Observed facts: vLLM says it runs on Vera Rubin NVL72 today and provides daily container builds. The post names DeepSeek, Moonshot AI, Z.ai, and MiniMax among supported models. It says Blackwell kernels are compatible with Rubin and identifies Rubin-tuned attention, GEMM, and MoE kernels through FlashInfer 0.7.0. It also describes applying CUDA 13.4 locality domains to split MoE weights so compute units access nearby memory. The reported 7.8× per-GPU throughput versus GB200 NVL72 is from AgentX at matched interactivity.
Mechanism: The initial path reuses a compatible kernel family, while hardware-specific kernels and locality-aware weight placement target bottlenecks exposed by Rubin. That distinction matters: basic bring-up broadens what can run, while tuning seeks to extract the new platform’s performance characteristics.
Why now: vLLM’s preview presents day-0 support as available now through daily builds, making the announcement a deployment signal rather than only a future roadmap item.
Implication: Builders evaluating Rubin can begin testing workloads and deployment paths on the runtime, but should treat benchmark claims as workload-specific rather than a forecast of their own results.
Evidence boundary: The source is a project preview with early results; it does not establish independent reproduction, production-wide availability, or comparable gains across workloads.
Watch: Whether later releases document stable support, reproducible benchmarks across workloads, and the operational requirements for locality-domain enablement.
The distinction between compatibility and tuning is the useful analytical record here: it separates an initial software bring-up milestone from evidence that the hardware yields repeatable serving gains. No prior New Runtime record was supplied for this branch.