April 2024
Intermediate to advanced
264 pages
6h 10m
English

What are the different multi-GPU training paradigms, and what are their respective advantages and disadvantages?
Multi-GPU training paradigms can be categorized into two groups: dividing data for parallel processing with multiple GPUs and dividing the model among multiple GPUs to handle memory constraints when the model size surpasses that of a single GPU. Data parallelism falls into the first category, while model parallelism and tensor parallelism fall into the second category. Techniques like pipeline parallelism borrow ideas from both categories. In addition, current software implementations such as DeepSpeed, ...
Read now
Unlock full access