Training on AMD Instinct GPUs - From Pre-training to Fine-tuning and Post-training Strategies

Learn comprehensive training strategies for large foundation models using AMD Instinct GPUs in this 31-minute conference talk from MLOps World. Explore the complete training pipeline from pre-training methodologies through fine-tuning solutions to advanced post-training optimization techniques. Discover how AMD Instinct GPUs support large language models, vision models, and multi-modal models across different training phases. Examine pre-training approaches with performance metrics and Multi-GPU setup benefits, then delve into fine-tuning solutions including Full Weight Fine-tuning and Parameter Efficient Tuning (PEFT) using Megatron-LM and HF-PEFT frameworks. Understand how the larger HBM memory of the MI300X enhances training performance and accuracy. Master post-training strategies featuring innovative distillation processes that convert multi-head attention (MHA) into more efficient architectures like Mamba and Multi-Head Latent Attention (MLA) layers for optimized model deployment. Gain practical insights into AMD's public training dockers and their key features designed to improve user experience and accessibility, while learning frameworks for implementing these advanced techniques to maximize performance on AMD GPU hardware.