vLLM Achieves Near-Zero-Downtime Scaling in Elastic Expert Parallelism

1 reports archived0 independent sourcesupdated 10/19/2026, 08:00:00

Synthesis & Latest Updates

vLLM reduced the downtime of Elastic Expert Parallelism (EP) scaling from tens of seconds to under 2 seconds by leveraging asynchronous preparation and NIXL EP. This optimization significantly enhances service continuity for MoE models like DeepSeek V4 Pro during dynamic resource adjustments.

LATEST/vLLM reduced the downtime of Elastic Expert Parallelism (EP) scaling from tens of seconds to under 2 seconds by leveraging asynchronous preparation and NIXL EP. This optimization significantly enhances service continuity for MoE models like DeepSeek V4 Pro during dynamic resource adjustments.

Coverage timeline

Total 1 reports · Latest first

vLLM 官方博客T1Oct 19, 08:00·78 pts
vLLM Achieves Near-Zero-Downtime Scaling in Elastic Expert Parallelism
- Reduced scale-up downtime from 100s to 0.82s (16 to 20 GPUs) using async prep and NIXL EP
- Reduced scale-down downtime from 56s to 1.96s (back to 16 GPUs)
- Enabled dynamic adjustment of expert parallelism for MoE models without service restart
View Related Reports & Analysis Original Source