vllm-project/vllm v0.19.0

Key points

• This release features 448 commits from 197 contributors (54 new)!1
• Zero-bubble async scheduling + speculative decoding: Async scheduling now supports speculative decoding with zero-bubble overlap, significantly improving throughput (#32951).1
• Model Runner V2 maturation: MRV2 gains piecewise CUDA graphs for pipeline parallelism (#35162), spec decode rejection sampler with greedy/logprobs support (#37238, #37237), multi-modal embeddings for spec decode (#36097), streaming inputs (#37028), and EPLB support (#37488).1
• General CPU KV cache offloading: A simple yet general CPU KV cache offloading mechanism for V1, with pluggable cache policy and block-level preemption handling (#37160, #37874, #34805, #36642, #37853).1

Sources (1)

• [1]vllm-project/vllm v0.19.0
  GitHub: vllm-project/vllm · Apr 3, 02:19 AM

  We recommend using pre-built docker image `vllm/vllm-openai:gemma4` for out of box usage.

  This release features 448 commits from 197 contributors (54 new)!

Extractive summary: sentences quoted from the sources.