[WIP] MQA Logits Gluon Path Activation and New Flag#794
Draft
cagrikymk wants to merge 1 commit into
Draft
Conversation
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
This PR relies on Aiter commit: 09e21f3ee9c6243b15e5995353dd793092fb6f05 that adds a skip mechanism to leave the logits beyond start and end indices untouched. Also, it introduces Gluon kernel for fp8 mqa logits that is used for gfx950.
For Gluon kernel to work, Triton >=3.6 needed, while best perf. requires ToT, 3.6 is the minimum version.
I did various testing using DeepSeek-V4-Flash on MI355 with this PR, Triton 3.5.1 and Triton 3.6 (official release versions) and provided the data below.
TLDR:
When triton_kernels is pinned to 3.5.1, just upgrading to Triton 3.6 gives around 6-9% improvement. This PR also improves prefill performance (8k/1) by another 2%.
Not sure if there are other hard constraints forcing us to use 3.5.1 related to other models / architectures.
Baseline (only #791) with Triton/triton_kernels 3.5.1:
Accuracy:
C=64, 1K/1K, TP=8:
C=64, 8K/1, TP=8:
Baseline (only #791) with Triton 3.6.0 and triton_kernels 3.5.1:
Accuracy:
C=64, 1K/1K, TP=8:
C=64, 8K/1, TP=8:
This PR on top of #791 with Triton 3.6.0 and triton_kernels 3.5.1:
C=64, 1K/1K, TP=8:
C=64, 8K/1, TP=8: