Skip to content

fix prepare_cos_sin invoke in RotaryEmbedding#1035

Merged
michalkuligowski merged 4 commits intohabana_mainfrom
dev/yu/rotary_emb_fix
Apr 17, 2025
Merged

fix prepare_cos_sin invoke in RotaryEmbedding#1035
michalkuligowski merged 4 commits intohabana_mainfrom
dev/yu/rotary_emb_fix

Conversation

@zhouyu5
Copy link
Copy Markdown

@zhouyu5 zhouyu5 commented Apr 9, 2025

Previously, we use if hasattr(self, "scaling_factors") to determine whether call prepare_cos_sin, however, some model's scaling factor in rope_scaling has a slightly different name, which is scaling_factor (note: no ending 's'), in this case, it will failed to call prepare_cos_sin, which is not expected.

This PR will help with the following class:

  • Llama3RotaryEmbedding
  • DynamicNTKScalingRotaryEmbedding
  • YaRNScalingRotaryEmbedding

Signed-off-by: zhouyu5 <yu.zhou@intel.com>
@zhouyu5
Copy link
Copy Markdown
Author

zhouyu5 commented Apr 9, 2025

@michalkuligowski @madamczykhabana Could you have a review?

@michalkuligowski
Copy link
Copy Markdown

/run-gaudi-tests

@kzawora-intel
Copy link
Copy Markdown

/run-gaudi-tests

@michalkuligowski
Copy link
Copy Markdown

/run-gaudi-tests

@michalkuligowski michalkuligowski merged commit 997af3e into habana_main Apr 17, 2025
45 checks passed
@michalkuligowski michalkuligowski deleted the dev/yu/rotary_emb_fix branch April 17, 2025 12:16
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants