Repack also experts#210
Merged
Merged
Conversation
This was referenced Feb 19, 2025
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
When I implemented run time repacking, I required the tensor to be 2D to be eligible for repacking, I guess to simplify the code. But I forgot about MoE models, where expert weights are in 3D tensors.
This PR fixes that. This leads to very significant performance gains. E.g., for DeepSeek-Lite quantized with
IQ4_XS, we getPP-512 = 545t/s on the main branch, andPP-512 = 677 t/swith this PR when using run time repacking.