Skip to content

Fix/model bugfixes - #1896

Merged
nastya236 merged 5 commits into
ml-explore:mainfrom
michalk8:fix/model-bugfixes
Sep 17, 2026
Merged

nastya236 merged 5 commits into
ml-explore:mainfrom
michalk8:fix/model-bugfixes

Conversation

@michalk8

Copy link
Copy Markdown
Collaborator

Some of the models have been incorrectly normalising q/k + on CUDA, some ops were returning wrong results if not contiguous (will submit MLX issues/PRs later)

  • ☑️ I understand it is strictly prohibited to use AI to write PR description
  • AI usage disclosure: bug finding

@michalk8 michalk8 self-assigned this Sep 16, 2026
@michalk8
michalk8 requested a review from zcbenz September 16, 2026 09:37
@michalk8 michalk8 added the bug label Sep 16, 2026
@michalk8
michalk8 requested a review from nastya236 September 16, 2026 12:33
Comment thread mlx_lm/models/bailing_moe_v3.py Outdated
Comment thread mlx_lm/models/switch_layers.py Outdated
@nastya236

Copy link
Copy Markdown
Collaborator

Amazing! Could I ask you please to link issues that this PR closes?

@michalk8

michalk8 commented Sep 16, 2026

Copy link
Copy Markdown
Collaborator Author

Amazing! Could I ask you please to link issues that this PR closes?

None open, but similar fixes have been scatter proposed in #1624 (the eps bug), #1431 (traditional rope), #768 (yarn stuff).
Most of the bugs fixed here were discovered when merging #1023 when fixing OLMo Hybrid.

@nastya236 nastya236 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Looks good to me!

@nastya236
nastya236 merged commit a63e24c into ml-explore:main Sep 17, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants