1
The case for disaggregated LLM serving
把大模型推理拆成预填充和解码两段，用网络传KV缓存，破解长上下文延迟痛点，值得一读。
Article URL: https://blog.doubleword.ai/when-to-disaggregate Comments URL: https://news.ycombinator.com/item?id=49269470 Points: 4 # Comments: 0
2
Local AI is re-reading its own prompt
苹果 M4 Pro 本地跑复合 AI：单个 9B 模型兼任路由、专家、规划，零云端调用，拆解“预填充税”对推理开销的隐性影响。
Article URL: https://github.com/thekiraproject/kira-project-site/blob/main/posts/2026-07-06-prefill-tax.md Comments URL: https://news.ycombinator.com/…
3
Towards Load-Aware Prefill Deflection for Disaggregated LLM Serving
负载感知的预填偏转设计，破解分离式LLM服务的时延难题
arXiv:2607.02043v1 Announce Type: cross Abstract: Disaggregated LLM serving runs prefill and decode on separate GPU pools to keep the two phases from …
4
VoltanaLLM: Energy-Efficient LLM Serving
LLM服务节能新方案：利用预填充/解码解耦架构，动态GPU频率控制实现36.3%能耗节省
Article URL: https://supercomputing-system-ai-lab.github.io/projects/voltana/ Comments URL: https://news.ycombinator.com/item?id=48655520 Points: 2 # …
5
Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models
提出预测预填充方法，让扩散语言模型高效处理超长上下文，解码速度显著提升。
arXiv:2606.10537v1 Announce Type: new Abstract: Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing r…
6
Copy-as-Decode: Grammar-Constrained Parallel Prefill for LLM Editing
无需重写整个输出，LLM编辑效率新突破：Copy-as-Decode用语法约束并行预填充，大幅加速文本代码编辑。
arXiv:2604.18170v2 Announce Type: replace-cross Abstract: LLMs edit text and code by autoregressively regenerating the full output, even when most tok…
7
Prism: Spectral-Aware Block-Sparse Attention
光谱感知的块稀疏注意力机制，突破长上下文LLM预填充效率瓶颈
arXiv:2602.08426v2 Announce Type: replace-cross Abstract: Block-sparse attention is promising for accelerating long-context LLM pre-filling, yet ident…
8
PreFT: Prefill-only finetuning for efficient inference
新微调方法PreFT仅优化预填充阶段，解决大模型个性化服务时PEFT导致的吞吐量瓶颈，理论与实证兼备。
arXiv:2605.14217v1 Announce Type: cross Abstract: Large language models can now be personalised efficiently at scale using parameter efficient finetun…