feat(scheduler): ratio-based decode reservation, chunked prefill and retract preemption#7038

Open

Foriv wants to merge 14 commits intoPaddlePaddle:developfrom

Commits on Mar 21, 2026

refactor(scheduler): replace block-based reservation with SGLang-aligned token-ratio mechanism
Foriv
committed
fix: add threshold check to RUNNING loop prefill path to prevent unbounded prefill admission
Foriv
committed

Commits on Mar 22, 2026

fix: align chunked prefill with SGLang single-chunked_req-per-step architecture
Foriv
committed

Commits on Mar 23, 2026

Commits on Mar 24, 2026

Commits on Mar 25, 2026

temp0325
Foriv
committed