[WIP] Upgrade FlashAttention to 2.0

[ghunkins]

What does this PR do?

This PR updates the symlink to FlashAttention 2.0 released in v1.0.9 of flash-attention. All credit goes to Tridao for the implementation.

According to reported results, this can enable a 2-4x speed-up.

This PR is a WIP and still needs testing and documentation updates. Additionally, the heuristics for choosing the underlying implementation likely needs to be updated per this comment.

Before submitting

• [Y] Did you have fun?
  • Make sure you had fun coding 🙃
• [Y] Did you read the contributor guideline?
• [Y] Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
  • [Y] (Flash Attention 2 #795)
• [N] Did you make sure to update the docs?
• [N] Did you write any new necessary tests?
• [N] Did you update the changelog? (if needed)

PR review

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

[facebook-github-bot]

Hi @ghunkins!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks!

[ghunkins]

Testing using the diffusers library, but I receive the below error.

import torch
from diffusers import DiffusionPipeline
from xformers.ops import MemoryEfficientAttentionFlashAttentionOp

# load pipe
pipe = DiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
).to('cuda')

# enable xformers with flash attention
pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None)

# run the pipe
image = pipe(prompt="a cute cat").images[0]

ValueError: Operator `memory_efficient_attention` does not support inputs:
     query       : shape=(16, 256, 1, 160) (torch.float16)
     key         : shape=(16, 256, 1, 160) (torch.float16)
     value       : shape=(16, 256, 1, 160) (torch.float16)
     attn_bias   : <class 'NoneType'>
     p           : 0.0
`flshattF` is not supported because:
    max(query.shape[-1] != value.shape[-1]) > 128

[ghunkins]

Closing. I think someone with more context needs eyes on this.

[Birch-san]

Testing using the diffusers library, but I receive the below error.

`flshattF` is not supported because:
    max(query.shape[-1] != value.shape[-1]) > 128

@ghunkins this is probably because SDXL's VAE uses a mid-block whose self-attention is 512-dim (albeit just 1 head). this exceeds the maximum head dim that flash attn 2's forward pass supports (256). nevermind; I see that this is SD1.5, you've already disabled it on the VAE, and this problem happened on a 160-dim layer. that is weird; 160-dim was supposed to be supported. why is it saying the max dim is 128? is that assertion a vestige from flash attn 1 that needs updating, or is flash attn 2 not detected at all?

[ghunkins]

@Birch-san The xformers team is tackling this, you can see #795 for context! I haven't had time to test yet, but based on the conversation it sounds like the pre-release wheels already support FlashAttention 2.0 on certain GPUs.

pip install xformers --pre