vLLM v0.25.0 makes Model Runner V2 default, removes PagedAttention

vLLM v0.25.0 makes Model Runner V2 the default execution path for all dense models and deletes PagedAttention entirely. The release includes 558 commits from 232 contributors, with 64 new contributors. The legacy attention implementation is removed now that the V1 and MRv2 backends are the standard path. MRv2 builds on the quantized-model support from the previous release and adds support for EVS, realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding compatible with full CUDA graphs.

Engine and model changes

The Transformers modeling backend now matches native vLLM performance, according to the release notes. It gained FP8 MoE support, CUDA graph and embed scaling fixes, and migrations for GPTBigCode, Starcoder2, and RoBERTa. New models 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. GLM-5 and DeepSeek-V3.2 were added to the model zoo, with GLM-5.2 tuning, and MiniMax-M3 gained pipeline parallelism and NVFP4 support.

A new Streaming Parser Engine provides a unified tool-call and reasoning parsing framework, with a new Kimi k2.5, k2.6, and k2.7 parser and ports of seed_oss and DeepSeek V4. The Rust frontend continues to mature with HTTPS and mTLS support, a DP supervisor, and profiler control routes. Universal speculative decoding for heterogeneous vocabularies, called TLI, is new, along with DSpark and DFlash drafters. DFlash includes per-layer RMSNorm fusion, CPU support, and SWA support for MiMo.

Hardware and performance

For NVIDIA Blackwell hardware, FlashInfer fused all-reduce is tuned for world size 16 on GB300, and NVFP4 swizzled-scale zero-init was restored to recover decode throughput. New kernels include Helion fused_qk_norm_rope and silu_and_mul_per_block_quant, plus a Triton MLA logits workspace. The GLM-5.2 and DeepSeek path includes a fused_indexer_q_rope_quant Triton kernel with 1.9 to 3.3 percent end-to-end throughput gain, reduce-scatter MoE all-reduce with 3.1 to 3.2 percent gain, and a token_to_req_indices cache for DSv4 with a 5 to 6 times kernel speedup.

AMD ROCm moved to torch 2.11 stable ABI, with AITER FlashAttention MLA prefill backend, fused shared-expert for GLM-4.5, 6, and 7 and MiniMax-M3, and INT3 quantization for quickreduce. Intel XPU gained a W8A8 FP8 linear kernel with multi-granularity quant and pipeline-parallel accuracy fixes. CPU support includes accelerated unquantized MoE for AArch64, a macOS Apple Silicon hang fix via OpenMP, compressed-tensor w8a8 int8 MoE, and chunked prefill plus prefix caching for Qwen3.5. RISC-V has an RVV path for W4A8 INT4 GEMM and BF16 on VLEN 256 hardware, and POWER gained fp16 support.

Serving and quantization

Sequence parallelism now works without requiring data parallelism, with 1.9 to 5.0 percent end-to-end throughput improvement. NCCL symmetric memory extends to AllGather and ReduceScatter, FlashInfer all-reduce defaults to MNNVL on single node, and a fault-tolerance backend detects all2all peer faults. PD disaggregation includes a secondary-tier implementation, Mooncake connector support for Qwen3.5 and DeepSeek-V4-Flash, and NIXL Mamba1 support. Data parallel features include throttling prefills based on local prefill work and rotating load-balancer tie-break to avoid engine bias.

Quantization adds 2, 3, 5, 6, and 7-bit pack-quantized weight-only inference, Triton INT4 per-token-head KV cache quantization, and NVFP4 fused weight dequantization in the MoE MLP Triton kernel. Marlin thread-tile padding extends to MoE with WNA16 plus FP8 or MXFP8, and FlashInfer MXINT4 MoE works for gated SiLU. The old FP8 online MoE quantization class is deprecated.

API and security

OpenAI compatibility adds a Responses API namespace for tools, a per-request timing metrics field on Chat and Completions responses, token offsets on render endpoints, and HTTP 422 for unprocessable image URLs. TorchCodec is a new video decoding backend. The CLI prints TTFT and TPS in vllm chat. Security fixes prevent image decompression-bomb OOM denial of service, an infinite loop in split_audio with NaN samples, and request-level GPU video backend selection. Removed models include Baichuan, Aquila, Grok, Tarsier, Tarsier2, AyaVision, MusicFlamingo, and Mantis.