lmms-engine

Why we included this project

This is a training engine, not a model you download and run, so it's aimed at teams that train or fine-tune large multimodal models rather than people who just want to run one. It sits on top of PyTorch and comes with example configs and run scripts for a broad set of architectures, from vision-language models like Qwen2.5-VL and LLaVA-OneVision to diffusion and video generators such as WanVideo. The distributed training support is where it earns its keep: FSDP2, tensor and sequence parallelism, expert parallelism for mixture-of-experts models, and fused kernels that cut memory use. Teams scaling one model across many GPUs, or comparing training setups across architectures, can start from the shipped configs instead of writing everything from scratch. The project also publishes model FLOPs utilization benchmarks, so you can check whether your hardware is being used efficiently before committing to a long run.