feat: support flash attention 2 in qwen2 vl vision blocks #2721
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
| ) | ||
| attn_output = attn_output.transpose(0, 1) | ||
| # calc maximum sequence length for any batch | ||
| max_seqlen = torch.max(cu_seqlens[1:] - cu_seqlens[:-1]) |
There was a problem hiding this comment.
It may be interesting to compute this value once in the Qwen2VisionModel.forward instead.
There was a problem hiding this comment.
great point, updated in the latest commit to only calculate the max_seqlen once!
|
optimistically merging as the Watching for regressions and will revet if required! |
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Outdated suggestions cannot be applied.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 adds support for flash attention within the vision blocks of qwen2vl. This improves latency and reduces spikes in memory usage from sdpa attention (helps avoid oom'ing on large images).
running
on a single
L4's with a small model this has a ~5% perf improvement: prev:~67msper token current:~63msper token.ps** it's likely the performance improvement scales with the size/number of layers however I have not benchmarked any larger models to measure the difference.