feat: CCCL CachingDeviceAllocator LD_PRELOAD — bypass CoreX expandable_segments ASSERT

从 CCCL upstream cub/cub/util_allocator.cuh 提取 CachingDeviceAllocator
核心算法，去掉所有 CUB/CCCL 宏依赖，编译为独立 .so。

用 LD_PRELOAD 拦截 cudaMalloc/cudaFree，路由到 CUB 的 geometric-bin
缓存分配器。同时在 constructor 中 strip PYTORCH_CUDA_ALLOC_CONF 里的
expandable_segments 配置，避免 CoreX CUDACachingAllocator.cpp:545 ASSERT。

BI-V100 调优参数:
  bin_growth=8, min_bin=3 (512B), max_bin=13 (~550MB)
  max_cached_bytes=4GB per device (32GB卡的合理上限)

真机测试步骤:
  1. bash build_cccl_preload.sh
  2. LD_PRELOAD=./libcccl_allocator.so CCCL_ALLOC_DEBUG=1 \
     PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
     python3 verify_preload.py

qwen3_6_scripts/cccl_preload/build_cccl_preload.sh

@@ -1,43 +1,41 @@
 #!/usr/bin/env bash
-# Build libcccl_allocator.so
-#
-# Full CCCL dependency chain (288 headers) in ./include/
-# Source: cccl_upstream/cub/cub/util_allocator.cuh + transitive deps
+# Build libcccl_allocator.so — LD_PRELOAD .so for CUB CachingDeviceAllocator
 #
 # 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"
 
-[[ -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 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
 
-# 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)
+# Find CUDA include path
 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}"
@@ -45,7 +43,7 @@ for candidate in \
     fi
 done
 
-# Find CUDA libs
+# Find CUDA lib path for linking
 CUDA_LIB=""
 for candidate in \
     /usr/local/corex/lib64 \
@@ -57,43 +55,57 @@ 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] 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
-)
+echo "[build] CUDA lib: ${CUDA_LIB:-system}"
+echo "[build] Source: ${SRC}"
+echo "[build] Output: ${OUT}"
 
+# Build as shared library
+# -x cuda or -x c++ depending on compiler
 if [[ "${CXX}" == *clang++* ]]; then
-    "${CXX}" "${COMMON_FLAGS[@]}" -x c++ -o "${OUT}" "${SRC}" 2>&1
+    # 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}"
 else
-    "${CXX}" "${COMMON_FLAGS[@]}" -x c++ -o "${OUT}" "${SRC}" 2>&1
+    # 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}"
 fi
 
 if [[ -f "${OUT}" ]]; then
-    SIZE=$(stat -c%s "${OUT}" 2>/dev/null || echo "?")
-    echo ""
+    SIZE=$(stat -c%s "${OUT}" 2>/dev/null || stat -f%z "${OUT}" 2>/dev/null || echo "?")
     echo "[build] SUCCESS: ${OUT} (${SIZE} bytes)"
     echo ""
-    echo "Test:"
-    echo "  LD_PRELOAD=${OUT} CCCL_ALLOC_DEBUG=1 \\"
-    echo "  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \\"
-    echo "  python3 verify_preload.py"
+    echo "Usage:"
+    echo "  LD_PRELOAD=${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"
 else
     echo "[build] FAILED"
     exit 1