Out of the Memory Barrier: A Highly Memory Efficient Training System for LLMs with Million-Token Contexts
This work addresses a critical memory bottleneck for researchers and practitioners training LLMs on long sequences, offering a substantial resource efficiency advance.
The paper tackles the prohibitive GPU memory overhead in training LLMs on long contexts by introducing OOMB, a memory-efficient system that reduces activation memory to constant O(1) and manages KV cache with optimizations, enabling training Qwen2.5-7B with a 4M-token context on a single H200 GPU with only 10MB overhead per 10K tokens.
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.