feat: flash attention v3 #479
Open
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
|
@cathalobrien great contribution. Please could you add some tiny docs or references in the configs to show how a user can control which attention is used? |
for more information, see https://pre-commit.ci
for more information, see https://pre-commit.ci
|
@mchantry |
|
TODO will merge both flash_attention wrappers into one wrapper to minimise config complexity |
Would be good to get output saying which version was chosen, for posterity. |
…e into feat/flash_attention_v3
done, good catch |
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Outdated suggestions cannot be applied.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.
Description
Adds a flash attention v3 option to models/layers/attention.py
flash attention v3 is optimised for hoppers and newer GPUs, with up to 2x faster attention kernels. For more info see here
you can use flash-attn v3 with the following config entry
model.processor.attention_implementation=flash_attention_v3I compared the loss (seed=42) for SDPA vs flash-attn v3 and it matched to the 7th decimal place
I also changed the error message reporting for the flash attention v2 wrapper, to actually print the error. This improves the user experience in the case that flash attn is installed but you need a newer gcc. Now anemoi will print the GLIBC error, rather then telling you to install flash attn when it's already installed
What problem does this change solve?
faster attention
As a contributor to the Anemoi framework, please ensure that your changes include unit tests, updates to any affected dependencies and documentation, and have been tested in a parallel setting (i.e., with multiple GPUs). As a reviewer, you are also responsible for verifying these aspects and requesting changes if they are not adequately addressed. For guidelines about those please refer to https://anemoi.readthedocs.io/en/latest/
By opening this pull request, I affirm that all authors agree to the Contributor License Agreement.