The popular repo for quantizing and running LLaMA is the GPTQ-for-llama repo on github, which mostly copies from the GPTQ authors. The CUDA kernels are needed to support the specific kind of quantization that GPTQ does.
Problem is, while those CUDA kernels are great at short prompt lengths, they fall apart at long prompt lengths. You could see people complaining about this, seeing their inference speeds slowly tanking as their chats/prompts/etc got longer.
So off I went, spending the last week or so re-writing the kernels in Triton. I've now got my kernels running faster than the CUDA kernels at all sizes [0]. And I'm busily optimizing and fusing other areas. The latest MLP fusion kernels gave another couple percentage boost in performance.
Yet I still haven't actually played with LLaMA and made those agents I wanted... sigh And now I'm debating diving into the Triton source code, because they removed integer unpacking instructions during one of their recent rewrites. So I had to use a hack in my kernels which causes them to use more bandwidth than they otherwise should. Think of the performance they could have with those! ... (someone please stop me...)