Unpadding sequences #56
Closed
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
Contributor
Author
|
To switch between flash attention and sdpa, we need to change _attn_implementation = "flash_attention_2",inside of |
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Outdated suggestions cannot be applied.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.
This PR adds sequence unpadding and repadding with flash attention 2 varlen for the attention and flash attention rope. Inside of the model,
attention_maskis created and sequences are unpadded before passing to decoder layers. After all the decoder layers and final norm, the unpadded sequences are padded again before passing tolm_head. We repad for the loss function calculation. In the future, we can get rid of repadding and do loss calculation ourselves as_unpad_modernbert_inputalso returns unpadded labels.Results for sdpa (green) vs. FA2 (red) on wandb: https://wandb.ai/local-research-group/smollm2-135m-training-avelina/workspace?nw=nwusergalopyz. Grey one uses FA2 with batch size of 8 instead of 4. Because there are no padding tokens, I could fit 8bs inside of RTX 3090. (24GB RAM)