Learn about the technical challenges and solutions in heterogeneous distributed training for Large Language Models (LLMs) in this 11-minute conference talk. Explore how integrating different computing resources for distributed parallel acceleration can support the development of LLMs with hundreds of billions of parameters. Discover the research conducted by China Mobile and industry partners to overcome challenges related to GPU architecture differences, memory constraints, and vendor hardware incompatibilities. Gain insights into the core functional components of a training system designed to enable heterogeneous GPUs to work together effectively, contributing to the advancement of the intelligent computing ecosystem.