Closed ZhaoyangLyu closed 2 years ago
I'm building with: Python 3.8.10, PyTorch version 1.10.0a0+git179b3ab, CUDA V11.0.221, gcc version 8.3.1
I can't reproduce that error on my end but you can try editing setup.py and change these lines:
extra_compile_args={
"nvcc": [
"-use_fast_math",
"-arch=sm_70",
"-std=c++17",
"-lineinfo",
]
}
to
extra_compile_args={
"nvcc": [
"-use_fast_math",
"-arch=sm_70",
"-std=c++17",
"-lineinfo",
],
"cxx": []
}
and see if that resolves it?
Thanks for your reply! I have switched to cuda 11.1, pytorch 1.10.1 and gcc 8.4.0. Now it works well for me.
May I know what pytorch version and cuda version are used to build the two CUDA PyTorch extensions? I was using pytorch 1.7.1, cuda 10.1 and gcc 5.5.0, but I got the error "torch/utils/cpp_extension.py", line 445, in unix_wrap_ninja_compile post_cflags = extra_postargs['cxx'] KeyError: 'cxx' when I tried to build the two CUDA PyTorch extensions. Any suggestions on how to solve the problem?