Post
2519
Recently, we open-sourced YaFSDP, Yandex’s tool for efficient distributed training of LLMs.
Here are some of the key ideas used in YaFSDP to provide speedup and memory savings over FSDP:
• Allocate and utilize just two buffers throughout the transformer for all collected weights to circumvent the torch memory allocator;
• Gather small normalization layers at the beginning of the iteration and average the gradients only at the end;
• Move gradient division to the very end of the backward pass.
To learn more about how YaFSDP works, check out our latest blog post: https://medium.com/yandex/yafsdp-a-tool-for-faster-llm-training-and-optimized-gpu-utilization-is-no-632b7539f5b3
Here are some of the key ideas used in YaFSDP to provide speedup and memory savings over FSDP:
• Allocate and utilize just two buffers throughout the transformer for all collected weights to circumvent the torch memory allocator;
• Gather small normalization layers at the beginning of the iteration and average the gradients only at the end;
• Move gradient division to the very end of the backward pass.
To learn more about how YaFSDP works, check out our latest blog post: https://medium.com/yandex/yafsdp-a-tool-for-faster-llm-training-and-optimized-gpu-utilization-is-no-632b7539f5b3