vLLM Scales DeepSeek-V4.1-Flash on GB300 NVL72, Matching DeepInfer Performance

1 reports archived1 independent sourcesupdated 10/23/2026, 08:00:00

Synthesis & Latest Updates

1 Sources Cross-Validated

vLLM, utilizing P/D disaggregation and Wide-EP, achieves 954K tokens/s per GPU on an NVIDIA GB300 NVL72 rack for DeepSeek-V4.1-Flash. Its decode throughput is within 6% of DeepSeek's specialized engine, DeepInfer, demonstrating that a general-purpose inference engine can match model-specific performance through targeted engineering at scale.

LATEST/vLLM, utilizing P/D disaggregation and Wide-EP, achieves 954K tokens/s per GPU on an NVIDIA GB300 NVL72 rack for DeepSeek-V4.1-Flash. Its decode throughput is within 6% of DeepSeek's specialized engine, DeepInfer, demonstrating that a general-purpose inference engine can match model-specific performance through targeted engineering at scale.

Hourly heat curve

3 fully observed hours

Now

6.8

Peak

7.204:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first

vLLM 官方博客T1Oct 23, 08:00·85 pts
vLLM Scales DeepSeek-V4.1-Flash on GB300 NVL72, Matching DeepInfer Performance
- vLLM achieves 954K tokens/s per GPU on GB300 NVL72, a 7.9x improvement over single-node results
- Decode performance matches DeepSeek's specialized engine (DeepInfer) within 6% via P/D disaggregation and Wide-EP
- CPU overhead becomes critical at NVL72 scale, while small DEP instances optimize prefill efficiency
View Related Reports & Analysis Original Source