vLLM v0.20.2 Ships Model Runner V2: 56% Throughput Improvement on GB200.

The open-source inference engine released v0.20.2 with GPU-native Triton kernels, async speculative decoding with zero-bubble overlap, and GPU-less render serving for separating multimodal preprocessing from GPU inference. FP8 is now standard for H100 and Blackwell. If you're serving LLMs in production and haven't upgraded, you're overpaying for compute.