fix: correct attention temperature scaling formula for Mistral3/Devstral #1
Conversation
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
Fix off-by-one error in the attention temperature scaling formula used by Mistral3 and Devstral models. The formula was computing: log(floor((pos + 1) / max_position_embeddings) + 1) * beta + 1 But the correct formula (matching the PyTorch reference implementation) is: 1 + beta * log(1 + floor(pos / max_position_embeddings)) This caused divergence from the reference implementation for positions >= original_max_position_embeddings - 1 (e.g., position 8191 for models with original_max_position_embeddings=8192). Fixes: ggml-org#17980 Co-authored-by: openhands <[email protected]>
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Outdated suggestions cannot be applied.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.
Summary
This PR fixes an off-by-one error in the attention temperature scaling formula used by Mistral3 and Devstral models.
Problem
The attention temperature scaling formula in
llama-graph.cppwas computing:But the correct formula (matching the PyTorch reference implementation) is:
Impact
This caused divergence from the reference implementation for positions >=
original_max_position_embeddings - 1. For example, withoriginal_max_position_embeddings=8192:Fix
Changed the formula to match the PyTorch reference implementation exactly.
Related Issue
Fixes: ggml-org#17980
@nlasky2000-dot can click here to continue refining the PR