chore: bump mlx-swift-lm for the non-indexed weights fix (#118) - #122
Merged
Conversation
Picks up SharpAI/mlx-swift-lm#43 (e60ccfd), which stops loadWeights from sweeping .safetensors that the weight index does not list, and guards the Qwen35 RMSNorm shift on the weights actually looking zero-centered. Fixes #118. mlx-community/Qwen3.6-27B-OptiQ-4bit previously loaded without error and generated noise, because its out-of-index optiq/mtp.safetensors made sanitize add 1 to every norm weight. Verified against that model on this commit: the text path answers correctly with MTP disabled and enabled, and the vision path no longer fails with `Unhandled keys ["mtp"]`. No change for Qwen3.5-4B-4bit or Qwen3.6-35B-A3B-4bit. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Re-points the bump at 717d77f9, which now carries both upstream fixes: - #43: non-indexed .safetensors ignored + RMSNorm shift guard (fixes #118) - #44: Gemma 4 KV-shared layers + quantizable scaled projections (fixes #120) With #44 included, the audio and omni integration jobs pass locally; opencode reaches its npm-install step (already continue-on-error). dflash remains red until #121 is fixed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…-lm-118-fix # Conflicts: # mlx-swift-lm
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #118 by picking up SharpAI/mlx-swift-lm#43 (
e60ccfd).What the submodule fix does
loadWeightsswept the model directory recursively, so any.safetensorsin a subdirectory was merged into the model's weights.mlx-community/Qwen3.6-27B-OptiQ-4bitshipsoptiq/mtp.safetensors, which is not inmodel.safetensors.index.json. That madeQwen35.sanitizeconclude the checkpoint stores zero-centered norms and add 1 to every RMSNorm weight — corrupting every normalization. The model loaded without error and generated noise.Two changes upstream:
SWIFTLM_MTP_ENABLE.Verified on this commit
mlx-community/Qwen3.6-27B-OptiQ-4bit:'},-idthlaf�(aren系_1!''Tokyo'SWIFTLM_MTP_ENABLE=1'Tokyo'Unhandled keys ["mtp"]¹ then stops on that repo's missing
processor_config.json— a separate packaging gap.Log line confirming the filter:
[loadWeights] skipping optiq/mtp.safetensors: not listed in model.safetensors.index.jsonNo behaviour change for
mlx-community/Qwen3.5-4B-4bitormlx-community/Qwen3.6-35B-A3B-4bit.CI expectation
This does not turn CI green on its own. The four persistently red checks have separate causes, both now filed: #120 (Gemma 4 KV-shared layers, blocks the three
integration_matrixjobs) and #121 (DFlashrope_theta).🤖 Generated with Claude Code