fix: honor YaRN scaling for Qwen3.5 MRoPE - #1582
Merged
Merged
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Qwen3.5-27B configurations containing both
mrope_sectionandrope_type: yarncurrently take the default rotary path, silently applying linear scaling instead of YaRN. Respect the explicit RoPE scaling type while retaining the multimodal position layout.Compute YaRN frequencies using
head_dim * partial_rotary_factor, store the half-width cache consumed by the rotary kernels, and apply magnitude scaling in FP32 before casting to the model dtype.