perf: use 1x4 warp layout for small query length#322
Merged
Conversation
yzh119
changed the title
yzh119
added a commit
that referenced
this pull request
Jun 21, 2024
🤖 I have created a release *beep* *boop* --- ## [0.0.6](v0.0.5...v0.0.6) (2024-06-21) ### Performance Improvements * use 1x4 warp layout for small query length ([#322](#322)) ([4e89b4d](4e89b4d)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: Zihao Ye <expye@outlook.com>
Merged
yzh119
added a commit
that referenced
this pull request
Jun 21, 2024
yzh119
added a commit
that referenced
this pull request
Jun 21, 2024
Disable #322 for v0.0.6 release because binary size is too large. v0.0.6 will only include bugfix at the moment.
yzh119
added a commit
that referenced
this pull request
Jul 4, 2024
Changes: 1. Prefetch page indices (we have already done such optimization on decode kernels, but not on append/prefill kernels which was used in GQA). 2. Unlock 1x4 warp layout in #322, we didn't enable this because the binary size is too large, we should further reduce some unnecessary template arguments. 3. Optimize `threadblock_sync_mdo_states` for efficient merging attention states of multiple warps in a threadblock. Our previous implementation assumes small shared memory size and interleaves shared memory reads/writes with computations, which is not as efficient as a bulk shared memory access. After this PR, the GQA kernel execution time (on H100) for setting `batch_size=128, seq_len=1024, num_qo_heads=32, num_kv_heads=4, head_dim=128` was improved from 133us to 103us.
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.
Duplicate of #304 and #185, just rebased on main.
This PR can accelerate GQA, we will release v0.0.6 after this PR gets merged (ETA: tonight).