Fix OLMo 3 scaled RoPE handling for sliding attention#45945
Draft
nurpax wants to merge 2 commits into
Draft
Conversation
cee279b to
b9bae6f
Compare
b9bae6f to
ddd085f
Compare
Contributor
|
[For maintainers] Suggested jobs to run (before merge) run-slow: olmo3, olmo_hybrid |
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
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
sliding_attentionlayers and configured RoPE forfull_attentionlayers.Gemma2RotaryEmbeddingforward behavior, including returningcos.to(dtype=x.dtype), sin.to(dtype=x.dtype).Context
Olmo3Modeldocuments that RoPE scaling is not applied to sliding-window attention layers, but the current implementation computes one configured rotary embedding and passes it to every layer. OLMo 3 needs layer-type-specific RoPE: default RoPE for sliding attention and configured/scaled RoPE for full attention.Test plan
PYTHONPATH=src python utils/check_modular_conversion.py --files src/transformers/models/olmo3/modular_olmo3.py --num_workers 1ruff check src/transformers/models/olmo3/modular_olmo3.py src/transformers/models/olmo3/modeling_olmo3.py tests/models/olmo3/test_modeling_olmo3.pygit diff --checkPYTHONPATH=src pytest tests/models/olmo3/test_modeling_olmo3.py -k sliding_attention_uses_default_rope_with_scaled_config -qAI assistance was used to prepare this PR. I reviewed the generated diff and am responsible for the change.