Unsloth
A fine-tuning framework that cuts the memory and time cost of adapting open-weight LLMs, using custom Triton kernels and memory optimizations. Covers training, fine-tuning, RL post-training and inference; it is one of the four tools the founding explainer names alongside Hugging Face PEFT, TRL and Axolotl finetuning-explained-kdnuggets.
AMD support
As of 2026-07-20 it officially supports AMD hardware — Instinct MI350/MI300/MI200, Radeon RX 9000/7000, and Ryzen AI Max "Strix Halo" — with an installer that pulls matched ROCm builds of PyTorch, llama.cpp, bitsandbytes and Triton unsloth-amd-support. Assembling that toolchain is the usual barrier to training outside CUDA, so packaging it is most of the announcement.
Claimed performance
Roughly 80% less VRAM (Gemma-4 in 8 GB, Qwen3.5 in 3 GB), and 1.39× faster / 1.33× leaner
than TRL + FlashAttention-2 on Llama-3.1-8B LoRA SFT with an identical loss curve
unsloth-amd-support. First-party numbers on vendor-selected hardware — see the benchmark caveat
in this spoke's CLAUDE.md.
Two vendor figures, both kept (re-checked 2026-10-03). The GitHub README (github.com/unslothai/unsloth, ~77k stars on that date) now headlines "2× faster with 70% less VRAM" for fine-tuning, against the ~80% in the July AMD post above. Neither states a method on the page, and the two may describe different models or setups, so this page keeps both rather than choosing one.
What changed since July (README, 2026-10-03)
- Three forms now: Unsloth Desktop (an app, including a macOS download), Unsloth Studio (a web UI, with dataset building from PDFs, CSVs and DOCX), and Unsloth Core, the library.
- Hardware listed: NVIDIA, AMD, Intel GPUs, CPUs and a Vulkan backend, on Windows, Linux, WSL and macOS. AMD was the July news; the list is now wider than that.
- Export formats named include GGUF, NVFP4 and FP8.
Design decision worth keeping
It refuses to merge LoRA adapters into base weights for serving and unmerge them to resume, because BF16 rounding is not associative and the base weights drift; it uses vLLM's LoRA path so the base is never edited unsloth-amd-support. See fine-tuning for why that generalizes.