vLLM v0.25.0 makes Model Runner V2 the default and drops PagedAttention

vLLM v0.25.0 is out, and Model Runner V2 is now the default for every dense model, not just the models that were opted in over the past few releases. The old PagedAttention implementation is gone entirely: deleted, not deprecated, not behind a flag. If you’re running anything that relied on the legacy attention path, v0.25.0 is the forcing function to deal with that.

The MRv2 promotion comes with real additions on top of the existing quantized-model support. Dynamic speculative decoding now works with full CUDA graphs, prefix caching lands for Mamba hybrid models, and realtime embeddings are supported. The Transformers modeling backend also caught up to native vLLM speed in this release and gained FP8 MoE support. New models in the zoo include Hy3, LLaVA-OneVision-2, and GLM-5/DeepSeek-V3.2. The release counts 558 commits from 232 contributors.

For local setups, the headline is the speculative decoding improvement. Dynamic spec decode with full CUDA graph support removes a friction point that made it harder to combine MRv2 with drafters on consumer hardware. Whether that translates to a meaningful throughput gain on a single 24GB card depends on the model and drafter combination, and the official benchmarks were run on GB200 hardware. Worth testing on your own setup before assuming the numbers carry over.