# microsoft/mup

maximal update parametrization (µP)

Repository: https://github.com/microsoft/mup
Canonical: https://ross.abutalabs.com/products/mup
Homepage: https://arxiv.org/abs/2203.03466
Language: Jupyter Notebook
License: MIT
License Family: permissive
Topics: python, pytorch, transformers, machine-learning, deep-learning, mutransfer, mup
Last push: 2024-07-17T11:54:02+00:00

## Health v2 (maintenance only)
Score: 23/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 0, release rhythm 8, longevity 100
- inputs: {"age_days": 1765, "days_push": 777, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1753, forks 105 (observed 2026-08-28T04:05:31.892729+00:00)

## What it is
The `mup` Python package implements Maximal Update Parametrization (μP) for PyTorch models, enabling optimal hyperparameters to remain stable across model widths. It supports μTransfer: tuning hyperparameters on a small proxy model and zero-shot transferring them to much larger models.

## Use cases
- tune hyperparameters on a small model and transfer them to a large transformer
- implement maximal update parametrization in PyTorch models
- reduce hyperparameter tuning cost for large neural networks
- run coord checks to verify parametrization correctness
- scale up MLPs or transformers without retuning learning rates

## When to choose
- you are pretraining large transformers or other wide neural networks and want cheap hyperparameter tuning
- you want to reduce fragility when scaling models from exploration to production size
- you use PyTorch and can modify your model's readout layer and optimizers

## When to avoid
- you train small models where hyperparameter tuning is already cheap
- you cannot modify your model architecture to use μP layers
- you use frameworks other than PyTorch

## Facets
- artifact type: library
- maturity: stable
- function: machine-learning, deep-learning, llm-training
- domain: machine-learning, deep-learning, large-language-models
- platform: python
- tags: pytorch, mup, mutransfer, hyperparameter-tuning, transformers, parametrization

## Member repositories
- microsoft/mup (main) score 23

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:31.892729+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:28:13.977801+00:00, confidence not recorded.
  - readme: https://github.com/microsoft/mup (fetched 2026-08-28T04:05:31.892729+00:00, sha 25ddd2bb1d96)
  - homepage: https://arxiv.org/abs/2203.03466 (fetched 2026-08-29T11:06:21.937377+00:00, sha c9eae2e78bc3)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T11:06:21.947084+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T11:06:21.950646+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T11:06:21.952544+00:00, sha a1f16f915a9a)
  - registry_pypi: https://pypi.org/pypi/mup/json (fetched 2026-08-29T11:06:21.954320+00:00, sha 18e1db92f4d5)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T11:06:21.949019+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.