Faster eval #298
Open
Faster eval #298
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
Open
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Outdated suggestions cannot be applied.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.
We might not want to merge this because it is hacky and there might be a usage that I don't foresee that could be impacted.
Problem:
Somewhere in llm-foundry or composer, the batches are appended with many stop tokens which makes the eval very slow.
Solution:
In
OpenLMforCausalLMtests for 0s at the end of the input_ids batch and remove them. Add back fake logits at the end....Consideration:
This hack could be implemented in
SimpleComposerOpenLMCausalLMinstead and would have less risk to have unforeseen impacts.There shouldn't really be a need for this but I could not understand what was happening in Composer or llm-foundry or find where/why the inputs are appended with 0s.