Research map / Architectures and efficiency / Attention and transformers
Mosaic: Unlocking Over 30$\times$ Context Length for Diffusion LLMs Inference via Global Memory Planning and Dynamic Peak Taming
Liang Zheng, Bowen Shi, Yitao Hu, Jiawei Zhang, Ruofan Li, Guotao Yang, Zhixin Zhao, Zhengchao Wang, Sheng Chen, Wenxin Li, Dezhi Ran, Tao Xie and 1 more
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 |
|---|---|---|---|
| add global memory manager (ablation) | Mosaic baselines · dLLM inference rp=0.5 | Lmax multiple: 1.41x-3.00x | Improved |
| add mask-only logits kernel (ablation) | Mosaic with global manager · dLLM inference rp=0.5 | Lmax multiple: 3.68x-8.04x | Improved |
| add lazy chunking optimizer (ablation) | Mosaic with global manager and mask-only · dLLM inference rp=0.5 | Lmax multiple: 15.30x-32.34x | Improved |
| use lazy chunking instead of static full chunking (ablation) | Mosaic · dLLM inference up to 32k | latency: 5.5%-21.5% reduction | Improved |
| use lazy chunking instead of static full chunking (ablation) | Mosaic · dLLM inference | Lmax: identical Lmax | No clear effect |
| bottleneck-driven search instead of brute force (ablation) | Mosaic · dLLM inference | search latency: 0.28%-0.32% of brute-force | Improved |
| first-fit planning instead of ILP (ablation) | Mosaic · dLLM inference | planning latency: 0.1%-4.3% of ILP | Improved |
| use first-fit instead of ILP (ablation) | Mosaic · dLLM inference | overall latency: 7.1%-36.7% reduction | Improved |
| add Mosaic to KV-cache dLLM frameworks | Fast-dLLM, dInfer, Elastic-Cache · LLaDA-8B RTX 3090 | context length: 3.75x, 3.00x, 2.27x | Improved |
| apply Mosaic under multi-GPU | LLaDA · 2 GPUs | context length: 12.30x (TP2), 14.92x (PP2) | Improved |
Closest papers by meaning
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- TEST-TIME SCALING IN DIFFUSION LLMS VIA HIDDEN SEMI-AUTOREGRESSIVE EXPERTS (ICLR 2026 · similarity 0.90)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.89)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.89)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- MemDLM: Memory-Enhanced DLM Training (NeurIPS 2026 · similarity 0.89)
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.
