Unsloth Officially Launches AMD GPU Support

The LLM fine-tuning library Unsloth has begun official support for AMD hardware. With this update, local inference, fine-tuning, reinforcement learning (RL), and deployment using AMD GPUs are now possible on Windows, Linux, and WSL environments.

Key support details:

• Hardware: Radeon RX 7000/9000 series, Instinct MI300/MI350, Strix Halo/Ryzen AI Max systems, and AMD CPUs (for inference without a GPU)
• Optimization technologies: Automatic installation support for optimized builds of ROCm, Triton, bitsandbytes, PyTorch, and llama.cpp
• Performance benefits: Reduces VRAM usage by up to 70% compared to before, and during reinforcement learning, reductions of up to 80% are possible
• Compatible models: Support for major models including Qwen, Gemma, DeepSeek, GLM, Kimi, MiniMax, and DiffusionGemma
• Other features: Models can be exported as GGUF, safetensors, or LoRA adapters, and can be connected to various agent tools such as Claude Code and Hermes Agent