Out of the Memory Barrier: A Highly Memory Efficient Training System for LLMs with Million-Token Contexts

2026-02-02

AI Summary

This work addresses the prohibitive GPU memory cost of training large language models on ultra-long contexts, where activation memory scales linearly with sequence length. The authors propose a chunked recursive training framework that integrates on-the-fly activation recomputation, paged memory management, asynchronous CPU offloading, and page-level sparse attention. This approach decouples activation memory from context length for the first time, achieving an overhead of only 10 MB per 10,000 tokens and enabling million-token-scale training. The method drastically reduces hardware requirements, successfully training the Qwen2.5-7B model with a 4-million-token context on a single H200 GPU—far exceeding the scale achievable by existing methods without large-scale clusters.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time. The primary culprits are the activations, whose memory footprints scale linearly with sequence length. We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier. Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint (O(1)) and shifts the primary bottleneck to the growing KV cache. To manage the KV cache, OOMB integrates a suite of synergistic optimizations: a paged memory manager for both the KV cache and its gradients to eliminate fragmentation, asynchronous CPU offloading to hide data transfer latency, and page-level sparse attention to reduce both computational complexity and communication overhead. The synergy of these techniques yields exceptional efficiency. Our empirical results show that for every additional 10K tokens of context, the end-to-end training memory overhead increases by a mere 10MB for Qwen2.5-7B. This allows training Qwen2.5-7B with a 4M-token context on a single H200 GPU, a feat that would otherwise require a large cluster using context parallelism. This work represents a substantial advance in resource efficiency for long-context LLM training. The source code is available at https://github.com/wenhaoli-xmu/OOMB.

Problem

• LLM training
• long-context
• GPU memory
• activation memory
• memory efficiency

Innovation

• memory-efficient training
• long-context LLMs
• activation recomputation
• paged KV cache
• sparse attention