DISTFLASHATTN: Distributed Memory-efficient Attention For Long-context Llms Training

Li Dacheng, Shao Rulin, Xie Anze, Xing Eric P., Ma Xuezhe, Stoica Ion, Gonzalez Joseph E., Zhang Hao. Arxiv 2023

[Paper] [Code]    
Attention Mechanism Fine Tuning Has Code Model Architecture Pretraining Methods Reinforcement Learning Training Techniques Transformer

FlashAttention (Dao, 2023) effectively reduces the quadratic peak memory usage to linear in training transformer-based large language models (LLMs) on a single GPU. In this paper, we introduce DISTFLASHATTN, a distributed memory-efficient attention mechanism optimized for long-context LLMs training. We propose three key techniques: token-level workload balancing, overlapping key-value communication, and a rematerialization-aware gradient checkpointing algorithm. We evaluate DISTFLASHATTN on Llama-7B and variants with sequence lengths from 32K to 512K. DISTFLASHATTN achieves 8x longer sequences, 4.45 - 5.64x speedup compared to Ring Self-Attention, 2 - 8x longer sequences, 1.24 - 2.01x speedup compared to Megatron-LM with FlashAttention. It achieves 1.67x and 1.26 - 1.88x speedup compared to recent Ring Attention and DeepSpeed-Ulysses. Code is available at https://github.com/RulinShao/LightSeq.

Similar Work