Vibeleaderboard
← All Intel
Intel / repo

unsloth v0.1.920-beta — AMD NPU + video diffusion LoRA

Source
github.com
Author
danielhanchen
Date

Context

Unsloth's v0.1.920-beta release lists models that users convert themselves in its NPU model list and loads them on the NPU in AMD's Ryzen AI processors, for text only. An NPU is a dedicated AI engine built into the chip, separate from the GPU, so a local model can run on it while the GPU stays free for other work. According to Unsloth's NPU guide, it covers Ryzen AI 300, Ryzen AI Max (Strix Halo) and Ryzen AI 400 chips on Windows 11 or Linux, and the NPU runs models in Q4NX, a 4-bit format built for AMD's FastFlowLM runtime, which Unsloth downloads during a one-time setup. Users could already convert ordinary GGUF model files to Q4NX. With this release, those converted models appear as Converted in the NPU model list and load on the NPU. The guide, when read for this explanation, still said loading your own conversions directly from the NPU model list was not yet supported, pointing users to a separate Trained section of the model picker instead, so on that point the release notes are the newer statement.

The release also makes it clearer what will work before a user commits to a conversion. Each NPU model now carries one of four labels, Runs on NPU, Convert then runs, Converts but won't run, or Not supported, read from the installed FastFlowLM, which Unsloth says stops newer models such as LFM3 from being hidden. Conversion now refuses source quantizations that would produce garbage, such as a Qwen3.5 model converted from a Q4_0 GGUF that answered only with slashes, and moves on to the next one. Unsloth's API and command-line tools now accept the NPU model names shown in the app. The release notes describe NPUs as 2 to 4 times more efficient per joule per token than GPUs, and the guide repeats the claim without describing how it was measured. The guide's speed table lists, for example, about 17 tokens per second for Qwen3.6-35B-A3B and about 41 for the small LFM3-1.7B.

Elsewhere the release adds training, a way to fine-tune by training a small set of add-on weights, for the LTX-2.5 video and MOSS-Image models, with diffusion LoRA training running 20 to 42% faster per step when the GPU has room to keep activations, and support for T4 and other GPUs without bf16.

Key takeaways · AI-distilled
  • Unsloth now loads Q4NX models converted on its Export page on AMD NPUs, text only, and tags each NPU model as runs, convert-then-runs, converts-but-won't-run or unsupported based on the installed FastFlowLM, so newer models such as LFM3 are no longer hidden.
  • Q4NX conversion now refuses GGUF quants that would load as garbage, such as a Qwen3.5 Q4_0 GGUF that answered only with slashes, and moves on to the next quant.
  • Diffusion LoRA training runs 20 to 42% faster per step when the GPU has room to keep activations, and FLUX.2-klein, Z-Image and Wan 2.2 5B LoRAs now train on T4 and other GPUs without bf16.
  • QLoRA training is up to 25% faster by keeping dequantized base weights in spare VRAM, and bf16 full fine-tuning now uses stochastic rounding by default so small updates are no longer lost.
  • Unsloth's /v1/chat/completions endpoint now applies a grammar or json_schema instead of silently ignoring it, and each API model alias can set its own , type and .
Terms in this piece · Glossary
  • LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
  • context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
  • KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
  • speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
More from danielhanchen
Recommended reads
Comments

Checking sign-in…

Loading comments…