Move loss generating token counting to the dataloader #1632
+93
−15
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
mvpatel2000
reviewed
Nov 3, 2024
mvpatel2000
approved these changes
Nov 3, 2024
dakinggg
added a commit
that referenced
this pull request
Nov 4, 2024
dakinggg
added a commit
that referenced
this pull request
Nov 4, 2024
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Outdated suggestions cannot be applied.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 previous PR to count loss generating tokens caused a throughput regression because the token counting is on the critical path in the Composer trainer. This PR moves the token counting into the collator so that it can be hidden in the dataloader. It does this by counting tokens per row in the batch, using the existing token counting function (not the most efficient, but doesn't matter because this is not dataloader bottlenecked, and allows reusing the existing code).
This PR should not change token counts or loss, and will barely affect throughput for larger models, but have a significant impact for smaller models.
125m model pretraining throughput, loss, token count, before and after this PR:
llama finetune throughput, loss, token count, before and after this PR: