Understand this story
On November 19, 2026, LMSYS announced optimizations for disaggregated prefill and decoding serving of DeepSeek-V4.1-Flash on H20-3e. Compared with the TP4 baseline, average end-to-end latency decreased by 65.5%, while output throughput rose to 2.91 times the baseline; the final configuration reduced average latency from 13.63 seconds to 4.70 seconds and increased output throughput from 847 to 2,465 tokens/s. LMSYS said the main gains came from prefill CP, decoding DP, and FP8 DeepGEMM, and also provided colocated serving results, item-by-item ablation data, and SGLang-based reproduction commands. There are currently no other substantiated developments.
AI-generated from public reports
Latest development · Report
LMSYS Analyzes PD and Co-located Serving Optimizations for DeepSeek-V4.1-Flash on H20-3eDevelopments
1 confirmed developments
LMSYS Optimizes Disaggregated Serving for DeepSeek-V4.1-Flash
Source:LMSYS Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 19
LMSYS BlogFirst-party
LMSYS Analyzes PD and Co-located Serving Optimizations for DeepSeek-V4.1-Flash on H20-3e
LMSYS optimized disaggregated prefill and decoding serving for DeepSeek-V4.1-Flash on H20-3e, reducing average end-to-end latency by 65.5% and increasing output throughput to 2.91× that of the TP4 baseline. The final configuration reduced average latency from 13.63 s to 4.70 s and increased output throughput from 847 to 2,465 tokens/s, with the main gains coming from prefill CP, decoding DP, and FP8 DeepGEMM. The article also presents co-located serving results, step-by-step ablation data, and SGLang-based reproduction commands.
AI Trends