vLLM Adds Support for NVIDIA Vera Rubin NVL72 with 7.8x Throughput via Tuned Kernels

1 reports archived0 independent sourcesupdated 10/9/2026, 08:00:00

Synthesis & Latest Updates

vLLM officially supports the NVIDIA Vera Rubin NVL72 platform, achieving 7.8x throughput over GB200 NVL72. This gain stems from hardware upgrades (5x NVFP4 FLOPS, 2.4x HBM bandwidth) and Rubin-tuned kernels via FlashInfer 0.7.0, ensuring compatibility with models like DeepSeek and Kimi.

LATEST/vLLM officially supports the NVIDIA Vera Rubin NVL72 platform, achieving 7.8x throughput over GB200 NVL72. This gain stems from hardware upgrades (5x NVFP4 FLOPS, 2.4x HBM bandwidth) and Rubin-tuned kernels via FlashInfer 0.7.0, ensuring compatibility with models like DeepSeek and Kimi.

HEAT TRENDHourly heat curve

53 fully observed hours

Now

0.0

Peak

10.711:00

24h change

–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first

vLLM 官方博客T1·Oct 9, 08:0068 pts
vLLM Adds Support for NVIDIA Vera Rubin NVL72 with 7.8x Throughput via Tuned Kernels
- NVIDIA Vera Rubin NVL72 offers 5x NVFP4 FLOPS and ~2.4x HBM bandwidth compared to GB200 NVL72.
- vLLM achieves Day-0 support via Blackwell architecture compatibility and gains Rubin-tuned kernels through FlashInfer 0.7.0.
- Supports models like DeepSeek and MiniMax, offering advanced configurations such as FP8 KV Cache.
View Evidence Original Source