remove fixed compute capabilities list

[wanzhenchn]

Removed fixed compute_capabilities = {80, 86, 89, 90} from setup.py when installing from source code.

It has been got by torch.cuda.get_device_capability() at runtime.

[casper-hansen]

@wanzhenchn Thank you for your contribution. Unfortunately, the setup.py will fail to build on some PCs without these compute capabilities specified, and GitHub workflows that build the pip package will also fail. See more in #13

As for the actual compute capabilities, {80, 86, 89, 90} is correct because 80 is the first version of Ampere and the AWQ CUDA kernel supports Ampere and later.

I will have to close this as it would cause errors when building.

[wanzhenchn]

@casper-hansen

When I installed it from source code on my single A100 GPUs with cuda 11.7, it only support compute_capabilities = 80, 86,
the 89 & 90 is not supported.

The issue is caused by a lower version of CUDA ?

If I removed the 89 & 90, everything is Ok.

[casper-hansen]

Unsupported gpu architecture

This error is most likely related to your installation because NVCC does not support compute capability 89 and 90 before CUDA 11.8