Unsloth on AMD GPUs (AMD developer blog)

An AMD developer-blog post (2026-07-20) announcing that Unsloth officially supports AMD hardware for training, fine-tuning, RL and inference.

Hardware and install path

Instinct MI350/MI300/MI200 (CDNA 2–4, Linux only) · Radeon RX 9000/7000 (RDNA 3–4) · Ryzen AI Max “Strix Halo” on Windows/WSL/Linux. An installer pulls matched ROCm builds of PyTorch, llama.cpp, bitsandbytes and Triton. Assembling that toolchain, rather than anything about the algorithms, is the practical barrier to training outside CUDA.

The claims, all first-party

~80% less VRAM from Triton kernels and memory tricks — Gemma-4 trained in 8 GB, Qwen3.5 in 3 GB.
1.39× faster and 1.33× leaner than TRL + FlashAttention-2 on Llama-3.1-8B LoRA SFT: 2.07 vs 2.87 s/step, 18.3 vs 24.3 GB peak, with an identical loss curve.
A GRPO step at 1.75 s vs 2.01 s with vLLM co-located.

Per this spoke’s benchmark caveat: these are the vendor’s numbers, on the vendor’s hardware, against a baseline the vendor configured. The identical loss curve is the load-bearing detail, since it is what separates a memory optimization from a quality trade. It too is self-reported.

The durable technical point

Beyond the benchmarks, one claim outlives the release cycle: for LoRA/QLoRA, Unsloth refuses the merge/de-merge trick (add W + sAB to serve, subtract W - sAB to resume training) because BF16 rounding is not associative, so the base weights drift across cycles. It routes through vLLM’s LoRA path instead, leaving base weights untouched. That is a numerical-precision argument, not a benchmark, and it generalizes to any adapter workflow that round-trips through merged weights.

Tier

T3. A vendor announcement: the hardware support and the install path are checkable facts, the performance claims are marketing until someone independent reproduces them. Recorded with the setup attached so a later source can check them.

Cross-spoke context

The training/serving seam runs straight through this source: Unsloth ships GGUF/llama.cpp inference and co-locates with vLLM, both of which ../llm-inference-wiki already pages (vllm, llama-cpp, quantization, flash-attention). The dominant substance here is train-time (LoRA/QLoRA SFT, GRPO, VRAM during backprop), which is why it sits in this spoke.

A second thread the hub flagged and did not cluster: AMD/ROCm as the open alternative to CUDA. ../platform-ops-wiki’s knod also runs on AMD GPUs, and the parked nvidia-doca record is adjacent silicon. Different subjects; revisit if such sources accumulate.

unsloth · fine-tuning · amd · finetuning-explained-kdnuggets · synthesis