feat: CCCL CachingDeviceAllocator preload — 完整依赖链 288 files

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

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