# flexflow/flexflow-train

Automatically Discovering Fast Parallelization Strategies for Distributed Deep Neural Network Training

Repository: https://github.com/flexflow/flexflow-train
Canonical: https://ross.abutalabs.com/products/flexflow-train
Homepage: https://flexflow.ai
Language: C++
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T19:10:43+00:00

## Health v2 (maintenance only)
Score: 67/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 8, longevity 100
- inputs: {"age_days": 2825, "days_push": 7, "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 1898, forks 255 (observed 2026-08-28T04:05:51.255591+00:00)

## What it is
FlexFlow Train is a deep learning framework that accelerates distributed DNN training by automatically searching for efficient parallelization strategies across samples, operators, attributes, and parameters. It provides a drop-in replacement for PyTorch and TensorFlow Keras and uses a simulator-based search algorithm to outperform manually designed parallelization strategies.

## Use cases
- speed up distributed training of large neural networks on multi-GPU clusters
- automatically find the best parallelization strategy instead of hand-tuning data or model parallelism
- import and optimize existing PyTorch models for parallel training
- train models with Keras-style APIs on distributed hardware
- benchmark parallelization strategies with an execution simulator

## When to choose
- you train large DNNs on multi-GPU or multi-node clusters and manual parallelization is slow or suboptimal
- you want automatic SOAP-dimension parallelization search for PyTorch or Keras models
- you need a research-grade framework for exploring parallelization strategies

## When to avoid
- you only train small models on a single GPU
- you need LLM inference or serving rather than training (use flexflow-serve)
- you require a mature production ecosystem like PyTorch DDP or DeepSpeed with broad community support

## Facets
- artifact type: framework
- maturity: active
- function: machine-learning, deep-learning, llm-training, gpu-computing
- domain: deep-learning, machine-learning, gpu-computing, microservices
- platform: python, cpp
- tags: distributed-training, parallelization, model-parallelism, data-parallelism, autotuning, cuda, pytorch, keras, linux, gpu, docker

## Member repositories
- flexflow/flexflow-train (main) score 67

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:51.255591+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 2026-08-30T03:12:12.912246+00:00, confidence not recorded.
  - readme: https://github.com/flexflow/flexflow-train (fetched 2026-08-28T04:05:51.255591+00:00, sha 1b9028f1bf5e)
  - homepage: https://flexflow.ai (fetched 2026-08-29T10:51:24.833989+00:00, sha 9a8530ccc753)
- Data as of 2026-08-30T08:39:29.467469+00:00.