SGLang Boosts GLM-5.2 Throughput to 500+/s
BanghuaZ · x · 2026-07-15
This deep dive explores how SGLang serves GLM-5.2's agentic coding workload on 8xB300, aiming to simultaneously improve single-user interactivity and high-concurrency throughput.
The article highlights key results: achieving 500+ tok/s/user at bs=1; boosting single-user interactivity by 18%–34% within two weeks of release; and increasing peak throughput under high concurrency by 6%–11%. The author attributes these gains to two areas: model-side architectural optimizations (GLM-5.2 uses IndexShare in the DSA layer and introduces a stronger MTP head, reusing IndexShare and KVShare) and serving-side optimizations.
A standout engineering detail is the new TopK-V2 kernel: it is 2.33x faster at 80K ISL, scales to 10.17x faster at 1M ISL, and keeps interactivity nearly flat for long contexts.
Related event: SGLang v0.5.15 tunes GLM-5.2 serving to 500+ tok/s on 8×B300(6 posts)→
More from coding & agent
- MathModelAgent gains traction: auto-solves math modeling and writes a submission-ready paper — jihe520 · 2026-09-11
- alphaXiv open-sources OpenResearch to run parallel research agents with any model — alphaXiv · 2026-09-11
- DeskcommCRM: open-source AI sales CRM with native agents and WhatsApp hits 1k stars — melgarafael · 2026-09-11
- hyperresearch: agent-driven knowledge base that turns web research into a searchable wiki — jordan-gibbs · 2026-09-11
- Forter's 13 lessons from its agent sprint: skip custom RAG, lean on mature enterprise search — bibryam · 2026-09-11
- Two real 'company brains' opened up live: Gorgias' in-house Cortex vs Slite — femke_plantinga · 2026-09-11