Crimson: Collaborative Parameter Updates for Efficient Pipeline Training of Large Language Models

Large language models (LLMs) have driven significant progress in natural language processing, yet their training and fine-tuning remain limited by memory constraints, particularly the substantial memory footprints of optimizer states. Existing solutions address this challenge by offloading optimizer states and update tasks to the CPU, but this often leads to increased GPU idleness due to the CPU's limited computational capabilities, especially in pipeline parallelism.