Strix Halo toolbox brings local LLM finetuning

This is the kind of practical local AI work that matters more than hype: it turns AMD’s big unified-memory Strix Halo boxes into credible fine-tuning rigs instead of mere inference toys.

• –The repo documents real memory and runtime envelopes, including Gemma 3 12B full fine-tuning on a 128GB Strix Halo system and GPT-OSS-20B LoRA runs in roughly an hour
• –Multi-node support is the standout update, with DDP for speed when models fit and FSDP for sharding larger jobs across two machines
• –The setup leans on Unsloth’s growing fine-tuning stack, which already supports Gemma, Qwen, and gpt-oss workflows across notebooks and local training paths
• –It is still a power-user project: ROCm nightlies, custom RCCL patching, kernel boot flags, RDMA device mapping, and Jupyter-first workflows mean this is not plug-and-play for beginners
• –For AI developers interested in local ownership, privacy, or avoiding expensive GPU rentals, this is a strong proof point that consumer AMD hardware is getting surprisingly usable for serious LLM customization