# alpa-projects/alpa

Training and serving large-scale neural networks with auto parallelization.

Repository: https://github.com/alpa-projects/alpa
Canonical: https://ross.abutalabs.com/products/alpa
Homepage: https://alpa.ai
Language: Python
License: Apache-2.0
License Family: permissive
Topics: deep-learning, machine-learning, compiler, distributed-training, high-performance-computing, alpa, jax, distributed-computing, llm, auto-parallelization
Archived: true
Last push: 2023-12-09T16:26:31+00:00

## Health v2 (maintenance only)
Score: 10/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 0, release rhythm 8, longevity 100
- inputs: {"age_days": 2018, "days_push": 998, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: archived
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3178, forks 361 (observed 2026-08-28T04:07:47.729364+00:00)

## What it is
Alpa is a Python system for training and serving large-scale neural networks by automatically parallelizing single-device code across distributed clusters using data, operator, and pipeline parallelism. It is built on JAX, XLA, and Ray, and is now unmaintained as a research artifact with its core algorithm merged into XLA.

## Use cases
- train multi-billion parameter models on a distributed cluster
- automatically parallelize single-device JAX training code
- serve large language models like OPT-175B across multiple devices
- run pipeline and operator parallelism without manual sharding
- scale deep learning training linearly on clusters

## When to choose
- you need automatic parallelization for JAX-based large model training and can work with a research artifact
- you want to study or extend auto-sharding and auto-parallelization algorithms
- you need to serve very large transformer models with a Hugging Face-style interface

## When to avoid
- you need actively maintained software with bug fixes and support
- you use PyTorch or non-JAX frameworks
- you want production distributed training - use the auto-sharding code merged into XLA instead

## Facets
- artifact type: library
- maturity: abandoned
- function: machine-learning, llm-training, llm-inference, compiler
- domain: deep-learning, machine-learning, large-language-models, microservices
- platform: python, cloud
- tags: auto-parallelization, jax, distributed-training, model-parallelism, pipeline-parallelism, research-artifact, linux, gpu

## Member repositories
- alpa-projects/alpa (main) score 10

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:47.729364+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 2026-08-30T07:25:07.416313+00:00, confidence not recorded.
  - readme: https://github.com/alpa-projects/alpa (fetched 2026-08-28T04:07:47.729364+00:00, sha 6838f79e3d9a)
  - registry_pypi: https://pypi.org/pypi/alpa/json (fetched 2026-08-29T09:39:50.575476+00:00, sha 1c8061bc897b)
- Data as of 2026-08-30T08:39:29.467469+00:00.