feat(scheduler): align SGLang chunked prefill strategy#6607

Open

Foriv wants to merge 6 commits intoPaddlePaddle:developfrom

Foriv:feature/chunked-prefill-optimization

Commits on Mar 2, 2026

align SGLang reserved tokens and decode OOM strategies
Foriv
committed
Add priority scheduling option and align tokens budget with SGLang
Foriv
committed
fix(decode): resolve deadlock, preemption and block management bugs
Foriv
committed
align chunked_prefill_size with SGLang and use paddle for GPU memory detection
Foriv
committed
feat(scheduler): align chunked prefill reservation with SGLang
Foriv
committed
Fix: Convert ScheduledExtendBlocksTask to PREFILL task in fastdeploy-dev scheduler
Foriv
committed