fix(build): use CoreX clang++ instead of nvcc — match working build scripts

qwen3_6_scripts/build_corex_batched_gemm.sh

@@ -1,99 +1,76 @@
-#!/bin/bash
+#!/usr/bin/env bash
 # Build corex_batched_gemm.so — CUTLASS batched GEMM pybind for MoE decode
 #
-# Run on BI-V100:
+# Verified: 2.462ms for 8-expert decode (issue #68)
-#   bash build_corex_batched_gemm.sh
 #
-# Output: corex_batched_gemm.so (deploy to vllm package dir)
+# Usage: bash build_corex_batched_gemm.sh VLLM_ROOT
-
+#   or:  bash build_corex_batched_gemm.sh  (outputs to prebuilt/)
 set -euo pipefail
 
-SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 PROJ_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
-EX_ENGINE="$PROJ_ROOT/ex_engine"
+COREX_ROOT=${COREX_ROOT:-/usr/local/corex-3.2.3}
+if [ ! -d "$COREX_ROOT" ]; then
+    COREX_ROOT=/usr/local/corex
+fi
+TORCH_ROOT=${TORCH_ROOT:-${COREX_ROOT}/lib64/python3/dist-packages/torch}
+if [ ! -d "$TORCH_ROOT" ]; then
+    TORCH_ROOT=$(python3 -c "import torch; import os; print(os.path.dirname(torch.__file__))" 2>/dev/null || echo "/usr/local/corex/lib/python3/dist-packages/torch")
+fi
+
+CUTLASS_INCLUDE="$COREX_ROOT/samples/cutlass/include"
+if [ ! -d "$CUTLASS_INCLUDE/cutlass" ]; then
+    CUTLASS_INCLUDE="$COREX_ROOT/include"
+fi
+
+# Output path
+if [ -n "${1:-}" ]; then
+    OUTPUT="${1}/corex_batched_gemm.so"
+else
+    OUTPUT="$SCRIPT_DIR/prebuilt/corex-3.2.3-ivcore10/corex_batched_gemm.so"
+fi
 
 # Source files
-BIND_CPP="$EX_ENGINE/xllm_kernels/cuda/bindings/corex_batched_gemm_bind.cpp"
+BIND_CPP="$PROJ_ROOT/ex_engine/xllm_kernels/cuda/bindings/corex_batched_gemm_bind.cpp"
-KERNEL_CU="$EX_ENGINE/xllm_kernels/cuda/corex_batched_gemm_kernel.cu"
+KERNEL_CU="$PROJ_ROOT/ex_engine/xllm_kernels/cuda/corex_batched_gemm_kernel.cu"
-
-# CUTLASS headers from cat_files (Iluvatar CoreX fork)
-CUTLASS_INCLUDE="/usr/local/corex/include"
-if [ ! -d "$CUTLASS_INCLUDE/cutlass" ]; then
-    # Fallback: check corex-samples
-    CUTLASS_INCLUDE="/usr/local/corex/samples/cutlass/include"
-fi
-
-# PyTorch/libtorch paths
-TORCH_DIR=$(python3 -c "import torch; print(torch.utils.cmake_prefix_path)" 2>/dev/null || echo "")
-TORCH_INCLUDE=$(python3 -c "import torch; print(torch.utils.cpp_extension.include_paths()[0])" 2>/dev/null || echo "/usr/local/corex/lib/python3/dist-packages/torch/include")
-TORCH_LIB=$(python3 -c "import torch; print(torch.utils.cpp_extension.library_paths()[0])" 2>/dev/null || echo "/usr/local/corex/lib/python3/dist-packages/torch/lib")
-PYTHON_INCLUDE=$(python3 -c "from sysconfig import get_path; print(get_path('include'))")
-CUDA_INCLUDE=$(python3 -c "import torch; print(torch.utils.cpp_extension.include_paths()[1])" 2>/dev/null || echo "/usr/local/corex/include")
-if [ ! -f "$CUDA_INCLUDE/cuda_runtime.h" ]; then
-    CUDA_INCLUDE="/usr/local/corex/include"
-fi
-if [ ! -f "$CUDA_INCLUDE/cuda_runtime.h" ]; then
-    CUDA_INCLUDE="$(dirname $(which nvcc 2>/dev/null || echo /usr/local/cuda/bin/nvcc))/../include"
-fi
 
+echo "[build] COREX_ROOT=$COREX_ROOT"
+echo "[build] TORCH_ROOT=$TORCH_ROOT"
 echo "[build] CUTLASS_INCLUDE=$CUTLASS_INCLUDE"
-echo "[build] CUDA_INCLUDE=$CUDA_INCLUDE"
+echo "[build] OUTPUT=$OUTPUT"
-echo "[build] TORCH_INCLUDE=$TORCH_INCLUDE"
-echo "[build] TORCH_LIB=$TORCH_LIB"
 
-BUILD_DIR="/tmp/build_corex_batched_gemm"
+"${COREX_ROOT}/bin/clang++" \
-mkdir -p "$BUILD_DIR"
+    -std=c++17 -O3 -shared -fPIC \
-OUT_SO="$SCRIPT_DIR/prebuilt/corex-3.2.3-ivcore10/corex_batched_gemm.so"
+    --cuda-path="${COREX_ROOT}" \
-
+    --cuda-gpu-arch=ivcore10 \
-# Step 1: Compile CUTLASS kernel .cu → .o
+    --no-cuda-version-check \
-echo "[build] compiling kernel..."
-nvcc -c "$KERNEL_CU" \
-    -o "$BUILD_DIR/kernel.o" \
-    -I "$CUTLASS_INCLUDE" \
-    -I "$TORCH_INCLUDE" \
-    -I "$TORCH_INCLUDE/torch/csrc/api/include" \
-    --gpu-architecture=ivcore10 \
-    -std=c++17 -O2 \
-    --expt-relaxed-constexpr \
-    -Xcompiler -fPIC
-
-# Step 2: Compile pybind .cpp → .o
-echo "[build] compiling pybind wrapper..."
-g++ -c "$BIND_CPP" \
-    -o "$BUILD_DIR/bind.o" \
-    -I "$TORCH_INCLUDE" \
-    -I "$TORCH_INCLUDE/torch/csrc/api/include" \
-    -I "$PYTHON_INCLUDE" \
-    -I "$CUDA_INCLUDE" \
-    -I "$CUTLASS_INCLUDE" \
-    -std=c++17 -O2 -fPIC \
     -D_GLIBCXX_USE_CXX11_ABI=0 \
-    -DTORCH_EXTENSION_NAME=corex_batched_gemm
+    -DTORCH_EXTENSION_NAME=corex_batched_gemm \
+    -DTORCH_API_INCLUDE_EXTENSION_H \
+    -I"${TORCH_ROOT}/include" \
+    -I"${TORCH_ROOT}/include/torch/csrc/api/include" \
+    -I"${TORCH_ROOT}/include/TH" \
+    -I"${TORCH_ROOT}/include/THC" \
+    -I"${CUTLASS_INCLUDE}" \
+    -I/usr/local/include/python3.10 \
+    "${KERNEL_CU}" "${BIND_CPP}" \
+    -L"${TORCH_ROOT}/lib" \
+    -L"${COREX_ROOT}/lib64" \
+    -Wl,-rpath,"${TORCH_ROOT}/lib" \
+    -Wl,-rpath,"${COREX_ROOT}/lib64" \
+    -ltorch_python -ltorch_cuda -ltorch_cpu -ltorch \
+    -lc10_cuda -lc10 -lcudart \
+    -o "${OUTPUT}"
 
-# Step 3: Link → .so
+echo "[build] ✓ built ${OUTPUT}"
-echo "[build] linking..."
+echo "[build] size: $(du -h "${OUTPUT}" | cut -f1)"
-g++ -shared \
-    "$BUILD_DIR/kernel.o" \
-    "$BUILD_DIR/bind.o" \
-    -o "$OUT_SO" \
-    -L "$TORCH_LIB" \
-    -ltorch -ltorch_cpu -ltorch_cuda -lc10 -lc10_cuda \
-    -L /usr/local/corex/lib64 -lcudart \
-    -Wl,-rpath,"$TORCH_LIB" \
-    -Wl,-rpath,/usr/local/corex/lib64
 
-echo "[build] ✓ built $OUT_SO"
-echo "[build] size: $(du -h "$OUT_SO" | cut -f1)"
-
-# Quick import test
 python3 -c "
-import torch
-torch.ops.load_library('$OUT_SO')
 import importlib.util
-spec = importlib.util.spec_from_file_location('corex_batched_gemm', '$OUT_SO')
+spec = importlib.util.spec_from_file_location('corex_batched_gemm', '${OUTPUT}')
 mod = importlib.util.module_from_spec(spec)
 spec.loader.exec_module(mod)
-print('[build] ✓ import OK, functions:', [x for x in dir(mod) if not x.startswith('_')])
+print('[build] ✓ import OK:', [x for x in dir(mod) if not x.startswith('_')])
-" 2>&1 || echo "[build] import test skipped (no GPU)"
+" 2>&1 || echo "[build] import test skipped"
 
 echo "[build] done"