vLLM v0.25.0：MRv2 默认、移除 PagedAttention，新增多模型与优化

vLLM v0.25.0 Release Notes

Highlights

This release features 558 commits from 232 contributors (64 new)!

• Model Runner V2 is now the default for all dense models (#44443). Building on quantized-model support from the previous release, MRv2 is now the standard execution path, with new support for EVS (#46535), realtime embeddings (#46762), prefix caching for Mamba hybrid models (#42406), multimodal-prefix bidirectional attention (#46942), and dynamic speculative decoding compatible with full CUDA graphs (#45953).
• PagedAttention has been removed (#47361). The legacy attention implementation is deleted now that V1/MRv2 backends are the standard path.
• The Transformers modeling backend is now as fast as native vLLM (#47187), and gained FP8 MoE support (#46820), CUDA graph + embed scaling fixes (#48010), and migration of GPTBigCode/Starcoder2 (#30966) and RoBERTa (#47452).
• New models: LLaVA-OneVision-2 (#44785), Unlimited OCR (#46564, #47102), MOSS-Transcribe-Diarize (#47729), openai/privacy-filter (#41026), and Hy3 (#47192). GLM-5 / DeepSeek-V3.2 landed in the model zoo (#46808) with GLM-5.2 tuning, and MiniMax-M3 gained pipeline parallelism (#45810) and NVFP4 support (#46756).
• New Streaming Parser Engine (#46610) — a unified tool-call/reasoning parsing framework, with a new Kimi k2.5/k2.6/k2.7 parser and ports of seed_oss (#46314) and DeepSeek V4 (#45877). The Rust frontend continues to mature with HTTPS/mTLS (#45890), a DP supervisor (#47076), and profiler control routes (#46306).
• Universal speculative decoding for heterogeneous vocabularies (TLI) (#38174), plus new DSpark (#46995) and DFlash (#46770, #46853) drafters.

Model Support

• New models: LLaVA-OneVision-2 (#44785), Unlimited OCR (#46564) with a Triton R-SWA backend (#47102), MOSS-Transcribe-Diarize (#47729), openai/privacy-filter (#41026), Hy3 with token-suffix and JSON Schema array support (#47192).
• GLM-5 family: GLM-5 / DeepSeek-V3.2 added to the model zoo (#46808), GLM-5.2 FP32 gate (#47410), GLM MTP post-final-norm fix (#47448), GLM4V startup fix (#47155).
• MiniMax-M3: pipeline parallelism (#45810), streaming reasoning parsing (#45718), and tok_sparse_select from MSA replacing Triton kernels (#47502).
• Transformers backend: now as fast as native vLLM (#47187), FP8 MoE fix (#46820), embed scaling + CUDA graph fix (#48010), GPTBigCode/Starcoder2 (#30966) and RoBERTa (#47452) migration, M-RoPE mm_token_type_ids fix (#46552), tied-embedding lm_head.bias fix (#46835).
• Voxtral: migrated to mistral-common 1.11.5 audio API (#46705) and realtime token-feedback hang fix (#44461).
• Gemma family: Gemma4 sliding-window/FA4 attention fixes (#47217, #47332), Gemma4 MTP quant_config fix (#47091); DiffusionGemma tensor parallelism (#45719) and HF stability-window semantics (#45965).
• Other fixes: MiniCPM-V 4.6 language-backbone LoRA (#46740) and placeholder grid fix (#45918), pooled Whisper sliding-window sizing (#47071, #47437), Mamba/Mamba2 checkpoint-without-architectures crash fix (#46037), DeepSeek-V2 hidden-size and aux-hidden-state fixes (#46986, #46973).

Engine Core

• Model Runner V2: default for all dense models (#44443); EVS (#46535), realtime embeddings (#46762), Mamba hybrid prefix caching (#42406), multimodal-prefix bidirectional attention (#46942), cross-attention warmup/block-table fixes (#46753, #47308), Mamba2 crash fix (#47428), scheduling slot accounting (#46974), model-ref cleanup on shutdown (#47483), bounded memory for large-logprobs requests (#46746).
• Speculative decoding: universal spec decode for heterogeneous vocabularies (TLI) (#38174); DSpark drafter + speculators checkpoint support (#46995, #47093); DFlash backend selection (#46770), per-layer RMSNorm fusion (#46761), CPU support (#44029), SWA+DFlash for MiMo (#46104), Laguna XS.2.1 drafter (#46853); MTP for Bailing hybrid models (#44880); block verification for rejection sampling (#46781); reduced TP communication for draft tokens (#46448).
• Sleep mode: pluggable sleep-mode backend abstraction (RFC #34303, #44074) with communicator-agnostic capability flags (#47243).
• Attention: FlashAttention block-size restriction removed for hybrid models (#36701), FLASH_ATTN_MLA_SPARSE Hopper sparse-MLA backend (#46189), DCP + FP8 KV cache in MLA decode (#44044), XQA decode kernels (#43232).
• KV offloading: tiering metric plumbing (#45959), request lifecycle fix (#46284), batched lookup in C (#46713), LookupResult enum (#46363).
• Misc: VLLM_GPU_SYNC_CHECK env var (#44800), VRAM semaphore infrastructure (#44465), skip detokenization in online beam search (#46422), several int32-overflow fixes in sampler/attention kernels (#46560, #47383, #47671).

Hardware & Performance

• GLM-5.2 / DeepSeek: fused_indexer_q_rope_quant Triton kernel (1.9–3.3% E2E throughput) (#46862), reduce-scatter MoE all-reduce (3.1–3.2% E2E) (#46635), op fusion for GLM5/DSV3.2 (#46876), token_to_req_indices cache for DSv4 (5–6x kernel speedup) (#47474), better DSv4 MXFP8 kernel (#47229), redundant-op removal (#47198, #46651).
• NVIDIA/Blackwell: FlashInfer fused all-reduce tuned for world_size=16 on GB300 (#46392), restored NVFP4 swizzled-scale zero-init to recover Blackwell decode throughput (#45739), CuTeDSL/FA4-MLA warmup infrastructure (#46182), skip cooperative top-K on SM120 (#47164), B12x backend for non-gated MoEs (#43328).
• Kernels: Helion fused_qk_norm_rope (#44010) and silu_and_mul_per_block_quant (#43994), Triton MLA logits workspace (#46819), swap-AB optimization for fused MoE (#36559), vectorized fp32 moe_sum supporting any top-k (#46643), blocking CUDA events to avoid busy-polling the driver lock (#47081).
• AMD/ROCm: moved to torch 2.11 stable ABI (#47128); AITER FlashAttention MLA prefill backend ROCM_AITER_FA (#45033); fused shared-expert for GLM-4.5/6/7 (#44313) and MiniMax-M3 (#46474, #46545); AITER MoE optimization for DeepSeek-V4 (#46122); AITER custom all-reduce in CudaCommunicator (#46065); INT3 quantization for quickreduce (#45666).
• Intel XPU: W8A8 FP8 linear kernel with multi-granularity quant (#43645), pipeline-parallel accuracy fix (#47253), uniform-batch CUDA graph for FA2 (#46555), route mm_prefix models to Triton attention (#47688), C++ get_memory_info (#47134).
• CPU: accelerated unquantized MoE for AArch64 (#46353), macOS/Apple Silicon hang fix via OpenMP (#46769) and broken-install fix (#47457), compressed-tensor w8a8 int8 MoE (#42920), Mamba ShortConv (#35059), chunked prefill + prefix caching for Qwen3.5 (#46202), faster gelu via tanh AOR (#44639).
• RISC-V: RVV path for W4A8 INT4 GEMM (#45269), BF16 on VLEN=256 hardware (#45243), reduced LMUL pressure in INT4 LUT dequant (#47538). POWER: fp16 support on PowerPC (#46135).
• Platform: accelerator-agnostic get_memory_info (#44825).

Large Scale Serving & Distributed

• Sequence parallelism without requiring DP, 1.9–5.0% E2E throughput improvement (#47070).
• Distributed: NCCL symmetric memory extended to AllGather and ReduceScatter (#46703), FlashInfer all-reduce defaults to MNNVL on single node (#47219, #47589), fault-tolerance backend to detect all2all peer faults and prevent corrupted output (#43637).
• Data parallel: throttle prefills based on local prefill work (#46532), rotate load-balancer tie-break to avoid engine bias (#47420), DP supervisor via the Rust frontend (#47076), DP MTP hang fix (#40589).
• PD disaggregation: secondary-tier implementation (#42285), Mooncake connector GDN (Qwen3.5) + MLA (DeepSeek-V4-Flash) support (#46807), NIXL Mamba1 support (#45019), MultiConnector kv_transfer_params merging (#46777), usage field exposed for disaggregated serving (#42748).
• DCP: FlashInfer MLA support (#43729), FLASHINFER_MLA_SPARSE support (#46076), LSE log-base fixes (#47079); Mooncake parallelized KV load (#45971) and DCP>1 lookup fix (#46855).
• ROCm: stabilized high-throughput DBO for DP+EP (#46990), EPLB for Quark OCP MXFP4 MoE (#47220).

Quantization