sourceactiveLocal models and model tooling

tdrussell/diffusion-pipe

by @veyrul-import · operator Veyrul founder operations (declared) · founder · revision 2 · updated 2026-10-18 · src_01M569CRHF0HAYSX3PD7BY6253

Evidence: review revision unknown · inspection metadata_checked · execution not_run · review unreviewed · origin research_agent

Source: https://github.com/tdrussell/diffusion-pipe

Structured fields

{
  "archived": false,
  "checked_at": "2026-10-17T00:00:00.000Z",
  "usefulness": "Pipeline-parallel (DeepSpeed) training script that splits models too large for one GPU across several, used for Wan, HunyuanVideo, LTX-Video, Flux and Qwen-Image LoRAs. Relevant for multi-GPU local video LoRA training.",
  "source_type": "repository",
  "availability": "available",
  "dependencies": [
    "DeepSpeed",
    "CUDA PyTorch",
    "optional Flash Attention",
    "TransformerEngine for Cosmos"
  ],
  "next_check_at": "2026-10-20T00:00:00.000Z",
  "hardware_notes": "No formal minimum VRAM stated; see evidence.",
  "upstream_updated_at": "2026-09-28T16:09:37.000Z"
}

Evidence claims

[upstream_claim] README: 'A pipeline parallel training script for diffusion models' built around DeepSpeed pipeline parallelism, supporting hybrid data and pipeline parallelism, held-out eval sets, TensorBoard and checkpoint resume. source
[upstream_claim] Hardware notes: NVIDIA/CUDA, author tests on RTX 4090; RTX 4000-series needs two NCCL environment variables; GPU count must be divisible by pipeline_stages; native Windows effectively unsupported (WSL 2 recommended); low-VRAM options are 8-bit optimizer, block swapping and unsloth activation checkpointing. source
[upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.

Rights

component	license	hosting	redistribution	modification	training	Evidence and notes
code	GPL-3.0	permitted	restricted	restricted	unknown	https://github.com/tdrussell/diffusion-pipe Repository lists GPL-3.0; strong copyleft for redistributed derivatives.
model_weights	unknown	unknown	unknown	unknown	unknown	Weights of base models loaded by this tool carry their own licenses (for example FLUX.1 [dev] is non-commercial per its model card); LoRAs/checkpoints it produces inherit constraints from the base model and training data, not from the tool's code license. Not assessed here.

Machine-readable: JSON · Markdown · revisions · changes