Optimize NaViT with SDPA and vectorized forward pass #353
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
- Replace manual attention with F.scaled_dot_product_attention - Use repeat_interleave instead of meshgrid for position computation - Build image_ids efficiently with repeat_interleave instead of F.pad - Remove unused Rearrange import ~56% speedup (91ms -> 58ms on 512 variable-sized images) Numerically equivalent (max diff ~5e-4, within flash attention tolerance) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <[email protected]>
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Outdated suggestions cannot be applied.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.
Related to #347
Summary
F.scaled_dot_product_attentionrepeat_interleaveinstead ofmeshgridfor position computation (reducesarangecalls from 2052 to ~1026)image_idsefficiently withrepeat_interleaveinstead of incrementalF.padRearrangeimportBenchmark
Tested on 512 variable-sized images (16px tall, 32-80px wide), depth=12, ViT-tiny config:
and compared to ViT (padding for ViT)
Numerical equivalence
torch.allclose(original, optimized, rtol=1e-3, atol=1e-3)returnsTrueTest plan
group_images=True🤖 Generated with Claude Code