Add low precision attention API from torchao to TorchAoConfig#13285
Open
howardzhang-cv wants to merge 1 commit intohuggingface:mainfrom
Open
Add low precision attention API from torchao to TorchAoConfig#13285howardzhang-cv wants to merge 1 commit intohuggingface:mainfrom
howardzhang-cv wants to merge 1 commit intohuggingface:mainfrom
Conversation
howardzhang-cv
force-pushed
the
feature/fp8_attn_ao
branch
from
125a7a8 to
ce14403
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
What does this PR do?
Adds low precision attention API from TorchAO to diffusers by updating TorchAoConfig with attn_backend option.
Note: this will require torchao 0.17.0
Todo:
Results:
flux.1-dev with 2048x2048 image size
Wan2.1-14B-Diffusers with 1280x720 frame size and 81 frames
Note that these results use a naive scheme, where every layer is quantized. Doing so results in subpar results when doing many inference steps (50 for WAN2.1). Using a better scheme (such as skipping early and late layers, quantizing 36/40 total layers) results in 1.12x speedup with much better quality. The VBench benchmarks for that are below: