search

triton-inference-server / fastertransformer_backend

BSD 3-Clause "New" or "Revised" License
411 stars 133 forks source link

When I compile ft_backend based on cuda10.2 report nvcc fatal : Unsupported gpu architecture 'compute_80' #109

Closed nercoeus closed 1 year ago

nercoeus commented 1 year ago

Description

branch: release/v1.2.1_tag
GPU: T4

Reproduced Steps

git clone https://github.com/triton-inference-server/fastertransformer_backend.git -b release/v1.2.1_tag \
    && cd fastertransformer_backend && mkdir build && cd build \
    && cmake -DTRITON_BUILD_CUDA_VERSION="10.2" -DTRITON_BUILD_CUDA_HOME="/usr/local/cuda/" -DTRITON_BUILD_CUDNN_HOME="/usr/local/cuda/" -DTRITON_ENABLE_GPU=ON -DCMAKE_C_COMPILER=/usr/lib64/openmpi/bin/mpicc -DCMAKE_CXX_COMPILER=/usr/lib64/openmpi/bin/mpicxx -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install -DTRITON_COMMON_REPO_TAG=r21.10 -DTRITON_CORE_REPO_TAG=r21.10 -DTRITON_BACKEND_REPO_TAG=r21.10 .. \
    && make -j 10 install
byshiue commented 1 year ago

Can you share the full log?

nercoeus commented 1 year ago 
cmake -DTRITON_BUILD_CUDA_VERSION="10.2" -DTRITON_BUILD_CUDA_HOME="/usr/local/cuda/" -DTRITON_BUILD_CUDNN_HOME="/usr/local/cuda/" -DTRITON_ENABLE_GPU=OFF -DCMAKE_C_COMPILER=/usr/lib64/openmpi/bin/mpicc -DCMAKE_CXX_COMPILER=/usr/lib64/openmpi/bin/mpicxx -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install -DTRITON_COMMON_REPO_TAG=r21.10 -DTRITON_CORE_REPO_TAG=r21.10 -DTRITON_BACKEND_REPO_TAG=r21.10 ..
make install

err log: [ 15%] Linking CUDA device code CMakeFiles/tensor.dir/cmake_device_link.o nvcc fatal : Unsupported gpu architecture 'compute_80' make[2]: [_deps/repo-ft-build/3rdparty/trt_fused_multihead_attention/CMakeFiles/trt_fused_multi_head_attention.dir/build.make:972: _deps/repo-ft-build/3rdparty/trt_fused_multihead_attention/CMakeFiles/trt_fused_multi_head_attention.dir/qkvToContext.cu.o] Error 1 make[2]: Waiting for unfinished jobs.... nvcc fatal : Unsupported gpu architecture 'compute_80' make[2]: [_deps/repo-ft-build/src/fastertransformer/utils/CMakeFiles/tensor.dir/build.make:97: _deps/repo-ft-build/src/fastertransformer/utils/CMakeFiles/tensor.dir/cmake_device_link.o] Error 1 make[1]: [CMakeFiles/Makefile2:1809: _deps/repo-ft-build/src/fastertransformer/utils/CMakeFiles/tensor.dir/all] Error 2 make[1]: [CMakeFiles/Makefile2:1652: _deps/repo-ft-build/3rdparty/trt_fused_multihead_attention/CMakeFiles/trt_fused_multi_head_attention.dir/all] Error 2 make: [Makefile:136: all] Error 2