SGLang Adds /v1/score and Multi-Item Scoring, Cuts 16-Candidate p95 to 20.6ms

hsu_byron · x · 2026-09-26

LMSYS published a blog on scaling JEV-like decision models with SGLang. Serving score-based decision models with Generate + top-k logprobs risks dropping the needed label and recomputing shared context per candidate. SGLang's new /v1/score endpoint returns scores for exact requested labels, while Multi-Item Scoring (MIS) computes the shared context once with candidates isolated. MIS latency stays nearly flat from 2 to 16 candidates: on Qwen3-8B, 16-candidate p95 drops from 54.1ms to 20.6ms, and stays under 100ms under load on Qwen3-0.6B vs seconds for alternatives. Contributed by LinkedIn.

Original post →

More from coding & agent

coding & agent channel →