2a7ca101d7
Sparse-checkout from NVIDIA/cccl main branch to complete cccl_upstream: Added: - python/cuda_cccl/ (226 files) — Python bindings for device-level algorithms Critical for muh toolchain: cuda.compute.reduce_into, scan, radix_sort, etc. Includes 204 .py files with full test coverage for all 27 algorithms - ci/ (163 files) — Build/test infrastructure build_cub.sh, test_cub.sh, build_and_test_targets.sh, matrix.yaml Directly maps to our [INFRA-CI] and [INFRA-BUILD] items - .agent/skills/ (7 files) — NVIDIA's own agent skills for CCCL cccl-style/SKILL.md, cccl-test/SKILL.md, sass-diff/SKILL.md - docs/ (491 files) — Official CCCL documentation CI references, CMake guides, Python compute docs, libcudacxx PTX docs - test/ (12 files) — Top-level integration tests (cuda_smoke, stdpar) - Root configs: .clang-format, .clang-tidy, CONTRIBUTING.md, pyproject.toml - CLAUDE.md symlink → AGENTS.md (NVIDIA's standard) cccl_upstream now mirrors full NVIDIA/cccl structure: Before: 42M (cub + thrust + libcudacxx + cudax + c + examples + benchmarks) After: 53M (+python +ci +docs +.agent +test +configs) This completes the CCCL base needed for: - [muh-bench] items: ci/util/build_and_test_targets.sh for targeted builds - [CCCL-verify] items: python/cuda_cccl/tests/ as reference implementations - [CCCL-test] items: ci/test_cub.sh, ci/test_thrust.sh - Agent workflow: .agent/skills/ for consistent style and test patterns
76 lines
3.6 KiB
Bash
Executable File
76 lines
3.6 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
|
|
set -euo pipefail
|
|
|
|
ci_dir="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
|
repo_root="$(cd "$ci_dir/.." && pwd)"
|
|
source "$ci_dir/pyenv_helper.sh"
|
|
|
|
# Parse common arguments
|
|
source "$ci_dir/util/python/common_arg_parser.sh"
|
|
parse_python_args "$@"
|
|
require_py_version "Usage: $0 -py-version <python_version>"
|
|
|
|
# Pin cuda-toolkit to the container's CTK minor and set cuda_version /
|
|
# cuda_major_version (-ctk-mode latest opts out). See pyenv_helper.sh.
|
|
pin_cuda_toolkit "${ctk_mode}"
|
|
|
|
# Setup Python environment
|
|
setup_python_env "${py_version}"
|
|
|
|
# Fetch or build the cuda_cccl wheel:
|
|
if [[ -n "${GITHUB_ACTIONS:-}" ]]; then
|
|
wheel_artifact_name=$("$ci_dir/util/workflow/get_wheel_artifact_name.sh")
|
|
"$ci_dir/util/artifacts/download.sh" "${wheel_artifact_name}" "${repo_root}/"
|
|
wheelhouse_dir="${repo_root}/wheelhouse"
|
|
else
|
|
"$ci_dir/build_cuda_cccl_python.sh" -py-version "${py_version}"
|
|
wheelhouse_dir="${repo_root}/wheelhouse"
|
|
fi
|
|
|
|
# Install cuda_cccl with the minimal CUDA extra. This intentionally avoids the
|
|
# full cu*/sysctk* extras because those pull in numba/numba-cuda. The flavor is
|
|
# "cu" (pip toolkit) or "sysctk" (system toolkit) per the -ctk-mode arg.
|
|
CUDA_CCCL_WHEEL_PATH="$(ls "${wheelhouse_dir}"/cuda_cccl-*.whl)"
|
|
ctk_flavor="$(ctk_extra_flavor "${ctk_mode}")"
|
|
python -m pip install "${CUDA_CCCL_WHEEL_PATH}[minimal-${ctk_flavor}${cuda_major_version}]"
|
|
python -m pip install pytest pytest-xdist
|
|
|
|
cd "${repo_root}/python/cuda_cccl/tests/"
|
|
python -m pytest -n 6 -v compute/test_no_numba.py
|
|
if [[ "${py_version}" == "3.14t" ]]; then
|
|
# Select only tests that support the minimal extra so pytest does not collect
|
|
# tests that import numba-cuda and re-enable the GIL. These tests provide their
|
|
# own worker threads, so keep pytest itself in a single process.
|
|
# The serialization node-ids are module-skipped on the v2 backend today and
|
|
# will start running there automatically once v2 gains serialization support.
|
|
python -m pytest -n 0 -v \
|
|
compute/test_free_threading_stress.py \
|
|
compute/test_multi_cc_serialization.py::test_aot_build_result_load_failure_is_shared_and_retryable \
|
|
compute/test_multi_cc_serialization.py::test_aot_serialization_waits_for_canonical_first_load
|
|
|
|
# Broad thread-safety sweep (pytest-run-parallel): re-run the numba-free
|
|
# functional suite with each test executed concurrently across threads
|
|
# (barrier-synchronized start), stressing the process-wide build cache,
|
|
# single-flight coordination, and the Cython bindings from many threads at
|
|
# once. Complements test_free_threading_stress.py above, which targets specific
|
|
# shared-object scenarios by hand. -n 0 so the threads share one interpreter.
|
|
#
|
|
# --parallel-threads=2 matches CuPy's free-threading CI (the closest GPU
|
|
# precedent); a small fixed count bounds GPU-memory pressure from concurrent
|
|
# kernels and stays reproducible across runners, unlike =auto (the runner's
|
|
# logical-core count).
|
|
#
|
|
# pytest-run-parallel is only used by this sweep, so install it on the 3.14t
|
|
# path rather than for every minimal (e.g. non-free-threaded 3.14) run.
|
|
python -m pip install pytest-run-parallel
|
|
|
|
# Fail fast if the interpreter is not actually GIL-free (wrong build /
|
|
# PYTHON_GIL=1): pytest-run-parallel does NOT catch a GIL that is enabled from
|
|
# the start -- it would run threads GIL-serialized and pass vacuously. (A GIL
|
|
# *re-enabled mid-run* by a non-free-threaded import IS caught by the plugin,
|
|
# which is why we do not pass --ignore-gil-enabled.)
|
|
python -c "import sys; assert not sys._is_gil_enabled(), 'GIL is enabled; parallel sweep has no signal'"
|
|
python -m pytest -n 0 -v --parallel-threads=2 compute/test_no_numba.py
|
|
fi
|