vLLM adds Vera Rubin NVL72 support, citing 7.8x throughput over GB200 NVL72
The vLLM team, working with Inferact, Red Hat and NVIDIA, says vLLM now runs on NVIDIA Vera Rubin NVL72 with daily container builds (vllm/vllm-openai:cu134-nightly on CUDA 13.4 and PyTorch 2.15) and can serve models including DeepSeek, Kimi, GLM and MiniMax. Because Rubin continues the Blackwell architecture family (sm100f kernels can run unmodified), early results claim up to 7.84x throughput per GPU versus GB200 NVL72 on SemiAnalysis AgentX with MiniMax M3 at matched interactivity, and up to 3.7x VLM throughput versus GB300 NVL72 in MLPerf Inference v6.1 using Qwen3-VL-235B-A22B. Rubin NVL72 is described as offering 5x NVFP4 FLOPS, about 2.4x HBM bandwidth and 1.7x sixth-generation NVLink bandwidth over GB200 NVL72, while vLLM uses FlashInfer 0.7.0 Rubin-tuned kernels and CUDA 13.4 locality domains to speed up MoE decode by roughly 1.2x in preliminary tests. All figures are early, self-reported results from the collaborating teams.
For teams planning to deploy open-weight models on the newest NVIDIA hardware, it signals the serving stack is largely usable, though the published numbers are early and self-reported.