# huggingface/nanotron

Minimalistic large language model 3D-parallelism training

Repository: https://github.com/huggingface/nanotron
Canonical: https://ross.abutalabs.com/products/nanotron
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-05-26T10:32:37+00:00

## Health v2 (maintenance only)
Score: 56/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 84, release rhythm 8, longevity 77
- inputs: {"age_days": 1087, "days_push": 99, "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 2800, forks 328 (observed 2026-08-28T04:07:22.551985+00:00)

## What it is
Nanotron is a minimalistic Python library from Hugging Face for pretraining large language models with 3D parallelism (data, tensor, and pipeline parallelism). It offers a simple, flexible API for training transformer models on custom datasets at scale.

## Use cases
- pretrain a Llama-style transformer from scratch
- train LLMs with data, tensor, and pipeline parallelism
- scale model training across multiple GPUs and nodes
- learn how large-scale LLM training works
- run distributed pretraining on custom datasets
- benchmark distributed training performance

## When to choose
- you need a lightweight, hackable codebase for LLM pretraining research
- you want to combine data, tensor, and pipeline parallelism in one tool
- you are learning distributed training techniques via the Ultrascale Playbook
- you want PyTorch-native training without heavyweight framework abstractions

## When to avoid
- you only need to fine-tune existing models rather than pretrain from scratch
- you want a turnkey training framework with broad model-type support
- your project requires features outside transformer pretraining
- you need long-term commercial support or extensive documentation

## Facets
- artifact type: library
- maturity: active
- function: llm-training, deep-learning, machine-learning
- domain: large-language-models, deep-learning, machine-learning, gpu-computing
- platform: python
- tags: 3d-parallelism, pretraining, transformers, distributed-training, huggingface, gpu, linux

## Member repositories
- huggingface/nanotron (main) score 56

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:22.551985+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 2026-08-30T08:14:36.064524+00:00, confidence not recorded.
  - readme: https://github.com/huggingface/nanotron (fetched 2026-08-28T04:07:22.551985+00:00, sha 4401e5fbc844)
- Data as of 2026-08-30T08:39:29.467469+00:00.