Skip to content

Tolerate a null rope_scaling config when constructing PhiMoE models - #1889

Open
gyanu2507 wants to merge 1 commit into
ml-explore:mainfrom
gyanu2507:fix/1884-phimoe-rope-scaling-none
Open

gyanu2507 wants to merge 1 commit into
ml-explore:mainfrom
gyanu2507:fix/1884-phimoe-rope-scaling-none

Conversation

@gyanu2507

@gyanu2507 gyanu2507 commented Sep 15, 2026

Copy link
Copy Markdown
Contributor

Fixes #1884

Phi-mini-MoE-instruct ships "rope_scaling": null, and ModelArgs.rope_scaling already defaults to None. Attention still did args.rope_scaling["short_factor"] and raised TypeError before any weights loaded.

SuScaledRoPE already defaults the factors to 1.0, so a missing or null config just uses those defaults.

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.

phimoe.py subscripts rope_scaling unconditionally, so Phi-mini-MoE-instruct dies with a bare TypeError

1 participant