ESC

Type to search articles...

No articles found.

↑ ↓ Navigate ↵ Open
Esc Close
Blog Tags GitHub
All tags

theory

1 post tagged with "theory"

July 3, 2026

LLM Serving Optimization with Variable Prefill and Decode Lengths

在固定 KV-cache 显存预算下研究异构 prefill/decode 长度的离线 LLM 服务调度:证明问题 NP-hard、FCFS/最短输出优先/总长度优先均有无界近似比;提出 Sorted-F 算法——用 F-metric(平均输出长度 / 批大小)平衡批并发与下游解码成本,证明常数因子近似比 ≤48;配三种 Phase-1 求解器(精确 DP、局部交换、分位贪心)与 LP 引导 / receding-horizon 变体。真实混合负载上相对 FCFS 提速 4.87×、相对 MC-SF 提速 2.09×,与 LP 下界差距仅 1.03-1.09×

inference-serving llm-inference scheduling kv-cache theory

Navigation

  • Work

Resources

  • Lexington Themes.

Socials

  • @Mike_Andreuzza
© 2025 MicroStudio. All rights reserved.

MicroStudio is not affiliated with Stripe, Breeew, Astro, or Tailwind Labs, nor is it endorsed or sponsored by them.