Need help building Pytorch from source

torch.compile
21 
nvcc fatal   : Unsupported gpu architecture 'compute_35'

Remove this deprecated architecture from your build and specify the needed ones via TORCH_CUDA_ARCH_LIST.

22

How to remove compute_35 from the build?

23

You can specify compute capability using:

TORCH_CUDA_ARCH_LIST="8.6" python setup.py install

Please change it to your desired version.

24

I tried with 8.6, 9.0, 12.2:

TORCH_CUDA_ARCH_LIST=“8.6” python setup.py install
TORCH_CUDA_ARCH_LIST=“9.0” python setup.py install
TORCH_CUDA_ARCH_LIST=“12.2” python setup.py install

All of them returned this error:

(base) notooth@localhost:pytorch$ TORCH_CUDA_ARCH_LIST="8.6" python setup.py install
Building wheel torch-2.1.0a0+gitfe1f26a
-- Building version 2.1.0a0+gitfe1f26a
cmake --build . --target install --config Release
[12/2618] Performing build step for 'nccl_external'
FAILED: nccl_external-prefix/src/nccl_external-stamp/nccl_external-build nccl/lib/libnccl_static.a /mnt/Archive/Downloads/1/pytorch/build/nccl_external-prefix/src/nccl_external-stamp/nccl_external-build /mnt/Archive/Downloads/1/pytorch/build/nccl/lib/libnccl_static.a 
cd /mnt/Archive/Downloads/1/pytorch/third_party/nccl/nccl && make -j12 -l12 CXX=/usr/bin/c++ CUDA_HOME=/usr/local/cuda-12.2 NVCC=/usr/local/cuda-12.2/bin/nvcc "NVCC_GENCODE=-gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_89,code=sm_89 -gencode=arch=compute_90,code=sm_90" BUILDDIR=/mnt/Archive/Downloads/1/pytorch/build/nccl VERBOSE=0 DEBUG=0 && /home/notooth/miniconda3/bin/cmake -E touch /mnt/Archive/Downloads/1/pytorch/build/nccl_external-prefix/src/nccl_external-stamp/nccl_external-build
make -C src build BUILDDIR=/mnt/Archive/Downloads/1/pytorch/build/nccl
make[1]: Entering directory '/mnt/Archive/Downloads/1/pytorch/third_party/nccl/nccl/src'
NVCC_GENCODE is -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_89,code=sm_89 -gencode=arch=compute_90,code=sm_90
Compiling  transport.cc                        > /mnt/Archive/Downloads/1/pytorch/build/nccl/obj/transport.o
make[2]: Entering directory '/mnt/Archive/Downloads/1/pytorch/third_party/nccl/nccl/src/collectives/device'
NVCC_GENCODE is -gencode=arch=compute_35,code=sm_35 -gencode=arch=compute_50,code=sm_50 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_86,code=sm_86 -gencode=arch=compute_89,code=sm_89 -gencode=arch=compute_90,code=sm_90
nvcc fatal   : Unsupported gpu architecture 'compute_35'
make[2]: *** [Makefile:53: /mnt/Archive/Downloads/1/pytorch/build/nccl/obj/collectives/device/sendrecv.dep] Error 1
make[2]: Leaving directory '/mnt/Archive/Downloads/1/pytorch/third_party/nccl/nccl/src/collectives/device'
make[1]: *** [Makefile:58: /mnt/Archive/Downloads/1/pytorch/build/nccl/obj/collectives/device/colldevice.a] Error 2
make[1]: Leaving directory '/mnt/Archive/Downloads/1/pytorch/third_party/nccl/nccl/src'
make: *** [Makefile:25: src.build] Error 2
[37/2618] Building CXX object third_party/ideep/mkl-dnn/src/cpu/x64/CMakeFiles/dnnl_cpu_x64.dir/gemm/s8x8s32/jit_avx512_core_u8_copy_at_kern_autogen.cpp.o
ninja: build stopped: subcommand failed.
25

Have you tried cleaning the previous build and then rebuilding it? If not, please do that.

26

I successfully built Pytorch from source, but I got an error when running it:

RuntimeError: The NVIDIA driver on your system is too old (found version 11040). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx Alternatively, go to: https://pytorch.org to install a PyTorch version that has been compiled with your version of the CUDA driver.
27

Looks like you need to update your CUDA driver. The link is provided in the error message.

28

I currently have NVIDIA driver 470.199.02 and CUDA toolkit 12.2 on my system. It is the latest driver for Quadro K600.