发生了什么: PD Serving of Qwen3.8-2.4T

vLLM achieved 5000 tokens per GPU throughput and 180 generated tokens per user latency serving Qwen3.8-2.4T on GB300 clusters.

为何重要: PD Serving of Qwen3.8-2.4T

Engineers optimizing large language model serving performance need reproducible tuning methodologies for disaggregated serving configurations.

Results are specific to Qwen3.8-2.4T on GB300 hardware; CUDA graph memory estimation often overallocates, reducing actual KV cache availability.

新闻

vLLM · vLLM Team · 发布于 2026年9月21日 · 阅读约3分钟

发生了什么: vLLM achieved 5000 tokens per GPU throughput and 180 generated tokens per user latency serving Qwen3.8-2.4T on GB300 clusters.
为何重要: Engineers optimizing large language model serving performance need reproducible tuning methodologies for disaggregated serving configurations.
注意: Results are specific to Qwen3.8-2.4T on GB300 hardware; CUDA graph memory estimation often overallocates, reducing actual KV cache availability.

每个模式都讲清楚技术如何运作、何时值得投入，以及在哪里会失效。

每周一个模式、一个权衡、一个生产事故案例。为构建智能体系统的人准备的每周简报。

每周一封邮件,一键退订。您的地址仅用于发送简报。