expose TORCH_CUDA_ARCH_LIST

[ErwannMillon]

No description provided.

[ErwannMillon]

make it easier to build SA in dockerfile

[woct0rdho]

I think if we set the cuda arch list, it should completely override the cuda archs detected from the GPUs rather than extend them, in case we need to do some cross compilation on a machine that has GPU. This is how TORCH_CUDA_ARCH_LIST behaves in PyTorch. I'm doing it like https://github.com/woct0rdho/SageAttention/blob/88e36fad70fd3ab40b8c8dfc0658c359bb3925ca/setup.py#L76

Also, we should split them by whitespace or semicolon, but not comma. See pytorch/audio#2050 (I'm not doing it yet but let's standardize it when making the PR)

[johnnynunez]

#160

[woct0rdho]

Hi @jt-zhang , as you've implemented a way to support TORCH_CUDA_ARCH_LIST in setup.py, I think all remaining PRs to support it can be closed.