decode bf16 smallm: support arbitrary 1<heads_per_group<=8 via direct Q/Y GMEM when TMA unsuitable#37
Open
Religious-J wants to merge 1 commit intoTencent:mainfrom
Open
Conversation
… Q/Y GMEM when TMA unsuitable
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
decode bf16 smallm support heads_per_group = 2,3,4,5,6,7,9
Main changes: in the decode stage, the amount of data involved in loading Q and storing Y is small and does not affect pipeline scheduling, direct memory access is used instead of TMA to support heads_per_group = 2, 3, 4, 5, 6, 7, 9.
Result
H20 GPU, cuda13.0
ps. hpc means splitk = false, hpc-splitk means splitk = true
sgl-flash_attn is flash_attn_v3
easy test: https://github.com/Religious-J/ops-beanchmark
Both can achieve optimal performance.