Exllama kernels

[IlyasMoutawwakil]

for casper-hansen/AutoAWQ#313

[casper-hansen]

I have a few questions about the ExLlama kernels.

1. Is it necessary to include v1 and v2, why not just v2 (seems they are better)?
2. Are these kernels updated with the latest commits from ExLlamaV2?

[IlyasMoutawwakil]

1. I think it's always best to have a fallback in case exllamav2 doesn't work or build, an example is exllamav2 not building on rocm6.0 https://github.com/AutoGPTQ/AutoGPTQ/pull/515/files
2. Not sure, the ones in the original repo contain a lot of unused components (e.g. moe specific stuff). I took the one in AutoGPTQ for now, which I believe to be minimal and fixed for AutoAWQ (bit overflow from removing and adding 1 to zeros).