Unsloth v0.1.920-beta Adds Major AMD NPU Support and Video Diffusion LoRA Training
The latest Unsloth update brings direct AMD NPU model loading, actionable NPU model support indicators, faster image generation on new AMD and NVIDIA hardware, improved LoRA training, and enhanced API and automation features.
What changed?
Unsloth v0.1.920-beta introduces direct support for running converted Q4NX models on AMD NPUs, with each NPU model now listing its run status. The OpenAI-compatible API now accepts model names from the app or FastFlowLM repository. New commands make it easier to list and manage NPU models and their compatibility. The release also delivers up to 42% faster diffusion LoRA training per step, support for LTX-2.5 and MOSS-Image LoRA training, and now can load NVIDIA Cosmos3 Nano-FP8 and Edge-FP8 video models without extra packages. Image generation is notably faster on AMD Strix Halo hardware and several GPU operations are improved. Additional updates include revised chat automation, improved translation and interface features, and extended API endpoint support for grammar and schema.
Why does it matter to an everyday developer?
Developers building or deploying AI models on AMD hardware can now leverage NPUs for more efficient inference and training workflows. The model interface now immediately indicates which models can run on the NPU, reducing trial and error. Faster LoRA and video model loading on both AMD and NVIDIA hardware can accelerate experimentation and deployment. Streamlined API changes and new automation in chat directly enhance developer productivity, especially when integrating with the OpenAI-compatible endpoints or automating workflows within the Unsloth UI.
What can the developer do now?
- Directly convert and run Q4NX models on compatible AMD NPUs and monitor model compatibility from the UI.
- Use `unsloth list-npu-models` to view model details including memory fit.
- Accelerate image and video generation and LoRA training, especially on AMD Strix Halo and NVIDIA Cosmos3 hardware.
