Seq1F1B: Efficient Sequence-Level Pipeline Parallelism for Large Language Model Training

TL;DR

Seq1F1B introduces a sequence-level pipeline parallelism method that significantly improves training efficiency and memory usage for large language models on long sequences, enabling training of 30B parameter models on sequences up to 64k without recomputation.

Contribution

The paper proposes Seq1F1B, a novel sequence-level pipeline scheduling method that reduces memory footprint and pipeline bubbles, enabling efficient training of large models on very long sequences.

Findings

Achieves higher throughput than baseline methods.

Reduces memory footprint during training.

Successfully trains a 30B parameter model on 64k sequences without recomputation.

Abstract

The emergence of large language models (LLMs) relies heavily on distributed training strategies, among which pipeline parallelism plays a crucial role. As LLMs' training sequence length extends to 32k or even 128k, the current pipeline parallel methods face severe bottlenecks, including high memory footprints and substantial pipeline bubbles, greatly hindering model scalability and training throughput. To enhance memory efficiency and training throughput, in this work, we introduce an efficient sequence-level one-forward-one-backward (1F1B) pipeline scheduling method tailored for training LLMs on long sequences named Seq1F1B. Seq1F1B decomposes batch-level schedulable units into finer sequence-level units, reducing bubble size and memory footprint. Considering that Seq1F1B may produce slight extra bubbles if sequences are split evenly, we design a computation-wise strategy to partition input sequences and mitigate this side effect. Compared to competitive pipeline baseline methods such as Megatron 1F1B pipeline parallelism, our method achieves higher training throughput with less memory footprint. Notably, Seq1F1B efficiently trains a LLM with 30B parameters on sequences up to 64k using 64 NVIDIA A100 GPUs without recomputation strategies, a feat unachievable with existing methods. Our source code is based on Megatron-LM, and now is avaiable at: https://github.com/MayDomine/Seq1F1B.git.