MedQA: Fine-Tuning a Clinical AI on AMD ROCm — No CUDA Required

Source: HuggingFace Date: 2026-05-08 URL: https://huggingface.co/blog/lablab-ai-amd-developer-hackathon/medqa

Summary

A lablab.ai AMD hackathon entry demonstrates fine-tuning Qwen3-1.7B for medical question-answering on AMD Instinct MI300X hardware using LoRA (rank=8, ~2.2M trainable parameters, 0.15% of total), trained on 2,000 MedMCQA samples in approximately 5 minutes with no quantization required. The HuggingFace ecosystem (Transformers, PEFT, TRL, Accelerate) ran on ROCm with only three environment variable changes. The resulting model produces both multiple-choice answers and clinical explanations; the weights and a live demo are published on HuggingFace.

Implications

• Hardware landscape / CUDA alternatives: confirms that mainstream HuggingFace fine-tuning tooling works on AMD ROCm MI300X with minimal friction — the integration gap that kept teams on NVIDIA is narrowing.
• Local/edge model fine-tuning: LoRA on a 1.7B model in 5 minutes on a single accelerator demonstrates how low the barrier to domain-specific fine-tuning has fallen; relevant to any team evaluating local model customization.
• MI300X positioning: AMD’s 192 GB HBM3 memory headroom means no quantization workarounds needed for models at this scale, which is a meaningful practical advantage over consumer NVIDIA cards for fine-tuning workflows.