# tdrussell/diffusion-pipe

A pipeline parallel training script for diffusion models.

Repository: https://github.com/tdrussell/diffusion-pipe
Canonical: https://ross.abutalabs.com/products/diffusion-pipe
Language: Python
License: GPL-3.0
License Family: copyleft
Last push: 2026-08-20T23:04:47+00:00

## Health v2 (maintenance only)
Score: 67/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 98, release rhythm 35, longevity 54
- inputs: {"age_days": 761, "days_push": 13, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 2015, forks 285 (observed 2026-08-28T04:06:05.513991+00:00)

## What it is
A Python training script for fine-tuning diffusion models (image and video generation) using DeepSpeed pipeline parallelism across multiple GPUs. It supports many popular models like SDXL, Flux, Wan, and HunyuanVideo with latent caching, checkpointing, and Tensorboard metrics.

## Use cases
- train a LoRA for Flux or SDXL on multiple GPUs
- fine-tune a video diffusion model like Wan2.1 or HunyuanVideo
- train a diffusion model too large to fit on a single GPU
- precompute latents and text embeddings for faster diffusion training
- resume diffusion model training from a checkpoint
- evaluate generalization of a fine-tuned diffusion model on a held-out set

## When to choose
- you want to fine-tune or train LoRAs for supported image/video diffusion models
- you need pipeline parallelism to train models larger than one GPU's memory
- you want efficient multi-GPU latent and text embedding caching
- you run Linux with NVIDIA GPUs and DeepSpeed

## When to avoid
- you need native Windows training - DeepSpeed support is partial
- you want a GUI-based trainer rather than TOML-configured CLI scripts
- your model is not among the supported architectures and you don't want to implement a subclass

## Facets
- artifact type: cli-tool
- maturity: active
- function: llm-training, deep-learning, machine-learning, gpu-computing
- domain: deep-learning, machine-learning, image-processing, artificial-intelligence
- platform: python, cli
- tags: diffusion-models, pipeline-parallelism, lora-training, deepspeed, fine-tuning, stable-diffusion, flux, video-generation, text-to-image, tensorboard, video, linux, gpu

## Member repositories
- tdrussell/diffusion-pipe (main) score 67

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:06:05.513991+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 2026-08-30T03:00:44.065379+00:00, confidence not recorded.
  - readme: https://github.com/tdrussell/diffusion-pipe (fetched 2026-08-28T04:06:05.513991+00:00, sha 6a58e4547770)
- Data as of 2026-08-30T08:39:29.467469+00:00.