Normal Attention fix to get good image quality without using flash attention #37
Conversation
This file contains 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
ayushtues
changed the title
This was referenced Jun 14, 2023
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Outdated suggestions cannot be applied.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.
Hi, I was playing around with this repo while helping to integrate it with HF Diffusers huggingface/diffusers#3492, and found that if we used normal attention
QKVAttentionLegacywe were not able to generate images of good quality, although using flash attentionQKVFlashAttentiongave the expected results.The reason for this turned out to be a slightly different ordering of the heads and keys in the flash attention and normal attention implementations. Fixing which allowed us to generate good images using normal attention, without flash attention ( which can be a pain to install ).
Sending this PR with the fix, would be good to integrate this since flash attention is hard to install and can't be installed in all devices.
The quantization might not be that important, its just that Flash attention works with fp16 only I think, so I did an explicit conversion in the normal attention to get exactly same results
TLDR : You can now generate good images from the pretrained checkpoints without using Flash attention