Unsloth Ships AMD GPU Support, Bringing 2x Faster Fine-Tuning to Radeon

Unsloth, the popular open-source LLM fine-tuning toolkit with 66.5k GitHub stars, just shipped official AMD GPU support. Built in collaboration with AMD's engineering team, the release brings Unsloth's signature speed and memory tricks to Radeon, Instinct, and Ryzen AI hardware, covering everything from consumer RX 9000 cards to MI350 data center GPUs. The pitch: up to 2x faster training and 70% less VRAM usage, with no accuracy loss.

The problem this actually solves

AMD GPU owners have been second-class citizens in the LLM fine-tuning world. The ROCm software stack has matured significantly in 2026, but the tooling gap remained painful. For years, running large language models locally meant buying an NVIDIA GPU, not because AMD hardware was bad, but because the software ecosystem lagged far behind CUDA in compatibility and stability. Unsloth's AMD release directly attacks that gap by porting its custom Triton kernels and memory-efficient training algorithms to ROCm.

The VRAM story is particularly compelling. Unsloth collaborated with AMD to enable up to 2x faster and 70% less VRAM training for all models with no accuracy loss. That means you can now train Gemma 4 models in 8GB VRAM or Qwen3.5 in 3GB VRAM on AMD hardware, numbers that were previously only achievable on NVIDIA.

What hardware is supported

Unsloth now officially supports AMD hardware, making it easy to train, fine-tune, run reinforcement learning, and deploy models locally on AMD GPUs, with support for AMD's Radeon RX 9000, 7000 series, Instinct MI350 and MI300 data center GPUs, Strix Halo powered Ryzen AI Max systems, and more.

The full breakdown by architecture:

• RDNA 4 (RX 9000 series): Full support on Windows, WSL, and Linux
• RDNA 3.5 (Ryzen AI 300/MAX, Strix Halo): Full support on all platforms
• RDNA 3 (RX 7000 series): Full support on all platforms
• CDNA 4/3/2 (Instinct MI350, MI300, MI200): Full support on Linux
• RDNA 2 (RX 6000 series): Limited, Linux only (gfx1030 only)

How it was built

Unsloth's custom Triton kernels were ported to AMD by working closely with the ROCm team, requiring a rework of the installer, hardware detection, and runtime. This was not a surface-level port. The team had to rebuild across five major areas:

• ROCm kernels: HIP/Triton ports tuned for RDNA and CDNA, with architecture-specific precision and stability fixes
• 4-bit training: bitsandbytes-based QLoRA support for Radeon and CDNA GPUs
• Automatic setup: robust AMD GPU detection and installation of the correct ROCm PyTorch wheel
• Windows and Strix Halo: native Windows ROCm installation, runtime compatibility patches, and unified-memory reporting
• ROCm inference: architecture-specific llama.cpp prebuilts with source-build fallback

The team also shipped 95 automated tests to make sure it stays working, covering the full ROCm stack and running entirely in CI without physical AMD hardware, so every PR gets validated before it ships.

The key technical wins

One of the more interesting engineering decisions is how Unsloth handles LoRA and QLoRA weight management during reinforcement learning. Standard RL engines merge and demerge LoRA weights into the base model during training, which introduces floating-point drift. Unsloth uses vLLM's LoRA path directly, so the base weights are never edited and will never drift. This matters more on AMD because BF16 has fewer mantissa bits, making rounding errors more likely to compound.

On reinforcement learning specifically, the memory savings are substantial. Unsloth supports up to 80% less VRAM usage for reinforcement learning such as GRPO, and weight sharing with vLLM to save 50% memory usage.

For standard supervised fine-tuning, the benchmarks on an AMD MI300X are concrete: on Llama-3.1-8B LoRA SFT, Unsloth trains at 2.07 s/step vs 2.87 s/step for TRL + FA2, and peaks at 18.3 GB versus 24.3 GB, that is 1.39x faster and 1.33x less memory usage, with no change in accuracy.

One notable limitation: on AMD GPUs, Flash Attention 2 is not available. Unsloth automatically falls back to Xformers, which provides equivalent performance on ROCm.

Getting started

Installation is a single command. The installer auto-detects your AMD GPU, pulls the correct ROCm PyTorch wheels, installs bitsandbytes, and launches Unsloth Studio, the browser-based UI for training and inference.

# macOS, Linux, WSL
curl -fsSL https://unsloth.ai/install.sh | sh
# Windows PowerShell
irm https://unsloth.ai/install.ps1 | iex

If you prefer the Python library without the UI:

uv pip install unsloth[amd]

Once installed, you can connect local AMD-hosted models directly to coding agents:

# Connect a local model to Claude Code
unsloth start claude --model unsloth/gemma-4-E2B-it-GGUF:UD-Q4_K_XL

The easiest way to experience Unsloth on AMD without any local setup is through AMD's Developer Cloud, which offers one-click notebook access powered by MI300X GPUs with 192GB VRAM. AMD also offers $100 in free credits through the AMD AI Developer Program.

What you can actually do with it

The feature set goes well beyond basic fine-tuning. Unsloth supports both inference and training for AMD, so if you don't have a GPU and only CPU, you can still run models with Unsloth. The full list of what works on AMD:

• SFT, QLoRA 4-bit fine-tuning, and GRPO reinforcement learning
• Inference for 500+ models including Qwen3.6, Gemma 4, DeepSeek-V4, GLM-5.2, Kimi K2.7, and DiffusionGemma
• Export to GGUF, merged safetensors, or LoRA adapters for deployment with llama.cpp, vLLM, or Ollama
• Multi-GPU training on Linux via RCCL (Windows multi-GPU not yet supported)
• Tool call healing for inference to increase accuracy by 50%, unlimited free web search, HTML canvases, secure HTTPS deployment via Cloudflare, and code execution
• Connect local models to Claude Code, Codex, Hermes, OpenClaw, and OpenCode agents

The bigger picture

This release lands at a moment when the ROCm ecosystem is genuinely maturing. For standard LLM inference with PyTorch and vLLM, ROCm on MI300X or MI355X reaches 90-95% of H100 throughput. Meta's 6-gigawatt AMD commitment signals that ROCm is production-validated at scale. The software gap that made AMD a second choice for AI workloads is closing fast.

But the gap is not closed yet. Where CUDA still dominates: workloads using TensorRT-LLM, FlashAttention 3, NVIDIA NIM containers, and any pipeline with CUDA-specific custom kernels. Performance on AMD still trails CUDA by about 15-25% in raw training throughput. Unsloth's custom kernels narrow that gap considerably, but they do not eliminate it entirely.

What Unsloth's AMD support actually changes is the access story. Previously, getting efficient fine-tuning running on AMD required significant manual effort: patching ROCm builds, managing bitsandbytes compatibility, and debugging architecture-specific kernel issues. The one-line installer now automatically detects your AMD GPU, installs the correct ROCm-compatible PyTorch wheels, and launches the Studio UI so you can get started immediately. That friction reduction is the real unlock, especially for the large installed base of AMD consumer GPU owners who have been locked out of the fine-tuning workflow entirely.

Multi-GPU distributed training on Windows remains a known gap, since AMD ROCm does not yet provide a GPU-collective backend on Windows, with GLOO/CPU-only support available, so Linux is still the recommended path for serious multi-GPU work. The team has flagged this as an active area of development alongside expanded hardware CI/CD coverage and further ROCm speed optimizations.