vLLM v0.25.0 drops PagedAttention, makes Model Runner V2 default, adds multimodal and speculative decoding features

The vLLM project released version 0.25.0 with Model Runner V2 (MRv2) now the default execution path for all dense models, replacing the previous standard path and removing PagedAttention. The release notes state that MRv2 introduces support for EVS, realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding compatible with full CUDA graphs.

The legacy attention implementation has been deleted as V1/MRv2 backends have become the standard. The Transformers modeling backend is now reported to be as fast as native vLLM, with added support for FP8 MoE, fixes for CUDA graph and embed scaling, and migrations for GPTBigCode/Starcoder2 and RoBERTa models.

New models added in this release include LLaVA-OneVision-2, Unlimited OCR with a Triton R-SWA backend, MOSS-Transcribe-Diarize, openai/privacy-filter, and Hy3 with token-suffix and JSON Schema array support. The model zoo also added GLM-5 and DeepSeek-V3.2, with GLM-5.2 tuning and MiniMax-M3 gaining pipeline parallelism and NVFP4 support.

A new Streaming Parser Engine provides a unified tool-call and reasoning parsing framework, including a Kimi k2.5/k2.6/k2.7 parser and ports of seed_oss and DeepSeek V4. The Rust frontend matured with HTTPS/mTLS support, a DP supervisor, and profiler control routes.

Universal speculative decoding for heterogeneous vocabularies (TLI) was introduced, along with new DSpark and DFlash drafters. The release includes 558 commits from 232 contributors, including 64 new contributors.

Engine core changes include EVS and realtime embeddings support, Mamba hybrid prefix caching, multimodal-prefix bidirectional attention, and fixes for cross-attention warmup and block-table management. Speculative decoding now supports universal decoding for heterogeneous vocabularies, with DSpark and DFlash drafters and backend selection improvements.