Skip to content

Remove unused shared_expert_combination_strategy from cohere2_moe#46408

Open
ArthurZucker wants to merge 1 commit into
mainfrom
cohere2-moe-rm-sum
Open

Remove unused shared_expert_combination_strategy from cohere2_moe#46408
ArthurZucker wants to merge 1 commit into
mainfrom
cohere2-moe-rm-sum

Conversation

@ArthurZucker
Copy link
Copy Markdown
Collaborator

The sum branch of shared_expert_combination_strategy is dead: every published cohere2_moe checkpoint (the command-a-plus-05-2026 family + its derivatives) uses average. Drop the config field and always average the shared-expert output.

Existing configs that still carry the key load unchanged (it's ignored). Tests: tests/models/cohere2_moe pass (TP tests need multi-GPU). AI-assisted, human-reviewed.

All published cohere2_moe checkpoints (the command-a-plus-05-2026 family)
use the 'average' strategy; the 'sum' branch is dead. Drop the config
field and always average the shared-expert output. Existing configs that
still carry the key load unchanged (it is ignored).
@github-actions
Copy link
Copy Markdown
Contributor

github-actions Bot commented Jun 4, 2026

[For maintainers] Suggested jobs to run (before merge)

run-slow: cohere2_moe

@HuggingFaceDocBuilderDev
Copy link
Copy Markdown

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

@ArthurZucker ArthurZucker requested a review from vasqu June 4, 2026 13:24
Copy link
Copy Markdown
Contributor

@vasqu vasqu left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fair enough, if worst comes to worst we can readd this but I would only expect that on future models with only that change (unlikely)

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.

3 participants