Updating the Flash Attention version to fix cross entropy loss #812
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
Pulling the latest commits from main fork
Pulling from the main repo
Pulling from mosaicml/llm-foundry main
Merging from mosaic main
Pulling from mosaic main
Pulling from mosaic main.
dakinggg
reviewed
Dec 19, 2023
|
throughput, memory, and loss before and after: |
dakinggg
approved these changes
Dec 20, 2023
Done. |
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Outdated suggestions cannot be applied.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.
The cross entropy loss of Flash Attention v2.3.2 (and lower) throws illegal memory access error when used with large (device train microbatch size X sequence length X vocabulary). To fix this we had reverted back to FA v1's CE loss in this PR (#795). However, we discovered that for very large (device train microbatch size X sequence length X vocabulary), FA v1's CE loss runs into numerical precision errors, causing divergence.
The newer versions of Flash Attention (v2.3.3 and higher) seem to have solved both of these problems, and hence in this PR, we update the repo to start using FA v2.3.6 (the latest version) instead of FA v2.3.2.
The blue loss curve below corresponds to the training run using FA v2.3.6's CE loss, and the pink curve corresponds to training run using FA v1's CE loss.