use flash attn fuse cross entropy loss to reduce metric memory usage #2987
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
|
@cli99 consider doing it as in https://github.com/mosaicml/llm-foundry/pull/575/files, to avoid introducing flash attn as a composer dependency. |
j316chuck
reviewed
Feb 9, 2024
Comment on lines
+100
to
+104
| try: | ||
| from flash_attn.losses.cross_entropy import CrossEntropyLoss as FusedCrossEntropyLoss | ||
| self.loss_fn = FusedCrossEntropyLoss(ignore_index=ignore_index, reduction='sum') | ||
| except ImportError: | ||
| self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=ignore_index, reduction='sum') |
There was a problem hiding this comment.
Should this live in llm-foundry? CC: @dakinggg
There was a problem hiding this comment.
i have a very old pr that i never merged
|
@cli99 should we close or is this still WIP |
7 tasks
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Outdated suggestions cannot be applied.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.
This PR uses fused cross entropy loss from flash attention in the metric LanguageCrossEntropy (also LanguagePerplexity).
The current torch.nn.CrossEntropyLoss call needs
6 * seq_len * vocab_sizeGPU memory, and can be the bottleneck memory usage when sequence length is long (where act ckpt is probably used). Using cross entropy loss from flash attn resolves this problem.Example test model with long sequence and full act ckpt:
with torch loss fn:
with flash_attn loss fn
