Metis - Fast Automatic Distributed Training on Heterogeneous GPUs

Explore a groundbreaking conference talk from USENIX ATC '24 that introduces Metis, an innovative system for automatic distributed training on heterogeneous GPUs. Delve into the challenges of expanding deep learning model sizes and the need to utilize diverse GPU types efficiently. Learn how Metis optimizes key system components to leverage the compute powers and memory capacities of various GPU types, enabling fine-grained distribution of training workloads. Discover the novel search algorithm developed to efficiently prune large search spaces and balance loads with heterogeneity-awareness. Examine the evaluation results showcasing Metis' superior performance in finding optimal parallelism plans for large models like GPT-3, MoE, and Wide-Resnet across multiple GPU types. Gain insights into how Metis achieves significant training speed-ups while reducing profiling and search overheads compared to traditional methods and oracle planning.

USENIX ATC '24 - Metis: Fast Automatic Distributed Training on Heterogeneous GPUs