vLLM v0.20.2 Ships Model Runner V2 — 56% Throughput Improvement on GB200 via GPU-Native Triton Kernels

GitHub·medium signal

The open-source LLM inference engine vLLM (80K+ GitHub stars) released v0.20.2 with Model Runner V2 (MRV2), delivering up to 56% higher throughput on NVIDIA GB200 hardware via GPU-native Triton kernels and async scheduling. Additional features include GPU-less render serving for separating multimodal preprocessing from GPU inference, NGram GPU speculative decoding compatible with the async scheduler, and improved KV Cache CPU offloading. FP8 inference is now standard for H100 and Blackwell GPUs.