Learn about innovative storage solutions for Large Language Model (LLM) training in this technical presentation from SK Hynix experts. Explore how checkpoint offloading SSD technology addresses performance bottlenecks and enhances scalability in LLM training environments. Discover methods for managing model states, including parameters, momentums, and variances, while reducing data movement between GPUs and storage. Examine experimental results demonstrating how AI storage solutions can optimize GPU memory usage and improve overall training efficiency by offloading optimizer operations to storage. Gain insights into practical approaches for handling interruptions and failures during LLM training through persistent storage strategies.
Checkpoint Offloading SSD - Enhancing Performance and Scalability in LLM Training
Open Compute Project via YouTube
Overview
Syllabus
Checkpoint Offloading SSD Enhancing Performance and Scalability in LLM Training
Taught by
Open Compute Project
Related Courses
-
Unlocking LLM Performance with eBPF - Optimizing Training and Inference Pipelines
-
LLM Service Revolution Through Memory Computing Fusion Technology - From Datacenter to On-Device AI
-
Handling Multi-Terabyte LLM Checkpoints - MLOps Podcast #228
-
MegaScale - Scaling Large Language Model Training to More Than 10,000 GPUs
-
LLM Post-Training Techniques - Lecture 14
-
LLM Post-Training - Lecture 14
