rocm support

[ambyjkl]

With a few tweaks I was able to get the cuda code to compile under rocm, but there seems to be some issue with inference, causing the output to always be !!!!!!.

Proper rocm support would be great.

[Qubitium]

@ambyjkl Check the ParoQuant kernels in GPT-QModel, we have both a cuda and triton kernel and aims to be accuracy compliant with official. We have not tested the cuda kernel on rocm but the triton kernel should run on rocm without issue for sure as compatibility is assured by triton compiler.