【源码编译错误】nvcc error : 'cudafe++' died due to signal 9 (Kill signal)

环境：

问题描述

我在linux中源码编译gpu版的fastdeploy，遇到如下错误：

[2025-09-16 19:22:33,441] [    INFO] spawn.py:38 - /usr/local/cuda/bin/nvcc -I/root/miniconda3/lib/python3.12/site-packages/paddle/include -I/root/miniconda3/lib/python3.12/site-packages/paddle/include/third_party -I/root/miniconda3/lib/python3.12/site-packages/paddle/include/paddle/phi/api/include/compat -I/root/miniconda3/lib/python3.12/site-packages/paddle/include/paddle/phi/api/include/compat/torch/csrc/api/include -I/usr/local/cuda/include -I/root/miniconda3/include/python3.12 -I/root/miniconda3/include/python3.12 -c /root/workspace/FastDeploy/custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/autogen/launch_gemm_kernel_block128x128x64_warp128x32x64_mma16x8x32_stage7.cu -o /root/workspace/FastDeploy/custom_ops/build/fastdeploy_ops/lib.linux-x86_64-cpython-312/launch_gemm_kernel_block128x128x64_warp128x32x64_mma16x8x32_stage7.cu.o -DPADDLE_WITH_CUDA -DEIGEN_USE_GPU -ccbin cc -Xcompiler -fPIC --expt-relaxed-constexpr -DNVCC -gencode arch=compute_89,code=sm_89 -gencode arch=compute_86,code=sm_86 -DPADDLE_DEV -DPADDLE_ON_INFERENCE -DPy_LIMITED_API=0x03090000 -Igpu_ops/cutlass_kernels -Ithird_party/cutlass/include -Ithird_party/cutlass/tools/util/include -Igpu_ops/fp8_gemm_with_cutlass -Igpu_ops -Ithird_party/nlohmann_json/include -DENABLE_SCALED_MM_C2X=1 -Igpu_ops/cutlass_kernels/w8a8 -DENABLE_BF16 -Igpu_ops/moe -DENABLE_FP8 -Igpu_ops/cutlass_kernels/fp8_gemm_fused/autogen -w -DPADDLE_WITH_CUSTOM_KERNEL -DPADDLE_EXTENSION_NAME=fastdeploy_ops -D_GLIBCXX_USE_CXX11_ABI=1 -std=c++17
nvcc error   : 'cudafe++' died due to signal 9 (Kill signal)

我的指令如下：

python3 -m pip install paddlepaddle-gpu==3.2.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/

build.sh 1 python false "[86,89]"

我想请问一下是否是内存不足的问题？如果是的话，应该如何设置呢？

一些尝试

1. 安装ccache无用
2. 按如下设置也无用

export MAX_JOBS=1
export OMP_NUM_THREADS=1

Start with build.sh and reproduce the failing nvcc command for the generated .cu file under custom_ops/gpu_ops/cutlass_kernels/fp8_gemm_fused/autogen. Check the build output and system diagnostics to determine why cudafe++ receives signal 9, then document a confirmed workaround or the missing build configuration.