Hosang comments

Repositories
Issues
Comments

Results 3 comments of


                                            Hosang

[ROCm] Fix kernel cache miss in Triton FA

Closing this PR since the changes have already been implemented and merged via https://github.com/vllm-project/vllm/pull/18226/files

[ROCm] Enable custom paged attention kernel for Navi3/4

Hi @tlrmchlsmth thanks for letting me know about new V1 kernel. As you mentioned, there seem to be significant performance improvements in V1 due to the new prefill/decode method. As...

[ROCm] Enable custom paged attention kernel for Navi3/4

Closing this PR in favor of new v1 support: https://github.com/vllm-project/vllm/pull/17004