diffusion-pipe

Splitting large diffusion models across several GPUs is what diffusion-pipe by tdrussell handles: it trains LoRAs and full models using pipeline parallelism through DeepSpeed, combining hybrid data and pipeline parallelism so models too big for one card can still be fine-tuned on consumer hardware. Supported architectures span more than 20 models, including SDXL, FLUX, LTX-Video, HunyuanVideo, and the Wan family, which made it a fixture of the video LoRA training community. Latents and text embeddings are precached to disk before training, cutting VRAM pressure and keeping text encoders out of GPU memory during the run. Setup involves cloning with submodules, a Python 3.12 conda environment, a separate PyTorch install, and the NVIDIA CUDA toolkit; development targets Linux, native Windows is essentially unsupported, and WSL 2 is the recommended path for Windows users, with RTX 4090 class cards used for testing. Configuration lives in commented example files in the examples directory. The project is GPL-3.0 licensed, has around 2,000 stars, and saw active development into 2026.