lmstudio-ai / lmstudio-ai/mlx-engine
[Feature Request] Support weightless RMSNorm (for FlashNorm weight folding trick)
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.2k
- Forks
- 133
- Avg merge
- 21h 6m
- Merged PRs (30d)
- 1
Description
Please add support for RMSNorm without normalization weights.
This is to support FlashNorm — a mathematically equivalent variant of RMSNorm that folds norm weights into the subsequent linear layer. See explainer video.
We have applied this weight folding trick to a few LLMs (Llama, Qwen, SMolLM) here:
https://huggingface.co/models?other=weightless-rmsnorm
Motivation
FlashNorm's removal of norm weights reduces inference overhead at zero accuracy cost, and we'd like to share these optimized models with the broader community.
Possible Implementation
Remove norm weights from your RMSNorm implementation. E.g., just skip norm weight multiplication if there are no norm weights provided.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by locating the Python RMSNorm implementation and the model code that constructs it. Add support for instances without normalization weights while preserving existing weighted behavior, then verify both forms through the repository's existing checks.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 58/100