SPPO:Efficient Long-sequence LLM Training via Adaptive Sequence Pipeline Parallel Offloading

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Chen, Qiaoling, Li, Shenggui, Gao, Wei, Sun, Peng, Wen, Yonggang, Zhang, Tianwei
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916651855773696
author Chen, Qiaoling
Li, Shenggui
Gao, Wei
Sun, Peng
Wen, Yonggang
Zhang, Tianwei
author_facet Chen, Qiaoling
Li, Shenggui
Gao, Wei
Sun, Peng
Wen, Yonggang
Zhang, Tianwei
contents In recent years, Large Language Models (LLMs) have exhibited remarkable capabilities, driving advancements in real-world applications. However, training LLMs on increasingly long input sequences imposes significant challenges due to high GPU memory and computational demands. Existing solutions face two key limitations: (1) memory reduction techniques, such as activation recomputation and CPU offloading, compromise training efficiency; (2) distributed parallelism strategies require excessive GPU resources, limiting the scalability of input sequence length. To address these gaps, we propose Adaptive Sequence Pipeline Parallel Offloading (SPPO), a novel LLM training framework that optimizes memory and computational resource efficiency for long-sequence training. SPPO introduces adaptive offloading, leveraging sequence-aware offloading, and two-level activation management to reduce GPU memory consumption without degrading the training efficiency. Additionally, SPPO develops an adaptive pipeline scheduling approach with a heuristic solver and multiplexed sequence partitioning to improve computational resource efficiency. Experimental results demonstrate that SPPO achieves up to 3.38x throughput improvement over Megatron-LM and DeepSpeed, realizing efficient training of a 7B LLM with sequence lengths of up to 4M tokens on only 128 A100 GPUs.
format Preprint
id arxiv_https___arxiv_org_abs_2503_10377
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SPPO:Efficient Long-sequence LLM Training via Adaptive Sequence Pipeline Parallel Offloading
Chen, Qiaoling
Li, Shenggui
Gao, Wei
Sun, Peng
Wen, Yonggang
Zhang, Tianwei
Distributed, Parallel, and Cluster Computing
In recent years, Large Language Models (LLMs) have exhibited remarkable capabilities, driving advancements in real-world applications. However, training LLMs on increasingly long input sequences imposes significant challenges due to high GPU memory and computational demands. Existing solutions face two key limitations: (1) memory reduction techniques, such as activation recomputation and CPU offloading, compromise training efficiency; (2) distributed parallelism strategies require excessive GPU resources, limiting the scalability of input sequence length. To address these gaps, we propose Adaptive Sequence Pipeline Parallel Offloading (SPPO), a novel LLM training framework that optimizes memory and computational resource efficiency for long-sequence training. SPPO introduces adaptive offloading, leveraging sequence-aware offloading, and two-level activation management to reduce GPU memory consumption without degrading the training efficiency. Additionally, SPPO develops an adaptive pipeline scheduling approach with a heuristic solver and multiplexed sequence partitioning to improve computational resource efficiency. Experimental results demonstrate that SPPO achieves up to 3.38x throughput improvement over Megatron-LM and DeepSpeed, realizing efficient training of a 7B LLM with sequence lengths of up to 4M tokens on only 128 A100 GPUs.
title SPPO:Efficient Long-sequence LLM Training via Adaptive Sequence Pipeline Parallel Offloading
topic Distributed, Parallel, and Cluster Computing
url https://arxiv.org/abs/2503.10377