Inference Infrastructure & ServingMicrosoft Research India (Agrawal et al.)Microsoft Research

Sarathi-Serve

Sarathi-Serve — chunked prefill, stall-free scheduling

Why it matters

Solved the problem of long prompts stalling decode; 2.6–5.6× the serving capacity of vLLM

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2403.02310

Added in Week 35, 2026 · See that week

Same week

India's Union Cabinet approves the IndiaAI MissionGovernmentIEEE Spectrum publishes AI Prompt Engineering Is DeadPressAnthropic announces the Claude 3 family (Haiku, Sonnet, Opus)Official

Nearby on the same axis

Blackwell platformOfficialDistServePaperLMSYS releases SGLang + RadixAttentionOfficialSGLang paperPaper