AMD ROCm support

[IlyasMoutawwakil]

• Adapt setup.py
• Adapt build.yaml workflow

[casper-hansen]

I have no AMD ROCm GPUs and no way of testing this. Is it correctly understood that it's exclusively ExLlama kernels that will work with AMD GPUs on ROCm 5 (but not 6)?

[IlyasMoutawwakil]

I tested on ROCm 5.6.1
@seungrokj do these changes build on both 5 and 6 without impacting performance ?
If so I can add them in this PR.

[seungrokj]

I tested on ROCm 5.6.1 @seungrokj do these changes build on both 5 and 6 without impacting performance ? If so I can add them in this PR.

@IlyasMoutawwakil I just checked AutoGPTQ/AutoGPTQ#515 in rocm5.7 env. It is compiled and there should be no perf issue because they are basically same syntaxes. (Somehow the old syntax is not compatible with rocm6.0 though)

[IlyasMoutawwakil]

@casper-hansen for the other part of your question, yes for now Exllama is going to be the only way to run AWQ model on AMD GPUs. Untill we get a hipifiable, ROCm native or performant triton GEMM.