Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do ?
This PR adds the exllamav2 kernels into auto-gptq. The integration is similar to the exllama kernel. Here's a quick benchmark with llama2-7B with the integration in optimum/transformers. We get slower speedup compared to the benchmark in exllamav2 repo because here we only replace the Linear layers with the quantized layer with exllamav2 kernel. I have confirmed that the tests were successful and that we have the same perplexity as exllama kernel. For now, we only support GPTQ format and not the new EXL2 format.
TLDR: for bs=4, we have the same speed as the llama model from exllamav2 repo and 40% faster than exllama kernel.
For exllama kernel, we see that we are not compute bound for bs=1 and bs=2 and memory/overhead bound for bs=4.
For exllamav2 kernel, we see that we are not compute bound for bs=1 and bs=2 and bs=4.
Benchmark using exllama2 repo with their optimized llama model:
cc @PanQiWei @fxmarty