Posts by Zhu Shan
Fine-Tuning LLMs with GRPO on AMD MI300X: Scalable RLHF with Hugging Face TRL and ROCm
- 18 June 2025
In this blog, you will learn how to implement GRPO-based RLHF on AMD MI300X using ROCm and Hugging Face TRL—streamlining alignment training while enhancing model reasoning and inference performance. Reinforcement Learning from Human Feedback (RLHF) constitutes a critical phase in the fine-tuning of large language models (LLMs) and multimodal architectures. Over time, RLHF methodologies have advanced beyond traditional techniques, progressing from Proximal Policy Optimization (PPO) to Direct Preference Optimization (DPO), and more recently, to Group Relative Policy Optimization (GRPO). RLHF aims to make LLMs’ output better aligned with human preferences. Reinforcement Learning (RL) is an important step to enhance LLM’s reasoning capabilities and for better inference/test-time scaling law. Apart from LLM, there is also DPO application in text-to-image generation.