vLLM 针对 Vera Rubin NVL72 优化 DeepSeek-V4.1-Flash，吞吐达 GB300 的 58 倍
vLLM 在新一代 NVIDIA 机架上把 DeepSeek-V4.1-Flash 跑到了每 GPU 170 万 tokens/s，比上一代 GB300 快 58 倍。自己部署 DeepSeek 的团队可以关注这套优化配置，硬件升级后的推理成本账值得重算一遍。
vLLM 团队宣布在 NVIDIA Vera Rubin NVL72 平台上优化 DeepSeek-V4.1-Flash，在 SemiAnalysis AgentX 基准上达到 1.7 million tokens/s/GPU 的吞吐和 387 tokens/s/user 的 P90 速度，为 GB300 NVL72 同一指标的 58 倍。
具体数据覆盖整条吞吐-延迟曲线：峰值吞吐 2.52 million tokens/s/GPU，是 GB300 的 2.7 倍；单并发会话（TP 模式）下可达 1,810 tokens/s。优化点在于针对新一代 NVL72 硬件的推理调度与并行策略调优，使高并发场景下单个用户的响应速度不再随吞吐提升而明显下降。
对部署方而言，这类开源推理栈针对新硬件的适配直接决定了 DeepSeek 系列模型自托管服务的单位成本，与美国主流闭源 API 服务在价格上的差距可能进一步拉开。
We optimized DeepSeek-V4.1-Flash on vLLM for Vera Rubin NVL72 and reached 58× the throughput of GB300 NVL72 at 1.7 million tokens/s/GPU and 387 tokens/s/user on SemiAnalysis AgentX.
Across the curve: 🔹 2.52 million tokens/s/GPU at peak throughput (2.7× GB300) 🔹 1.7 million tokens/s/GPU at P90 387 tokens/s/user (58× GB300) 🔹 1,810 tokens/s at one concurrent session with TP
1/4
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include DeepSeek, vLLM, Vera Rubin, 开源模型.
Source attribution: curated from vLLM (@vllm_project). Cross-source corroboration: 11 independent sources have covered this development (including IT之家, Jim Fan, a16z), confirming industry-wide signal strength.
