Added a CMake variable USER_CUDA_ARCH_LIST to allow users to specify CUDA architectures manually.

If this variable is not set, CMake will automatically detect the CUDA architecture of the underlying machine and build the dynamic library accordingly.

Set the TORCH_CUDA_ARCH_LIST environment variable to the desired architecture. CMake will automatically read this value from the environment.

for example:

dynamic library is built for sm_75 and sm_80:

export TORCH_CUDA_ARCH_LIST="7.5 80"
python3 setup.py develop 2>&1 | tee build.log

dynamic library is built for local GPU:

python3 setup.py develop 2>&1 | tee build.log

Should I update export TORCH_CUDA_ARCH_LIST="7.5 80" as export TORCH_CUDA_ARCH_LIST="7.5 8.0" to aviod error?

Added a CMake variable USER_CUDA_ARCH_LIST to allow users to specify CUDA architectures manually.
If this variable is not set, CMake will automatically detect the CUDA architecture of the underlying machine and build the dynamic library accordingly.
Set the TORCH_CUDA_ARCH_LIST environment variable to the desired architecture. CMake will automatically read this value from the environment.
for example:
dynamic library is built for sm_75 and sm_80:

export TORCH_CUDA_ARCH_LIST="7.5 80"
python3 setup.py develop 2>&1 | tee build.log

dynamic library is built for local GPU:

python3 setup.py develop 2>&1 | tee build.log

Should I update export TORCH_CUDA_ARCH_LIST="7.5 80" as export TORCH_CUDA_ARCH_LIST="7.5 8.0" to aviod error?

my typo. It should be "export TORCH_CUDA_ARCH_LIST="7.5 8.0" required by cuda_select_nvcc_arch_flags

Thanks for your help!