Research map / Architectures and efficiency / Attention and transformers
Out of the Memory Barrier: A Highly Memory-Efficient Training System for LLMs with Million-Token Contexts
Wenhao Li, Daohai Yu, Gen Luo, Yuxin Zhang, Yifan Wu, Jiaxin Liu, Ziyang Gong, Zimu Liao, Fei Chao, Rongrong Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OOMB dense attention instead of FlashAttention baseline | Qwen2.5-7B training · 64K context | per-iteration latency: 76.1 -> 68.9 | Improved |
| add CPU offload (ablation) | OOMB dense attention · 256K context | peak memory (MB): 76650 -> 37129 | Improved |
| add CPU offload (ablation) | OOMB dense attention · 256K context | per-iteration latency: 1015.9 -> 985.8 | No clear effect |
| use page-level sparse attention (4K+512) instead of dense (ablation) | OOMB · 256K context | per-iteration latency: 1015.9 -> 75.7 | Improved |
| add CPU offload to sparse attention (ablation) | OOMB sparse 4K+512 · 256K context | per-iteration latency: 75.7 -> 96.9 | Hurt |
| use OOMB sparse attention instead of Ring Flash Attention | long-context training · 256K context | per-device throughput: 218.41 -> 1301.60 | Improved |
| use sparse attention with larger retrieval budget (ablation) | OOMB · 16K-256K context | training loss: 1.50000 -> 1.32031 | Improved |
| use sparse attention instead of dense (ablation) | OOMB · 1M context | training loss: 4.31250 -> 4.78125 | Hurt |
| chunk size above 4096 (ablation) | OOMB · chunk size ablation | throughput | No clear effect |
Closest papers by meaning
- A Little Goes a Long Way: Efficient Long Context Training and Inference with Partial Contexts (ICLR 2025 · similarity 0.93)
- Scaling Instruction-tuned LLMs to Million-token Contexts via Hierarchical Synthetic Data Generation (ICLR 2025 · similarity 0.91)
- GaLore: Memory-Efficient LLM Training by Gradient Low-Rank Projection (ICML 2024 · similarity 0.90)
- InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory (NeurIPS 2024 · similarity 0.90)
- SCBench: A KV Cache-Centric Analysis of Long-Context Methods (ICLR 2025 · similarity 0.90)
- EE-LLM: Large-Scale Training and Inference of Early-Exit Large Language Models with 3D Parallelism (ICML 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
