DynamiQ: Accelerating Gradient Synchronization using Compressed Multi-hop All-reduce

TL;DR

DynamiQ is a novel gradient compression framework that accelerates multi-hop all-reduce in large model training by optimizing partial sum representation and execution, achieving up to 34.2% faster training with near-baseline accuracy.

Contribution

It introduces a new quantization technique and fused kernel design specifically optimized for multi-hop aggregation in distributed training.

Findings

Achieves up to 34.2% speedup over state-of-the-art methods.

Maintains near-baseline accuracy (99.9% of BF16 baseline).

Supports various LLMs, tasks, and scales with consistent improvements.

Abstract

Multi-hop all-reduce is the de facto backbone of large model training. As the training scale increases, the network often becomes a bottleneck, motivating reducing the volume of transmitted data. Accordingly, recent systems demonstrated significant acceleration of the training process using gradient quantization. However, these systems are not optimized for multi-hop aggregation, where entries are partially summed multiple times along their aggregation topology. This paper presents DynamiQ, a quantization framework that bridges the gap between quantization best practices and multi-hop aggregation. DynamiQ introduces novel techniques to better represent partial sums, co-designed with a decompress-accumulate-recompress fused kernel to facilitate fast execution. We extended PyTorch DDP to support DynamiQ over NCCL P2P, and across different LLMs, tasks, and scales, we demonstrate consistent improvement of up to 34.2% over the best among state-of-the-art methods such as Omni-Reduce, THC, and emerging standards such as MXFP4, MXFP6, and MXFP8. Further, DynamiQ is the only evaluated method that consistently reaches near-baseline accuracy (e.g., 99.9% of the BF16 baseline) and does so while significantly accelerating the training.