In the news
Optimizing DeepSeek-V4.1 on H200 with SGLang PD Disaggregation
LMSYS · Published · 3 min read
In 30 seconds
- What happened
- SGLang optimized DeepSeek-V4.1 serving on H200 GPUs by fixing Indexer scoring, reducing unnecessary computation, and improving communication patterns.
- Why it matters
- Engineers deploying large language models in production need efficient inference serving, especially when balancing prefill and decode workloads across multiple GPUs.
- Watch out
- Combined optimizations reduced accuracy from 96.74% to 95.38%, a 1.36 percentage point drop. Results are configuration-specific and may not generalize to different hardware or deployment topologies.
- deepseek
The patterns behind this
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.