
ROCm Studio is a guided web application that makes fine-tuning open-source language models on AMD GPUs feel accessible, repeatable, and demo-ready. Instead of asking developers to manually configure ROCm, Hugging Face, PEFT, TRL, dataset formatting, LoRA training, adapter packaging, and inference serving, ROCm Studio turns the entire workflow into a step-by-step product experience. Users can upload documents, extract and clean the text, generate fine-tuning examples with an Agentic cleaning mode, select a Gemma model and training recipe, fine-tune a LoRA adapter on an AMD MI300X GPU through AMD Developer Cloud, and then either download the adapter bundle or deploy it to a live Try-It inference endpoint. The app also includes live GPU/job progress, queueing for GPU-heavy tasks, reloadable project URLs, adapter ZIP packaging, and a chat panel for testing the trained model against uploaded documents. ROCm Studio is built with React, FastAPI, PyTorch ROCm, Hugging Face Transformers, TRL, and PEFT. It demonstrates how AMD Instinct GPUs can support practical end-to-end model customization workflows for developers, educators, teams, and domain experts who want specialized AI models without becoming infrastructure engineers.
10 May 2026