Default to only applying QK norm over the head dimension#1367
Open
christianazinn wants to merge 1 commit intoEleutherAI:mainfrom
Open
Default to only applying QK norm over the head dimension#1367christianazinn wants to merge 1 commit intoEleutherAI:mainfrom
christianazinn wants to merge 1 commit intoEleutherAI:mainfrom
Conversation
|
|
Member
|
@christianazinn why is this a better default than the current defualt? |
Author
I had assumed that most models used this normalization style, but it seems that's not the case (e.g. Olmo2). Perhaps it ought to be changed to a toggleable option that is enforced for GQA. |
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
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Before this change, the default application of QK norm was to apply it to dimensions
[*, N, H]. This caused problems with GQA, in which the K tensor is of shape[T, B, N_kv, H], which doesn't match. This PR fixes that by changing QK norm to normalize over[*, H]by default, as in e.g. Gemma3 and the original QK norm paper, with the option to go back to using[*, N, H]available asqk_layernorm_over_heads.