT_FAST64_MAX UINT64_MAX
+#  else // ^^^ _CCCL_OS(WINDOWS) ^^^ / vvv !_CCCL_OS(WINDOWS) vvv
+#    define INT_FAST8_MIN   INT8_MIN
+#    define INT_FAST16_MIN  INT16_MIN
+#    define INT_FAST32_MIN  INT32_MIN
+#    define INT_FAST64_MIN  INT64_MIN
+#    define INT_FAST8_MAX   INT8_MAX
+#    define INT_FAST16_MAX  INT16_MAX
+#    define INT_FAST32_MAX  INT32_MAX
+#    define INT_FAST64_MAX  INT64_MAX
+#    define UINT_FAST8_MAX  UINT8_MAX
+#    define UINT_FAST16_MAX UINT16_MAX
+#    define UINT_FAST32_MAX UINT32_MAX
+#    define UINT_FAST64_MAX UINT64_MAX
+#  endif // !_CCCL_OS(WINDOWS)
+
+#  define INT_LEAST8_MIN   INT8_MIN
+#  define INT_LEAST16_MIN  INT16_MIN
+#  define INT_LEAST32_MIN  INT32_MIN
+#  define INT_LEAST64_MIN  INT64_MIN
+#  define INT_LEAST8_MAX   INT8_MAX
+#  define INT_LEAST16_MAX  INT16_MAX
+#  define INT_LEAST32_MAX  INT32_MAX
+#  define INT_LEAST64_MAX  INT64_MAX
+#  define UINT_LEAST8_MAX  UINT8_MAX
+#  define UINT_LEAST16_MAX UINT16_MAX
+#  define UINT_LEAST32_MAX UINT32_MAX
+#  define UINT_LEAST64_MAX UINT64_MAX
+
+#  define INTPTR_MIN  INT64_MIN
+#  define INTPTR_MAX  INT64_MAX
+#  define UINTPTR_MAX UINT64_MAX
+
+#  define INTMAX_MIN  INT64_MIN
+#  define INTMAX_MAX  INT64_MAX
+#  define UINTMAX_MAX UINT64_MAX
+
+#  define PTRDIFF_MIN INT64_MIN
+#  define PTRDIFF_MAX INT64_MAX
+
+#  define SIZE_MAX UINT64_MAX
+
+#  define INT8_C(X)   ((::int_least8_t) (X))
+#  define INT16_C(X)  ((::int_least16_t) (X))
+#  define INT32_C(X)  ((::int_least32_t) (X))
+#  define INT64_C(X)  ((::int_least64_t) (X))
+#  define UINT8_C(X)  ((::uint_least8_t) (X))
+#  define UINT16_C(X) ((::uint_least16_t) (X))
+#  define UINT32_C(X) ((::uint_least32_t) (X))
+#  define UINT64_C(X) ((::uint_least64_t) (X))
+
+#  define INTMAX_C(X)  ((::intmax_t) (X))
+#  define UINTMAX_C(X) ((::uintmax_t) (X))
+#endif // _CCCL_COMPILER(NVRTC) || (_CCCL_HOSTJIT() && _CCCL_OS(WINDOWS))
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+using ::int16_t;
+using ::int32_t;
+using ::int64_t;
+using ::int8_t;
+using ::uint16_t;
+using ::uint32_t;
+using ::uint64_t;
+using ::uint8_t;
+
+using ::int_fast16_t;
+using ::int_fast32_t;
+using ::int_fast64_t;
+using ::int_fast8_t;
+using ::uint_fast16_t;
+using ::uint_fast32_t;
+using ::uint_fast64_t;
+using ::uint_fast8_t;
+
+using ::int_least16_t;
+using ::int_least32_t;
+using ::int_least64_t;
+using ::int_least8_t;
+using ::uint_least16_t;
+using ::uint_least32_t;
+using ::uint_least64_t;
+using ::uint_least8_t;
+
+using ::intptr_t;
+using ::uintptr_t;
+
+using ::intmax_t;
+using ::uintmax_t;
+
+#if _CCCL_HAS_INT128()
+using __cccl_intmax_t  = __int128_t;
+using __cccl_uintmax_t = __uint128_t;
+#else // ^^^ _CCCL_HAS_INT128() ^^^ / vvv !_CCCL_HAS_INT128() vvv
+using __cccl_intmax_t  = long long;
+using __cccl_uintmax_t = unsigned long long;
+#endif // !_CCCL_HAS_INT128()
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD_CSTDINT
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/detail/__config b/qwen3_6_scripts/cccl_preload/include/cuda/std/detail/__config
new file mode 100644
index 00000000..a3ebabc5
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/detail/__config
@@ -0,0 +1,22 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef __cuda_std__
+#define __cuda_std__
+
+#include <cuda/std/__internal/atomic.h>
+#include <cuda/std/__internal/cpp_dialect.h>
+#include <cuda/std/__internal/features.h>
+#include <cuda/std/__internal/namespaces.h>
+#include <cuda/std/__internal/pstl_config.h>
+#include <cuda/std/__internal/thread_api.h>
+#include <cuda/std/__internal/version.h>
+
+#endif //__cuda_std__
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/initializer_list b/qwen3_6_scripts/cccl_preload/include/cuda/std/initializer_list
new file mode 100644
index 00000000..29eca098
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/initializer_list
@@ -0,0 +1,48 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_INITIALIZER_LIST
+#define _CUDA_STD_INITIALIZER_LIST
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if !_CCCL_COMPILER(NVRTC)
+#  include <initializer_list>
+#else // ^^^ !_CCCL_COMPILER(NVRTC) ^^^ / vvv _CCCL_COMPILER(NVRTC) vvv
+#  if !defined(__NV_BUILTIN_INITIALIZER_LIST)
+#    error "libcu++ requires -builtin-initializer-list=true when compiling with NVRTC"
+#  endif // !__NV_BUILTIN_INITIALIZER_LIST
+#endif // ^^^ _CCCL_COMPILER(NVRTC) ^^^
+
+#include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA_STD
+
+using ::std::initializer_list;
+
+template <class _Tp>
+inline constexpr bool __is_cuda_std_initializer_list = false;
+
+template <class _Ep>
+inline constexpr bool __is_cuda_std_initializer_list<initializer_list<_Ep>> = true;
+
+_CCCL_END_NAMESPACE_CUDA_STD
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CUDA_STD_INITIALIZER_LIST
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/limits b/qwen3_6_scripts/cccl_preload/include/cuda/std/limits
new file mode 100644
index 00000000..91f0c78a
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/limits
@@ -0,0 +1,28 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_LIMITS
+#define _CUDA_STD_LIMITS
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__limits/numeric_limits.h>
+#include <cuda/std/__limits/numeric_limits_ext.h>
+#include <cuda/std/version>
+
+#endif // _CUDA_STD_LIMITS
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/std/version b/qwen3_6_scripts/cccl_preload/include/cuda/std/version
new file mode 100644
index 00000000..7a8785b3
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/std/version
@@ -0,0 +1,243 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_STD_VERSION
+#define _CUDA_STD_VERSION
+
+#include <cuda/std/detail/__config> // IWYU pragma: export
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+// We need to define our own macros to not conflict with the host stl.
+// At the same time we want bring in all feature test macros from host
+#if __has_include(<version>) // <version> should be the smallest include possible
+#  include <version>
+#elif _CCCL_HOSTED()
+#  include <ciso646> // otherwise go for the smallest possible header
+#endif // _CCCL_HOSTED()
+
+#define __cccl_lib_bitops                       201907L
+#define __cccl_lib_bool_constant                201505L
+#define __cccl_lib_bounded_array_traits         201902L
+#define __cccl_lib_byte                         201603L
+#define __cccl_lib_byteswap                     202110L
+#define __cccl_lib_clamp                        201603L
+#define __cccl_lib_endian                       201907L
+#define __cccl_lib_forward_like                 202207L
+#define __cccl_lib_gcd_lcm                      201606L
+#define __cccl_lib_integer_comparison_functions 202002L
+#define __cccl_lib_integer_sequence             202511L
+#define __cccl_lib_integral_constant_callable   201304L
+#define __cccl_lib_is_final                     201402L
+#define __cccl_lib_is_nothrow_convertible       201806L
+#define __cccl_lib_is_null_pointer              201309L
+#define __cccl_lib_is_scoped_enum               202011L
+#define __cccl_lib_is_swappable                 201603L
+#if defined(_CCCL_BUILTIN_IS_VIRTUAL_BASE_OF)
+#  define __cccl_lib_is_virtual_base_of 202406L
+#endif // _CCCL_BUILTIN_IS_VIRTUAL_BASE_OF
+#define __cccl_lib_launder        201606L
+#define __cccl_lib_logical_traits 201510L
+#define __cccl_lib_math_constants 201907L
+#if defined(_CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY) \
+  && defined(_CCCL_BUILTIN_REFERENCE_CONVERTS_FROM_TEMPORARY)
+#  define __cccl_lib_reference_from_temporary 202202L
+#endif // _CCCL_BUILTIN_REFERENCE_CONSTRUCTS_FROM_TEMPORARY && _CCCL_BUILTIN_REFERENCE_CONVERTS_FROM_TEMPORARY
+#define __cccl_lib_shift                        201806L
+#define __cccl_lib_to_address                   201711L
+#define __cccl_lib_to_array                     201907L
+#define __cccl_lib_to_underlying                202102L
+#define __cccl_lib_transparent_operators        201210L
+#define __cccl_lib_transformation_trait_aliases 201304L
+#define __cccl_lib_tuple_element_t              201402L
+// #define __cccl_lib_tuple_like                    202311L // P2819R2 is implemented, but P2165R4 is not yet
+#define __cccl_lib_type_identity                 201806L
+#define __cccl_lib_type_trait_variable_templates 201510L
+#define __cccl_lib_unreachable                   202202L
+#define __cccl_lib_void_t                        201411L
+
+#define __cccl_lib_as_const     201510L
+#define __cccl_lib_bit_cast     201806L
+#define __cccl_lib_chrono_udls  201304L
+#define __cccl_lib_complex_udls 201309L
+#ifdef _CCCL_BUILTIN_IS_CONSTANT_EVALUATED
+#  define __cccl_lib_constexpr_complex 201711L
+#endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED
+#define __cccl_lib_concepts          202002L
+#define __cccl_lib_exchange_function 201304L
+#define __cccl_lib_expected          202606L
+// # define __cccl_lib_generic_associative_lookup           201304L
+#define __cccl_lib_invoke                201411L
+#define __cccl_lib_invoke_r              202106L
+#define __cccl_lib_is_invocable          201703L
+#define __cccl_lib_make_reverse_iterator 201402L
+// # define __cccl_lib_make_unique                          201304L
+#define __cccl_lib_mdspan         202207L
+#define __cccl_lib_null_iterators 201304L
+#define __cccl_lib_optional       202506L
+// # define __cccl_lib_quoted_string_io                     201304L
+#define __cccl_lib_result_of_sfinae            201210L
+#define __cccl_lib_robust_nonmodifying_seq_ops 201304L
+// #   define __cccl_lib_shared_timed_mutex                 201402L
+#define __cccl_lib_source_location 201907L
+#define __cccl_lib_span            202311L
+// # define __cccl_lib_string_udls                          201304L
+#define __cccl_lib_submdspan      202207L
+#define __cccl_lib_tuples_by_type 201304L
+
+#ifdef _CCCL_BUILTIN_ADDRESSOF
+#  define __cccl_lib_addressof_constexpr 201603L
+#endif // _CCCL_BUILTIN_ADDRESSOF
+// # define __cccl_lib_allocator_traits_is_always_equal     201411L
+// # define __cccl_lib_any                                  201606L
+#define __cccl_lib_apply                      201603L
+#define __cccl_lib_array_constexpr            201603L
+#define __cccl_lib_atomic_is_always_lock_free 201603L
+#define __cccl_lib_bind_front                 201907L
+// # define __cccl_lib_boyer_moore_searcher                 201603L
+#define __cccl_lib_chrono 201611L
+// # define __cccl_lib_enable_shared_from_this              201603L
+// # define __cccl_lib_execution                            201603L
+// # define __cccl_lib_filesystem                           201703L
+// #  define __cccl_lib_hardware_interference_size 201703L
+#define __cccl_lib_has_unique_object_representations 201606L
+#define __cccl_lib_hypot                             201603L
+// # define __cccl_lib_incomplete_container_elements        201505L
+#ifndef _LIBCUDACXX_HAS_NO_IS_AGGREGATE
+#  define __cccl_lib_is_aggregate 201703L
+#endif // _LIBCUDACXX_HAS_NO_IS_AGGREGATE
+#if defined(_CCCL_BUILTIN_IS_LAYOUT_COMPATIBLE) && defined(_CCCL_BUILTIN_IS_CORRESPONDING_MEMBER)
+#  define __cccl_lib_is_layout_compatible 201907L
+#endif // _CCCL_BUILTIN_IS_LAYOUT_COMPATIBLE && _CCCL_BUILTIN_IS_CORRESPONDING_MEMBER
+#if defined(_CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_BASE_OF) \
+  && defined(_CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_WITH_CLASS)
+#  define __cccl_lib_is_pointer_interconvertible 201907L
+#endif // _BUILTIN_IS_POINTER_INTERCONVERTIBLE_BASE_OF && _CCCL_BUILTIN_IS_POINTER_INTERCONVERTIBLE_WITH_CLASS
+#define __cccl_lib_make_from_tuple 201606L
+// # define __cccl_lib_map_try_emplace                      201411L
+// # define __cccl_lib_math_special_functions               201603L
+// # define __cccl_lib_memory_resource                      201603L
+// # define __cccl_lib_node_extract                         201606L
+// # define __cccl_lib_nonmember_container_access           201411L
+#define __cccl_lib_not_fn 201603L
+// # define __cccl_lib_parallel_algorithm                   201603L
+// # define __cccl_lib_raw_memory_algorithms                201606L
+// # define __cccl_lib_sample                               201603L
+// # define __cccl_lib_scoped_lock                          201703L
+// #   define __cccl_lib_shared_mutex                       201505L
+// # define __cccl_lib_shared_ptr_arrays                    201611L
+// # define __cccl_lib_shared_ptr_weak_type                 201606L
+#define __cccl_lib_string_view 201803L
+// # define __cccl_lib_to_chars                             201611L
+// #  define __cccl_lib_uncaught_exceptions           201411L
+// #  define __cccl_lib_unordered_map_try_emplace     201411L
+#define __cccl_lib_variant 201606L
+
+#define __cccl_lib_saturation_arithmetic 202603L
+
+#if _CCCL_STD_VER >= 2020
+#  undef __cccl_lib_array_constexpr
+#  define __cccl_lib_array_constexpr 201811L
+// # define __cccl_lib_assume_aligned                       201811L
+#  define __cccl_lib_atomic_flag_test              201907L
+#  define __cccl_lib_atomic_float                  201711L
+#  define __cccl_lib_atomic_lock_free_type_aliases 201907L
+#  define __cccl_lib_atomic_ref                    202603L
+// # define __cccl_lib_atomic_shared_ptr                    201711L
+#  define __cccl_lib_atomic_value_initialization 201911L
+#  define __cccl_lib_atomic_wait                 201907L
+#  define __cccl_lib_barrier                     201907L
+#  define __cccl_lib_bit_cast                    201806L
+#  if _CCCL_HAS_CHAR8_T()
+#    define __cccl_lib_char8_t 201811L
+#  endif // _CCCL_HAS_CHAR8_T()
+// # define __cccl_lib_constexpr_algorithms                 201806L
+// # define __cccl_lib_constexpr_dynamic_alloc              201907L
+#  define __cccl_lib_constexpr_functional 201907L
+// # define __cccl_lib_constexpr_iterator                   201811L
+// # define __cccl_lib_constexpr_memory                     201811L
+// # define __cccl_lib_constexpr_misc                       201811L
+// # define __cccl_lib_constexpr_numeric                    201911L
+// # define __cccl_lib_constexpr_string                     201907L
+// # define __cccl_lib_constexpr_swap_algorithms            201806L
+// # define __cccl_lib_constexpr_tuple                      201811L
+// # define __cccl_lib_constexpr_utility                    201811L
+// # define __cccl_lib_constexpr_vector                     201907L
+// # define __cccl_lib_coroutine                            201902L
+#  if __cpp_impl_destroying_delete >= 201806L && __cpp_lib_destroying_delete >= 201806L
+#    define __cccl_lib_destroying_delete 201806L
+#  endif // __cpp_impl_destroying_delete >= 201806L && __cpp_lib_destroying_delete >= 201806L
+// # define __cccl_lib_erase_if                             201811L
+// # undef  __cccl_lib_execution
+// # define __cccl_lib_execution                            201902L
+// #   define __cccl_lib_format                             202106L
+// # define __cccl_lib_generic_unordered_lookup             201811L
+// # define __cccl_lib_int_pow2                             202002L
+// # define __cccl_lib_interpolate                          201902L
+#  ifdef _CCCL_BUILTIN_IS_CONSTANT_EVALUATED
+#    define __cccl_lib_is_constant_evaluated 201811L
+#  endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED
+// #   define __cccl_lib_jthread                            201911L
+// #   define __cccl_lib_latch                              201907L
+// # define __cccl_lib_list_remove_return_type              201806L
+// # define __cccl_lib_polymorphic_allocator                201902L
+// # define __cccl_lib_ranges                               201811L
+// # define __cccl_lib_remove_cvref                         201711L
+// #   define __cccl_lib_semaphore                          201907L
+// # undef  __cccl_lib_shared_ptr_arrays
+// # define __cccl_lib_shared_ptr_arrays                    201707L
+// # define __cccl_lib_smart_ptr_for_overwrite              202002L
+// # define __cccl_lib_source_location                      201907L
+// # define __cccl_lib_ssize                                201902L
+// # define __cccl_lib_starts_ends_with                     201711L
+// # define __cccl_lib_syncbuf                              201803L
+// # define __cccl_lib_three_way_comparison                 201907L
+#  define __cccl_lib_unwrap_ref 201811L
+
+// todo(dabayer): Uncomment this once constant_wrapper is exposed.
+// #  define __cccl_lib_constant_wrapper 202603L
+#endif // _CCCL_STD_VER >= 2020
+
+#if _CCCL_STD_VER >= 2023
+// # define __cccl_lib_adaptor_iterator_pair_constructor    202106L
+// # define __cccl_lib_allocate_at_least                    202106L
+// # define __cccl_lib_associative_heterogeneous_erasure    202110L
+// # define __cccl_lib_bind_back                            202202L
+// # define __cccl_lib_constexpr_bitset                     202207L
+// # define __cccl_lib_constexpr_charconv                   202207L
+// # define __cccl_lib_constexpr_cmath                      202202L
+// # undef  __cccl_lib_constexpr_memory
+// # define __cccl_lib_constexpr_memory                     202202L
+// # define __cccl_lib_constexpr_typeinfo                   202106L
+// # define __cccl_lib_move_only_function                   202110L
+// # define __cccl_lib_out_ptr                              202106L
+// # define __cccl_lib_ranges_chunk                         202202L
+// # define __cccl_lib_ranges_chunk_by                      202202L
+// # define __cccl_lib_ranges_iota                          202202L
+// # define __cccl_lib_ranges_join_with                     202202L
+// # define __cccl_lib_ranges_slide                         202202L
+// # define __cccl_lib_ranges_starts_ends_with              202106L
+// # define __cccl_lib_ranges_to_container                  202202L
+// # define __cccl_lib_ranges_zip                           202110L
+// # define __cccl_lib_reference_from_temporary             202202L
+// # define __cccl_lib_spanstream                           202106L
+// # define __cccl_lib_stacktrace                           202011L
+// # define __cccl_lib_stdatomic_h                          202011L
+// # define __cccl_lib_string_contains                      202011L
+// # define __cccl_lib_string_resize_and_overwrite          202110L
+#endif // _CCCL_STD_VER >= 2023
+
+#endif // _CUDA_STD_VERSION
diff --git a/qwen3_6_scripts/cccl_preload/include/cuda/version b/qwen3_6_scripts/cccl_preload/include/cuda/version
new file mode 100644
index 00000000..e99928f8
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/cuda/version
@@ -0,0 +1,16 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA_VERSION
+#define _CUDA_VERSION
+
+#include <cuda/std/version> // IWYU pragma: export
+
+#endif // _CUDA_VERSION
diff --git a/qwen3_6_scripts/cccl_preload/include/nv/detail/__preprocessor b/qwen3_6_scripts/cccl_preload/include/nv/detail/__preprocessor
new file mode 100644
index 00000000..30657e6e
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/nv/detail/__preprocessor
@@ -0,0 +1,169 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#if defined(__GNUC__)
+#  pragma GCC system_header
+#endif
+
+// For all compilers and dialects this header defines:
+//  _NV_EVAL
+//  _NV_IF
+//  _NV_CONCAT_EVAL
+// For C++11 and up it defines:
+//  _NV_STRIP_PAREN
+//  _NV_DISPATCH_N_ARY
+//  _NV_FIRST_ARG
+//  _NV_REMOVE_PAREN
+
+#define _NV_EVAL1(...) __VA_ARGS__
+#define _NV_EVAL(...)  _NV_EVAL1(__VA_ARGS__)
+
+#define _NV_CONCAT_EVAL1(l, r) _NV_EVAL(l##r)
+#define _NV_CONCAT_EVAL(l, r)  _NV_CONCAT_EVAL1(l, r)
+
+#define _NV_IF_0(t, f) f
+#define _NV_IF_1(t, f) t
+
+#define _NV_IF_BIT(b)           _NV_EVAL(_NV_IF_##b)
+#define _NV_IF__EVAL(fn, t, f)  _NV_EVAL(fn(t, f))
+#define _NV_IF_EVAL(cond, t, f) _NV_IF__EVAL(_NV_IF_BIT(cond), t, f)
+
+#define _NV_IF1(cond, t, f) _NV_IF_EVAL(cond, t, f)
+#define _NV_IF(cond, t, f)  _NV_IF1(_NV_EVAL(cond), _NV_EVAL(t), _NV_EVAL(f))
+
+// The below mechanisms were derived from: https://gustedt.wordpress.com/2010/06/08/detect-empty-macro-arguments/
+
+#define _NV_ARG32(...) _NV_EVAL(_NV_ARG32_0(__VA_ARGS__))
+#define _NV_ARG32_0( \
+  _0,                \
+  _1,                \
+  _2,                \
+  _3,                \
+  _4,                \
+  _5,                \
+  _6,                \
+  _7,                \
+  _8,                \
+  _9,                \
+  _10,               \
+  _11,               \
+  _12,               \
+  _13,               \
+  _14,               \
+  _15,               \
+  _16,               \
+  _17,               \
+  _18,               \
+  _19,               \
+  _20,               \
+  _21,               \
+  _22,               \
+  _23,               \
+  _24,               \
+  _25,               \
+  _26,               \
+  _27,               \
+  _28,               \
+  _29,               \
+  _30,               \
+  _31,               \
+  ...)               \
+  _31
+
+#define _NV_HAS_COMMA(...) \
+  _NV_ARG32(__VA_ARGS__, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0)
+
+#define _NV_TRIGGER_PARENTHESIS_(...) ,
+
+/*
+This tests a variety of conditions for determining what the incoming statement is.
+1. test if there is just one argument
+2. test if _TRIGGER_PARENTHESIS_ together with the argument adds a comma
+3. test if the argument together with a parenthesis adds a comma
+4. test if placing it between _TRIGGER_PARENTHESIS_ and the parenthesis adds a comma
+*/
+#define _NV_ISEMPTY(...)                                                      \
+  _NV_ISEMPTY0(_NV_EVAL(_NV_HAS_COMMA(__VA_ARGS__)),                          \
+               _NV_EVAL(_NV_HAS_COMMA(_NV_TRIGGER_PARENTHESIS_ __VA_ARGS__)), \
+               _NV_EVAL(_NV_HAS_COMMA(__VA_ARGS__(/*empty*/))),               \
+               _NV_EVAL(_NV_HAS_COMMA(_NV_TRIGGER_PARENTHESIS_ __VA_ARGS__(/*empty*/))))
+
+#define _NV_PASTE5(_0, _1, _2, _3, _4) _0##_1##_2##_3##_4
+#define _NV_ISEMPTY0(_0, _1, _2, _3)   _NV_HAS_COMMA(_NV_PASTE5(_NV_IS_EMPTY_CASE_, _0, _1, _2, _3))
+#define _NV_IS_EMPTY_CASE_0001         ,
+
+#define _NV_REMOVE_PAREN(...) _NV_REMOVE_PAREN1(__VA_ARGS__)
+#define _NV_REMOVE_PAREN1(...) \
+  _NV_STRIP_PAREN(_NV_IF(_NV_TEST_PAREN(__VA_ARGS__), (_NV_STRIP_PAREN(__VA_ARGS__)), (__VA_ARGS__)))
+
+#define _NV_STRIP_PAREN2(...) __VA_ARGS__
+#define _NV_STRIP_PAREN1(...) _NV_STRIP_PAREN2 __VA_ARGS__
+#define _NV_STRIP_PAREN(...)  _NV_STRIP_PAREN1(__VA_ARGS__)
+
+#define _NV_TEST_PAREN(...)  _NV_TEST_PAREN1(__VA_ARGS__)
+#define _NV_TEST_PAREN1(...) _NV_TEST_PAREN2(_NV_TEST_PAREN_DUMMY __VA_ARGS__)
+#define _NV_TEST_PAREN2(...) _NV_TEST_PAREN3(_NV_CONCAT_EVAL(_, __VA_ARGS__))
+#define _NV_TEST_PAREN3(...) _NV_EVAL(_NV_FIRST_ARG(__VA_ARGS__))
+
+#define __NV_PAREN_YES 1
+#define __NV_PAREN_NO  0
+
+#define _NV_TEST_PAREN_DUMMY(...) _NV_PAREN_YES
+#define __NV_TEST_PAREN_DUMMY     __NV_PAREN_NO,
+
+#define _NV_FIRST_ARG1(x, ...) x
+#define _NV_FIRST_ARG(x, ...)  _NV_FIRST_ARG1(x)
+
+#define _NV_REMOVE_FIRST_ARGS1(...)   __VA_ARGS__
+#define _NV_REMOVE_FIRST_ARGS(x, ...) _NV_REMOVE_FIRST_ARGS1(__VA_ARGS__)
+
+#define _NV_NUM_ARGS(...)  _NV_NUM_ARGS0(__VA_ARGS__)
+#define _NV_NUM_ARGS0(...) _NV_EVAL(_NV_NUM_ARGS1(__VA_ARGS__))
+#define _NV_NUM_ARGS1(...) _NV_IF(_NV_ISEMPTY(__VA_ARGS__), 0, _NV_NUM_ARGS2(__VA_ARGS__))
+#define _NV_NUM_ARGS2(...) \
+  _NV_ARG32(               \
+    __VA_ARGS__,           \
+    31,                    \
+    30,                    \
+    29,                    \
+    28,                    \
+    27,                    \
+    26,                    \
+    25,                    \
+    24,                    \
+    23,                    \
+    22,                    \
+    21,                    \
+    20,                    \
+    19,                    \
+    18,                    \
+    17,                    \
+    16,                    \
+    15,                    \
+    14,                    \
+    13,                    \
+    12,                    \
+    11,                    \
+    10,                    \
+    9,                     \
+    8,                     \
+    7,                     \
+    6,                     \
+    5,                     \
+    4,                     \
+    3,                     \
+    2,                     \
+    1,                     \
+    0)
+
+#define _NV_DISPATCH_N_IMPL1(name, ...)        _NV_EVAL(name(__VA_ARGS__))
+#define _NV_DISPATCH_N_IMPL0(depth, name, ...) _NV_DISPATCH_N_IMPL1(_NV_CONCAT_EVAL(name, depth), __VA_ARGS__)
+#define _NV_DISPATCH_N_IMPL(name, ...)         _NV_DISPATCH_N_IMPL0(_NV_NUM_ARGS(__VA_ARGS__), name, __VA_ARGS__)
+#define _NV_DISPATCH_N_ARY(name, ...)          _NV_DISPATCH_N_IMPL(name, __VA_ARGS__)
diff --git a/qwen3_6_scripts/cccl_preload/include/nv/detail/__target_macros b/qwen3_6_scripts/cccl_preload/include/nv/detail/__target_macros
new file mode 100644
index 00000000..40a72e90
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/nv/detail/__target_macros
@@ -0,0 +1,794 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _NV__TARGET_MACROS
+#define _NV__TARGET_MACROS
+
+#include <nv/detail/__preprocessor>
+
+#if defined(__GNUC__)
+#  pragma GCC system_header
+#endif
+
+#define _NV_TARGET_ARCH_TO_SELECTOR_350  nv::target::sm_35
+#define _NV_TARGET_ARCH_TO_SELECTOR_370  nv::target::sm_37
+#define _NV_TARGET_ARCH_TO_SELECTOR_500  nv::target::sm_50
+#define _NV_TARGET_ARCH_TO_SELECTOR_520  nv::target::sm_52
+#define _NV_TARGET_ARCH_TO_SELECTOR_530  nv::target::sm_53
+#define _NV_TARGET_ARCH_TO_SELECTOR_600  nv::target::sm_60
+#define _NV_TARGET_ARCH_TO_SELECTOR_610  nv::target::sm_61
+#define _NV_TARGET_ARCH_TO_SELECTOR_620  nv::target::sm_62
+#define _NV_TARGET_ARCH_TO_SELECTOR_700  nv::target::sm_70
+#define _NV_TARGET_ARCH_TO_SELECTOR_720  nv::target::sm_72
+#define _NV_TARGET_ARCH_TO_SELECTOR_750  nv::target::sm_75
+#define _NV_TARGET_ARCH_TO_SELECTOR_800  nv::target::sm_80
+#define _NV_TARGET_ARCH_TO_SELECTOR_860  nv::target::sm_86
+#define _NV_TARGET_ARCH_TO_SELECTOR_870  nv::target::sm_87
+#define _NV_TARGET_ARCH_TO_SELECTOR_880  nv::target::sm_88
+#define _NV_TARGET_ARCH_TO_SELECTOR_890  nv::target::sm_89
+#define _NV_TARGET_ARCH_TO_SELECTOR_900  nv::target::sm_90
+#define _NV_TARGET_ARCH_TO_SELECTOR_1000 nv::target::sm_100
+#define _NV_TARGET_ARCH_TO_SELECTOR_1030 nv::target::sm_103
+#define _NV_TARGET_ARCH_TO_SELECTOR_1070 nv::target::sm_107
+#define _NV_TARGET_ARCH_TO_SELECTOR_1100 nv::target::sm_110
+#define _NV_TARGET_ARCH_TO_SELECTOR_1200 nv::target::sm_120
+#define _NV_TARGET_ARCH_TO_SELECTOR_1210 nv::target::sm_121
+
+#define _NV_TARGET_ARCH_TO_SM_350  35
+#define _NV_TARGET_ARCH_TO_SM_370  37
+#define _NV_TARGET_ARCH_TO_SM_500  50
+#define _NV_TARGET_ARCH_TO_SM_520  52
+#define _NV_TARGET_ARCH_TO_SM_530  53
+#define _NV_TARGET_ARCH_TO_SM_600  60
+#define _NV_TARGET_ARCH_TO_SM_610  61
+#define _NV_TARGET_ARCH_TO_SM_620  62
+#define _NV_TARGET_ARCH_TO_SM_700  70
+#define _NV_TARGET_ARCH_TO_SM_720  72
+#define _NV_TARGET_ARCH_TO_SM_750  75
+#define _NV_TARGET_ARCH_TO_SM_800  80
+#define _NV_TARGET_ARCH_TO_SM_860  86
+#define _NV_TARGET_ARCH_TO_SM_870  87
+#define _NV_TARGET_ARCH_TO_SM_880  88
+#define _NV_TARGET_ARCH_TO_SM_890  89
+#define _NV_TARGET_ARCH_TO_SM_900  90
+#define _NV_TARGET_ARCH_TO_SM_1000 100
+#define _NV_TARGET_ARCH_TO_SM_1030 103
+#define _NV_TARGET_ARCH_TO_SM_1070 107
+#define _NV_TARGET_ARCH_TO_SM_1100 110
+#define _NV_TARGET_ARCH_TO_SM_1200 120
+#define _NV_TARGET_ARCH_TO_SM_1210 121
+
+// Only enable when compiling for CUDA/stdpar
+#if defined(_NV_COMPILER_NVCXX) && defined(_NVHPC_CUDA)
+
+#  define _NV_TARGET_VAL_SM_35  nv::target::sm_35
+#  define _NV_TARGET_VAL_SM_37  nv::target::sm_37
+#  define _NV_TARGET_VAL_SM_50  nv::target::sm_50
+#  define _NV_TARGET_VAL_SM_52  nv::target::sm_52
+#  define _NV_TARGET_VAL_SM_53  nv::target::sm_53
+#  define _NV_TARGET_VAL_SM_60  nv::target::sm_60
+#  define _NV_TARGET_VAL_SM_61  nv::target::sm_61
+#  define _NV_TARGET_VAL_SM_62  nv::target::sm_62
+#  define _NV_TARGET_VAL_SM_70  nv::target::sm_70
+#  define _NV_TARGET_VAL_SM_72  nv::target::sm_72
+#  define _NV_TARGET_VAL_SM_75  nv::target::sm_75
+#  define _NV_TARGET_VAL_SM_80  nv::target::sm_80
+#  define _NV_TARGET_VAL_SM_86  nv::target::sm_86
+#  define _NV_TARGET_VAL_SM_87  nv::target::sm_87
+#  define _NV_TARGET_VAL_SM_88  nv::target::sm_88
+#  define _NV_TARGET_VAL_SM_89  nv::target::sm_89
+#  define _NV_TARGET_VAL_SM_90  nv::target::sm_90
+#  define _NV_TARGET_VAL_SM_100 nv::target::sm_100
+#  define _NV_TARGET_VAL_SM_103 nv::target::sm_103
+#  define _NV_TARGET_VAL_SM_107 nv::target::sm_107
+#  define _NV_TARGET_VAL_SM_110 nv::target::sm_110
+#  define _NV_TARGET_VAL_SM_120 nv::target::sm_120
+#  define _NV_TARGET_VAL_SM_121 nv::target::sm_121
+
+#  define _NV_TARGET___NV_IS_HOST   nv::target::is_host
+#  define _NV_TARGET___NV_IS_DEVICE nv::target::is_device
+
+#  define _NV_TARGET___NV_ANY_TARGET (nv::target::any_target)
+#  define _NV_TARGET___NV_NO_TARGET  (nv::target::no_target)
+
+#  if defined(NV_TARGET_SM_INTEGER_LIST)
+#    define NV_TARGET_MINIMUM_SM_SELECTOR _NV_FIRST_ARG(NV_TARGET_SM_SELECTOR_LIST)
+#    define NV_TARGET_MINIMUM_SM_INTEGER  _NV_FIRST_ARG(NV_TARGET_SM_INTEGER_LIST)
+#    define __CUDA_MINIMUM_ARCH__         _NV_CONCAT_EVAL(_NV_FIRST_ARG(NV_TARGET_SM_INTEGER_LIST), 0)
+#  endif
+
+#  define _NV_TARGET_PROVIDES(q)   nv::target::provides(q)
+#  define _NV_TARGET_IS_EXACTLY(q) nv::target::is_exactly(q)
+
+#elif defined(_NV_COMPILER_NVCC) || defined(_NV_COMPILER_CLANG_CUDA)
+
+#  define _NV_TARGET_VAL_SM_35  350
+#  define _NV_TARGET_VAL_SM_37  370
+#  define _NV_TARGET_VAL_SM_50  500
+#  define _NV_TARGET_VAL_SM_52  520
+#  define _NV_TARGET_VAL_SM_53  530
+#  define _NV_TARGET_VAL_SM_60  600
+#  define _NV_TARGET_VAL_SM_61  610
+#  define _NV_TARGET_VAL_SM_62  620
+#  define _NV_TARGET_VAL_SM_70  700
+#  define _NV_TARGET_VAL_SM_72  720
+#  define _NV_TARGET_VAL_SM_75  750
+#  define _NV_TARGET_VAL_SM_80  800
+#  define _NV_TARGET_VAL_SM_86  860
+#  define _NV_TARGET_VAL_SM_87  870
+#  define _NV_TARGET_VAL_SM_88  880
+#  define _NV_TARGET_VAL_SM_89  890
+#  define _NV_TARGET_VAL_SM_90  900
+#  define _NV_TARGET_VAL_SM_100 1000
+#  define _NV_TARGET_VAL_SM_103 1030
+#  define _NV_TARGET_VAL_SM_107 1070
+#  define _NV_TARGET_VAL_SM_110 1100
+#  define _NV_TARGET_VAL_SM_120 1200
+#  define _NV_TARGET_VAL_SM_121 1210
+
+#  if defined(__CUDA_ARCH__)
+#    define _NV_TARGET_VAL                __CUDA_ARCH__
+#    define NV_TARGET_MINIMUM_SM_SELECTOR _NV_CONCAT_EVAL(_NV_TARGET_ARCH_TO_SELECTOR_, __CUDA_ARCH__)
+#    define NV_TARGET_MINIMUM_SM_INTEGER  _NV_CONCAT_EVAL(_NV_TARGET_ARCH_TO_SM_, __CUDA_ARCH__)
+#    define __CUDA_MINIMUM_ARCH__         __CUDA_ARCH__
+#  endif
+
+#  if defined(__CUDA_ARCH__)
+#    define _NV_TARGET_IS_HOST   0
+#    define _NV_TARGET_IS_DEVICE 1
+#  else
+#    define _NV_TARGET_IS_HOST   1
+#    define _NV_TARGET_IS_DEVICE 0
+#  endif
+
+#  if defined(_NV_TARGET_VAL)
+#    define _NV_DEVICE_CHECK(q) (q)
+#  else
+#    define _NV_DEVICE_CHECK(q) (0)
+#  endif
+
+#  define _NV_TARGET_PROVIDES(q)   _NV_DEVICE_CHECK(_NV_TARGET_VAL >= q)
+#  define _NV_TARGET_IS_EXACTLY(q) _NV_DEVICE_CHECK(_NV_TARGET_VAL == q)
+
+// NVCC/NVCXX not being used, only host dispatches allowed
+#else
+
+#  define _NV_COMPILER_NVCC
+
+#  define _NV_TARGET_VAL_SM_35  350
+#  define _NV_TARGET_VAL_SM_37  370
+#  define _NV_TARGET_VAL_SM_50  500
+#  define _NV_TARGET_VAL_SM_52  520
+#  define _NV_TARGET_VAL_SM_53  530
+#  define _NV_TARGET_VAL_SM_60  600
+#  define _NV_TARGET_VAL_SM_61  610
+#  define _NV_TARGET_VAL_SM_62  620
+#  define _NV_TARGET_VAL_SM_70  700
+#  define _NV_TARGET_VAL_SM_72  720
+#  define _NV_TARGET_VAL_SM_75  750
+#  define _NV_TARGET_VAL_SM_80  800
+#  define _NV_TARGET_VAL_SM_86  860
+#  define _NV_TARGET_VAL_SM_87  870
+#  define _NV_TARGET_VAL_SM_88  880
+#  define _NV_TARGET_VAL_SM_89  890
+#  define _NV_TARGET_VAL_SM_90  900
+#  define _NV_TARGET_VAL_SM_100 1000
+#  define _NV_TARGET_VAL_SM_103 1030
+#  define _NV_TARGET_VAL_SM_107 1070
+#  define _NV_TARGET_VAL_SM_110 1100
+#  define _NV_TARGET_VAL_SM_120 1200
+#  define _NV_TARGET_VAL_SM_121 1210
+
+#  define _NV_TARGET_VAL 0
+
+#  define _NV_TARGET_IS_HOST   1
+#  define _NV_TARGET_IS_DEVICE 0
+
+#  define _NV_DEVICE_CHECK(q) (false)
+
+#  define _NV_TARGET_PROVIDES(q)   _NV_DEVICE_CHECK(_NV_TARGET_VAL >= q)
+#  define _NV_TARGET_IS_EXACTLY(q) _NV_DEVICE_CHECK(_NV_TARGET_VAL == q)
+
+#endif
+
+#define _NV_TARGET___NV_PROVIDES_SM_35  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_35))
+#define _NV_TARGET___NV_PROVIDES_SM_37  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_37))
+#define _NV_TARGET___NV_PROVIDES_SM_50  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_50))
+#define _NV_TARGET___NV_PROVIDES_SM_52  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_52))
+#define _NV_TARGET___NV_PROVIDES_SM_53  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_53))
+#define _NV_TARGET___NV_PROVIDES_SM_60  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_60))
+#define _NV_TARGET___NV_PROVIDES_SM_61  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_61))
+#define _NV_TARGET___NV_PROVIDES_SM_62  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_62))
+#define _NV_TARGET___NV_PROVIDES_SM_70  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_70))
+#define _NV_TARGET___NV_PROVIDES_SM_72  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_72))
+#define _NV_TARGET___NV_PROVIDES_SM_75  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_75))
+#define _NV_TARGET___NV_PROVIDES_SM_80  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_80))
+#define _NV_TARGET___NV_PROVIDES_SM_86  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_86))
+#define _NV_TARGET___NV_PROVIDES_SM_87  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_87))
+#define _NV_TARGET___NV_PROVIDES_SM_88  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_88))
+#define _NV_TARGET___NV_PROVIDES_SM_89  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_89))
+#define _NV_TARGET___NV_PROVIDES_SM_90  (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_90))
+#define _NV_TARGET___NV_PROVIDES_SM_100 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_100))
+#define _NV_TARGET___NV_PROVIDES_SM_103 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_103))
+#define _NV_TARGET___NV_PROVIDES_SM_107 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_107))
+#define _NV_TARGET___NV_PROVIDES_SM_110 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_110))
+#define _NV_TARGET___NV_PROVIDES_SM_120 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_120))
+#define _NV_TARGET___NV_PROVIDES_SM_121 (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_121))
+
+#define _NV_TARGET___NV_IS_EXACTLY_SM_35  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_35))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_37  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_37))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_50  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_50))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_52  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_52))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_53  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_53))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_60  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_60))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_61  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_61))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_62  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_62))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_70  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_70))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_72  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_72))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_75  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_75))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_80  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_80))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_86  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_86))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_87  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_87))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_88  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_88))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_89  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_89))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_90  (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_90))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_100 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_100))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_103 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_103))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_107 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_107))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_110 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_110))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_120 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_120))
+#define _NV_TARGET___NV_IS_EXACTLY_SM_121 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_121))
+
+#define NV_PROVIDES_SM_35  __NV_PROVIDES_SM_35
+#define NV_PROVIDES_SM_37  __NV_PROVIDES_SM_37
+#define NV_PROVIDES_SM_50  __NV_PROVIDES_SM_50
+#define NV_PROVIDES_SM_52  __NV_PROVIDES_SM_52
+#define NV_PROVIDES_SM_53  __NV_PROVIDES_SM_53
+#define NV_PROVIDES_SM_60  __NV_PROVIDES_SM_60
+#define NV_PROVIDES_SM_61  __NV_PROVIDES_SM_61
+#define NV_PROVIDES_SM_62  __NV_PROVIDES_SM_62
+#define NV_PROVIDES_SM_70  __NV_PROVIDES_SM_70
+#define NV_PROVIDES_SM_72  __NV_PROVIDES_SM_72
+#define NV_PROVIDES_SM_75  __NV_PROVIDES_SM_75
+#define NV_PROVIDES_SM_80  __NV_PROVIDES_SM_80
+#define NV_PROVIDES_SM_86  __NV_PROVIDES_SM_86
+#define NV_PROVIDES_SM_87  __NV_PROVIDES_SM_87
+#define NV_PROVIDES_SM_88  __NV_PROVIDES_SM_88
+#define NV_PROVIDES_SM_89  __NV_PROVIDES_SM_89
+#define NV_PROVIDES_SM_90  __NV_PROVIDES_SM_90
+#define NV_PROVIDES_SM_100 __NV_PROVIDES_SM_100
+#define NV_PROVIDES_SM_103 __NV_PROVIDES_SM_103
+#define NV_PROVIDES_SM_107 __NV_PROVIDES_SM_107
+#define NV_PROVIDES_SM_110 __NV_PROVIDES_SM_110
+#define NV_PROVIDES_SM_120 __NV_PROVIDES_SM_120
+#define NV_PROVIDES_SM_121 __NV_PROVIDES_SM_121
+
+#define NV_IS_EXACTLY_SM_35  __NV_IS_EXACTLY_SM_35
+#define NV_IS_EXACTLY_SM_37  __NV_IS_EXACTLY_SM_37
+#define NV_IS_EXACTLY_SM_50  __NV_IS_EXACTLY_SM_50
+#define NV_IS_EXACTLY_SM_52  __NV_IS_EXACTLY_SM_52
+#define NV_IS_EXACTLY_SM_53  __NV_IS_EXACTLY_SM_53
+#define NV_IS_EXACTLY_SM_60  __NV_IS_EXACTLY_SM_60
+#define NV_IS_EXACTLY_SM_61  __NV_IS_EXACTLY_SM_61
+#define NV_IS_EXACTLY_SM_62  __NV_IS_EXACTLY_SM_62
+#define NV_IS_EXACTLY_SM_70  __NV_IS_EXACTLY_SM_70
+#define NV_IS_EXACTLY_SM_72  __NV_IS_EXACTLY_SM_72
+#define NV_IS_EXACTLY_SM_75  __NV_IS_EXACTLY_SM_75
+#define NV_IS_EXACTLY_SM_80  __NV_IS_EXACTLY_SM_80
+#define NV_IS_EXACTLY_SM_86  __NV_IS_EXACTLY_SM_86
+#define NV_IS_EXACTLY_SM_87  __NV_IS_EXACTLY_SM_87
+#define NV_IS_EXACTLY_SM_88  __NV_IS_EXACTLY_SM_88
+#define NV_IS_EXACTLY_SM_89  __NV_IS_EXACTLY_SM_89
+#define NV_IS_EXACTLY_SM_90  __NV_IS_EXACTLY_SM_90
+#define NV_IS_EXACTLY_SM_100 __NV_IS_EXACTLY_SM_100
+#define NV_IS_EXACTLY_SM_103 __NV_IS_EXACTLY_SM_103
+#define NV_IS_EXACTLY_SM_107 __NV_IS_EXACTLY_SM_107
+#define NV_IS_EXACTLY_SM_110 __NV_IS_EXACTLY_SM_110
+#define NV_IS_EXACTLY_SM_120 __NV_IS_EXACTLY_SM_120
+#define NV_IS_EXACTLY_SM_121 __NV_IS_EXACTLY_SM_121
+
+// Disable SM_90a support on non-supporting compilers.
+// Will re-enable for nvcc below.
+#define NV_HAS_FEATURE_SM_90a  NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_100a NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_103a NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_107a NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_110a NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_120a NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_121a NV_NO_TARGET
+
+#define NV_HAS_FEATURE_SM_100f NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_103f NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_107f NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_110f NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_120f NV_NO_TARGET
+#define NV_HAS_FEATURE_SM_121f NV_NO_TARGET
+
+#define NV_IS_HOST   __NV_IS_HOST
+#define NV_IS_DEVICE __NV_IS_DEVICE
+
+#define NV_ANY_TARGET __NV_ANY_TARGET
+#define NV_NO_TARGET  __NV_NO_TARGET
+
+// Platform invoke mechanisms
+#if defined(_NV_COMPILER_NVCXX) && defined(_NVHPC_CUDA)
+
+#  define _NV_ARCH_COND(q) (_NV_TARGET_##q)
+
+#  define _NV_BLOCK_EXPAND(...) _NV_REMOVE_PAREN(__VA_ARGS__)
+
+#  define _NV_TARGET_IF(cond, t, ...) \
+    (if target _NV_ARCH_COND(cond) { _NV_BLOCK_EXPAND(t) } else {_NV_BLOCK_EXPAND(__VA_ARGS__)})
+
+#elif defined(_NV_COMPILER_NVCC) || defined(_NV_COMPILER_CLANG_CUDA)
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_35)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_35 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_35 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_37)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_37 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_37 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_50)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_50 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_50 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_52)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_52 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_52 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_53)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_53 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_53 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_60)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_60 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_60 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_61)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_61 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_61 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_62)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_62 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_62 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_70)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_70 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_70 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_72)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_72 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_72 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_75)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_75 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_75 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_80)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_80 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_80 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_86)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_86 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_86 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_87)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_87 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_87 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_88)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_88 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_88 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_89)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_89 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_89 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_90)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_90 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_90 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_100)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_103)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_103 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_103 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_107)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_107 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_107 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_110)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_110 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_110 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_120)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_120 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_120 0
+#  endif
+
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_121)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_121 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_121 0
+#  endif
+
+//----------------------------------------------------------------------------------------------------------------------
+// architecture-specific SM versions
+
+#  if defined(__CUDA_ARCH_SPECIFIC__)
+#    define _NV_TARGET_CHECK_ARCH_SPECIFIC(_MAJOR, _MINOR) (__CUDA_ARCH_SPECIFIC__ == (_MAJOR * 100 + _MINOR * 10))
+#  else
+#    define _NV_TARGET_CHECK_ARCH_SPECIFIC(_MAJOR, _MINOR) (0)
+#  endif
+
+// Re-enable sm_90a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_90a
+#  define NV_HAS_FEATURE_SM_90a __NV_HAS_FEATURE_SM_90a
+#  if defined(__CUDA_ARCH_FEAT_SM90_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(9, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_90a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_90a 0
+#  endif
+
+// Re-enable sm_100a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_100a
+#  define NV_HAS_FEATURE_SM_100a __NV_HAS_FEATURE_SM_100a
+#  if defined(__CUDA_ARCH_FEAT_SM100_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(10, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100a 0
+#  endif
+
+// Re-enable sm_103a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_103a
+#  define NV_HAS_FEATURE_SM_103a __NV_HAS_FEATURE_SM_103a
+#  if defined(__CUDA_ARCH_FEAT_SM103_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(10, 3)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103a 0
+#  endif
+
+// Re-enable sm_107a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_107a
+#  define NV_HAS_FEATURE_SM_107a __NV_HAS_FEATURE_SM_107a
+#  if defined(__CUDA_ARCH_FEAT_SM103_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(10, 7)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_107a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_107a 0
+#  endif
+
+// Re-enable sm_110a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_110a
+#  define NV_HAS_FEATURE_SM_110a __NV_HAS_FEATURE_SM_110a
+#  if defined(__CUDA_ARCH_FEAT_SM110_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(11, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110a 0
+#  endif
+
+// Re-enable sm_120a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_120a
+#  define NV_HAS_FEATURE_SM_120a __NV_HAS_FEATURE_SM_120a
+#  if defined(__CUDA_ARCH_FEAT_SM120_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(12, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120a 0
+#  endif
+
+// Re-enable sm_121a support in nvcc.
+#  undef NV_HAS_FEATURE_SM_121a
+#  define NV_HAS_FEATURE_SM_121a __NV_HAS_FEATURE_SM_121a
+#  if defined(__CUDA_ARCH_FEAT_SM121_ALL) || _NV_TARGET_CHECK_ARCH_SPECIFIC(12, 1)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121a 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121a 0
+#  endif
+
+//----------------------------------------------------------------------------------------------------------------------
+// family-specific SM versions
+
+#  if defined(__CUDA_ARCH_FAMILY_SPECIFIC__)
+#    define _NV_TARGET_CHECK_FAMILY_SPECIFIC(_MAJOR, _MINOR)         \
+      (__CUDA_ARCH_FAMILY_SPECIFIC__ >= (_MAJOR * 100 + _MINOR * 10) \
+       && __CUDA_ARCH_FAMILY_SPECIFIC__ < (_MAJOR + 1) * 100)
+#  else
+#    define _NV_TARGET_CHECK_FAMILY_SPECIFIC(_MAJOR, _MINOR) (0)
+#  endif
+
+// Re-enable sm_100f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_100f
+#  define NV_HAS_FEATURE_SM_100f __NV_HAS_FEATURE_SM_100f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(10, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_100f 0
+#  endif
+
+// Re-enable sm_103f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_103f
+#  define NV_HAS_FEATURE_SM_103f __NV_HAS_FEATURE_SM_103f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(10, 3)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_103f 0
+#  endif
+
+// Re-enable sm_107f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_107f
+#  define NV_HAS_FEATURE_SM_107f __NV_HAS_FEATURE_SM_107f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(10, 7)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_107f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_107f 0
+#  endif
+
+// Re-enable sm_110f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_110f
+#  define NV_HAS_FEATURE_SM_110f __NV_HAS_FEATURE_SM_110f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(11, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_110f 0
+#  endif
+
+// Re-enable sm_120f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_120f
+#  define NV_HAS_FEATURE_SM_120f __NV_HAS_FEATURE_SM_120f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(12, 0)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_120f 0
+#  endif
+
+// Re-enable sm_121f support in nvcc.
+#  undef NV_HAS_FEATURE_SM_121f
+#  define NV_HAS_FEATURE_SM_121f __NV_HAS_FEATURE_SM_121f
+#  if _NV_TARGET_CHECK_FAMILY_SPECIFIC(12, 1)
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121f 1
+#  else
+#    define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_121f 0
+#  endif
+
+#  if (_NV_TARGET_IS_HOST)
+#    define _NV_TARGET_BOOL___NV_IS_HOST   1
+#    define _NV_TARGET_BOOL___NV_IS_DEVICE 0
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_HOST   0
+#    define _NV_TARGET_BOOL___NV_IS_DEVICE 1
+#  endif
+
+#  define _NV_TARGET_BOOL___NV_ANY_TARGET 1
+#  define _NV_TARGET_BOOL___NV_NO_TARGET  0
+
+// NVCC Greater than stuff
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_35)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_35 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_35 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_37)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_37 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_37 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_50)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_50 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_50 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_52)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_52 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_52 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_53)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_53 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_53 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_60)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_60 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_60 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_61)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_61 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_61 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_62)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_62 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_62 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_70)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_70 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_70 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_72)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_72 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_72 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_75)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_75 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_75 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_80)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_80 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_80 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_86)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_86 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_86 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_87)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_87 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_87 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_88)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_88 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_88 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_89)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_89 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_89 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_90)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_90 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_90 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_100)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_103)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_103 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_103 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_107)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_107 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_107 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_110)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_110 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_110 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_120)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_120 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_120 0
+#  endif
+
+#  if (_NV_TARGET___NV_PROVIDES_SM_121)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_121 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_121 0
+#  endif
+
+#  define _NV_ARCH_COND_CAT1(cond) _NV_TARGET_BOOL_##cond
+#  define _NV_ARCH_COND_CAT(cond)  _NV_EVAL(_NV_ARCH_COND_CAT1(cond))
+
+#  define _NV_TARGET_EMPTY_PARAM ;
+
+#  define _NV_BLOCK_EXPAND(...)       {_NV_REMOVE_PAREN(__VA_ARGS__)}
+#  define _NV_TARGET_IF(cond, t, ...) _NV_IF(_NV_ARCH_COND_CAT(cond), t, __VA_ARGS__)
+
+#endif // _NV_COMPILER_NVCC
+
+#define _NV_TARGET_DISPATCH_HANDLE0()
+#define _NV_TARGET_DISPATCH_HANDLE2(q, fn)       _NV_TARGET_IF(q, fn)
+#define _NV_TARGET_DISPATCH_HANDLE4(q, fn, ...)  _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE2(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE6(q, fn, ...)  _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE4(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE8(q, fn, ...)  _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE6(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE10(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE8(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE12(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE10(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE14(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE12(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE16(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE14(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE18(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE16(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE20(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE18(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE22(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE20(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE24(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE22(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE26(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE24(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE28(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE26(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE30(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE28(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE32(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE30(__VA_ARGS__))
+#define _NV_TARGET_DISPATCH_HANDLE34(q, fn, ...) _NV_TARGET_IF(q, fn, _NV_TARGET_DISPATCH_HANDLE32(__VA_ARGS__))
+
+#define _NV_TARGET_DISPATCH(...) _NV_BLOCK_EXPAND(_NV_DISPATCH_N_ARY(_NV_TARGET_DISPATCH_HANDLE, __VA_ARGS__))
+
+// NV_IF_TARGET supports a false statement provided as a variadic macro
+#define NV_IF_TARGET(cond, ...)       _NV_BLOCK_EXPAND(_NV_TARGET_IF(cond, __VA_ARGS__))
+#define NV_IF_ELSE_TARGET(cond, t, f) _NV_BLOCK_EXPAND(_NV_TARGET_IF(cond, t, f))
+#define NV_DISPATCH_TARGET(...)       _NV_TARGET_DISPATCH(__VA_ARGS__)
+
+#endif // _NV__TARGET_MACROS
diff --git a/qwen3_6_scripts/cccl_preload/include/nv/target b/qwen3_6_scripts/cccl_preload/include/nv/target
new file mode 100644
index 00000000..31d2518c
--- /dev/null
+++ b/qwen3_6_scripts/cccl_preload/include/nv/target
@@ -0,0 +1,250 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2023 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+// This header contains a preview of a portability system that enables
+// CUDA C++ development with NVC++, NVCC, and supported host compilers.
+// These interfaces are not guaranteed to be stable.
+
+#ifndef __NV_TARGET_H
+#define __NV_TARGET_H
+
+#if defined(__NVCC__) || defined(__CUDACC_RTC__)
+#  define _NV_COMPILER_NVCC
+#elif defined(__NVCOMPILER) && __cplusplus >= 201103L
+#  define _NV_COMPILER_NVCXX
+#elif defined(__clang__) && defined(__CUDA__) && defined(__CUDA_ARCH__)
+// clang compiling CUDA code, device mode.
+#  define _NV_COMPILER_CLANG_CUDA
+#endif
+
+// Hide `if target` support from NVRTC and C
+// Some toolkit headers use <nv/target> in true C contexts
+#if !defined(__CUDACC_RTC__) && defined(__cplusplus)
+
+#  if defined(_NV_COMPILER_NVCXX)
+#    define _NV_BITSET_ATTRIBUTE [[nv::__target_bitset]]
+#  else
+#    define _NV_BITSET_ATTRIBUTE
+#  endif
+
+// This must be a nested namespace as <nv/target> works with older dialects.
+namespace nv // NOLINT(modernize-concat-nested-namespaces)
+{
+namespace target // NOLINT(modernize-concat-nested-namespaces)
+{
+namespace detail
+{
+using base_int_t = unsigned long long;
+
+// No host specialization
+constexpr base_int_t all_hosts = 1;
+
+// NVIDIA GPUs
+constexpr base_int_t sm_35_bit  = 1 << 1;
+constexpr base_int_t sm_37_bit  = 1 << 2;
+constexpr base_int_t sm_50_bit  = 1 << 3;
+constexpr base_int_t sm_52_bit  = 1 << 4;
+constexpr base_int_t sm_53_bit  = 1 << 5;
+constexpr base_int_t sm_60_bit  = 1 << 6;
+constexpr base_int_t sm_61_bit  = 1 << 7;
+constexpr base_int_t sm_62_bit  = 1 << 8;
+constexpr base_int_t sm_70_bit  = 1 << 9;
+constexpr base_int_t sm_72_bit  = 1 << 10;
+constexpr base_int_t sm_75_bit  = 1 << 11;
+constexpr base_int_t sm_80_bit  = 1 << 12;
+constexpr base_int_t sm_86_bit  = 1 << 13;
+constexpr base_int_t sm_87_bit  = 1 << 14;
+constexpr base_int_t sm_88_bit  = 1 << 15;
+constexpr base_int_t sm_89_bit  = 1 << 16;
+constexpr base_int_t sm_90_bit  = 1 << 17;
+constexpr base_int_t sm_100_bit = 1 << 18;
+constexpr base_int_t sm_103_bit = 1 << 19;
+constexpr base_int_t sm_107_bit = 1 << 20;
+constexpr base_int_t sm_110_bit = 1 << 21;
+constexpr base_int_t sm_120_bit = 1 << 22;
+constexpr base_int_t sm_121_bit = 1 << 23;
+constexpr base_int_t all_devices =
+  sm_35_bit | sm_37_bit | sm_50_bit | sm_52_bit | sm_53_bit | sm_60_bit | sm_61_bit | sm_62_bit | sm_70_bit | sm_72_bit
+  | sm_75_bit | sm_80_bit | sm_86_bit | sm_87_bit | sm_88_bit | sm_89_bit | sm_90_bit | sm_100_bit | sm_103_bit
+  | sm_107_bit | sm_110_bit | sm_120_bit | sm_121_bit;
+
+// Store a set of targets as a set of bits
+struct _NV_BITSET_ATTRIBUTE target_description
+{
+  base_int_t targets;
+
+  constexpr target_description(base_int_t a)
+      : targets(a)
+  {}
+};
+
+// The type of the user-visible names of the NVIDIA GPU targets
+enum class sm_selector : base_int_t
+{
+  sm_35  = 35,
+  sm_37  = 37,
+  sm_50  = 50,
+  sm_52  = 52,
+  sm_53  = 53,
+  sm_60  = 60,
+  sm_61  = 61,
+  sm_62  = 62,
+  sm_70  = 70,
+  sm_72  = 72,
+  sm_75  = 75,
+  sm_80  = 80,
+  sm_86  = 86,
+  sm_87  = 87,
+  sm_88  = 88,
+  sm_89  = 89,
+  sm_90  = 90,
+  sm_100 = 100,
+  sm_103 = 103,
+  sm_107 = 107,
+  sm_110 = 110,
+  sm_120 = 120,
+  sm_121 = 121,
+};
+
+constexpr base_int_t toint(sm_selector a)
+{
+  return static_cast<base_int_t>(a);
+}
+
+constexpr base_int_t bitexact(sm_selector a)
+{
+  return toint(a) == 35  ? sm_35_bit
+       : toint(a) == 37  ? sm_37_bit
+       : toint(a) == 50  ? sm_50_bit
+       : toint(a) == 52  ? sm_52_bit
+       : toint(a) == 53  ? sm_53_bit
+       : toint(a) == 60  ? sm_60_bit
+       : toint(a) == 61  ? sm_61_bit
+       : toint(a) == 62  ? sm_62_bit
+       : toint(a) == 70  ? sm_70_bit
+       : toint(a) == 72  ? sm_72_bit
+       : toint(a) == 75  ? sm_75_bit
+       : toint(a) == 80  ? sm_80_bit
+       : toint(a) == 86  ? sm_86_bit
+       : toint(a) == 87  ? sm_87_bit
+       : toint(a) == 88  ? sm_88_bit
+       : toint(a) == 89  ? sm_89_bit
+       : toint(a) == 90  ? sm_90_bit
+       : toint(a) == 100 ? sm_100_bit
+       : toint(a) == 103 ? sm_103_bit
+       : toint(a) == 107 ? sm_107_bit
+       : toint(a) == 110 ? sm_110_bit
+       : toint(a) == 120 ? sm_120_bit
+       : toint(a) == 121 ? sm_121_bit
+                         : 0;
+}
+
+constexpr base_int_t bitrounddown(sm_selector a)
+{
+  return toint(a) >= 121 ? sm_121_bit
+       : toint(a) >= 120 ? sm_120_bit
+       : toint(a) >= 110 ? sm_110_bit
+       : toint(a) >= 107 ? sm_107_bit
+       : toint(a) >= 103 ? sm_103_bit
+       : toint(a) >= 100 ? sm_100_bit
+       : toint(a) >= 90  ? sm_90_bit
+       : toint(a) >= 89  ? sm_89_bit
+       : toint(a) >= 88  ? sm_88_bit
+       : toint(a) >= 87  ? sm_87_bit
+       : toint(a) >= 86  ? sm_86_bit
+       : toint(a) >= 80  ? sm_80_bit
+       : toint(a) >= 75  ? sm_75_bit
+       : toint(a) >= 72  ? sm_72_bit
+       : toint(a) >= 70  ? sm_70_bit
+       : toint(a) >= 62  ? sm_62_bit
+       : toint(a) >= 61  ? sm_61_bit
+       : toint(a) >= 60  ? sm_60_bit
+       : toint(a) >= 53  ? sm_53_bit
+       : toint(a) >= 52  ? sm_52_bit
+       : toint(a) >= 50  ? sm_50_bit
+       : toint(a) >= 37  ? sm_37_bit
+       : toint(a) >= 35  ? sm_35_bit
+                         : 0;
+}
+
+// Public API for NVIDIA GPUs
+
+constexpr target_description is_exactly(sm_selector a)
+{
+  return target_description(bitexact(a));
+}
+
+constexpr target_description provides(sm_selector a)
+{
+  return target_description(~(bitrounddown(a) - 1) & all_devices);
+}
+
+// Boolean operations on target sets
+
+constexpr target_description operator&&(target_description a, target_description b)
+{
+  return target_description(a.targets & b.targets);
+}
+
+constexpr target_description operator||(target_description a, target_description b)
+{
+  return target_description(a.targets | b.targets);
+}
+
+constexpr target_description operator!(target_description a)
+{
+  return target_description(~a.targets & (all_devices | all_hosts));
+}
+} // namespace detail
+
+using detail::sm_selector;
+using detail::target_description;
+
+// The predicates for basic host/device selection
+constexpr target_description is_host    = target_description(detail::all_hosts);
+constexpr target_description is_device  = target_description(detail::all_devices);
+constexpr target_description any_target = target_description(detail::all_hosts | detail::all_devices);
+constexpr target_description no_target  = target_description(0);
+
+// The public names for NVIDIA GPU architectures
+constexpr sm_selector sm_35  = sm_selector::sm_35;
+constexpr sm_selector sm_37  = sm_selector::sm_37;
+constexpr sm_selector sm_50  = sm_selector::sm_50;
+constexpr sm_selector sm_52  = sm_selector::sm_52;
+constexpr sm_selector sm_53  = sm_selector::sm_53;
+constexpr sm_selector sm_60  = sm_selector::sm_60;
+constexpr sm_selector sm_61  = sm_selector::sm_61;
+constexpr sm_selector sm_62  = sm_selector::sm_62;
+constexpr sm_selector sm_70  = sm_selector::sm_70;
+constexpr sm_selector sm_72  = sm_selector::sm_72;
+constexpr sm_selector sm_75  = sm_selector::sm_75;
+constexpr sm_selector sm_80  = sm_selector::sm_80;
+constexpr sm_selector sm_86  = sm_selector::sm_86;
+constexpr sm_selector sm_87  = sm_selector::sm_87;
+constexpr sm_selector sm_88  = sm_selector::sm_88;
+constexpr sm_selector sm_89  = sm_selector::sm_89;
+constexpr sm_selector sm_90  = sm_selector::sm_90;
+constexpr sm_selector sm_100 = sm_selector::sm_100;
+constexpr sm_selector sm_103 = sm_selector::sm_103;
+constexpr sm_selector sm_107 = sm_selector::sm_107;
+constexpr sm_selector sm_110 = sm_selector::sm_110;
+constexpr sm_selector sm_120 = sm_selector::sm_120;
+constexpr sm_selector sm_121 = sm_selector::sm_121;
+
+using detail::is_exactly;
+using detail::provides;
+} // namespace target
+} // namespace nv
+
+#endif // C++  && !defined(__CUDACC_RTC__)
+
+#include <nv/detail/__target_macros>
+
+#endif // __NV_TARGET_H