VAULT2077
Vault 资讯瀑布官方2026.07.27 09:06 UTC+8

vllm-project/vllm 发布 v0.26.0

vLLM v0.26.0 发布，411 个提交，新增 Inkling 模型系列、DeepSeek-V4 性能优化、fp32 lm_head、灵活注意力后端、KV 卸载增强、Rust 前端等。

中文处理结果

vLLM v0.26.0 发布说明

亮点

本版本包含来自 212 位贡献者（其中 61 位新贡献者）的 411 个提交！

  • 全新 Inkling 模型系列，附带完整支持栈：基础建模（#48799）、分段 CUDA 图支持（#48822）、Hopper FA4 相对注意力（#48858）、MTP=1 推测解码（#48869）、LoRA（#48884）以及标准 ModelOpt NVFP4 量化（#48990）。
  • 跨厂商的 DeepSeek-V4 性能提升：专用路由内核（2.94% E2E TPOT，#48660）、fused_topk_bias（内核性能提升 1.5–2 倍，#47463）、冗余重复/拷贝消除（1.8% E2E TPOT，#48137），外加 ROCm 双级压缩器用于 HCA 预填充（#47718）、稀疏解码/预填充优化（#48519、#48788、#46275），以及 AMD（#47419）和 XPU（#47677）上的 DSpark 推测解码。
  • 通过 head_dtype 为生成模型提供 fp32 lm_head（#48390），已扩展到 LoRA 路径（#48525）并提供 ROCm torch.mm 快速路径（#48688），提高生成头的精度。
  • 灵活的注意力后端：现在可以按 KV 缓存组选择注意力后端（#48012），滑窗支持现在是一个明确的后端能力（#48011）——增强对混合模型的支持。
  • KV 卸载与分层二级存储 进一步成熟：卸载指标（#45958、#47666、#47679）、分层所有事件处理（#46544、#47923）、带有工作负载身份的对象存储二级层（#47063、#47274、#48150）、DP 副本感知分层（#47987），以及包括 CPU 卸载在内的编码器缓存（EC）连接器（#42433、#47423）。
  • Rust 前端 增加多模态视频（#47959）和音频（#48554）、Seed-OSS 工具解析器（#47741），以及原生 vllm-bench 移植（#48107）。
  • Transformers 5.13.0（#47867），更多模型迁移到 Transformers 建模后端：Olmo/Olmo2（#48100）、MistralLarge3（#48153）和 HunyuanVL（#47872）。

模型支持

  • 新模型：Inkling 系列（#48799、#48822、#48858、#48869、#48884、#48990）、BertForMaskedLM（#48463）、RobertaForTokenClassification / XLMRobertaForTokenClassification（#47991）、LongCat-Flash-Lite n-gram 嵌入（#47857）、Cosmos3 Edge Reasoner（#48291）和 Cosmos3-Super 注册（#48211）、TranslateGemma-12b-it（#41599）。
  • Transformers 后端迁移：Olmo/Olmo2（#48100）、MistralLarge3 迁移到 AutoWeightsLoader（#48153）、面向 transformers 5.13 的 HunyuanVL 原生 transformers 处理器（#47872）。
  • GLM5.2：将 MoE 序列并行支持迁移到非 torch 编译路径（#47881）。
  • LoRA：针对 BF16 模型的 FlashInfer MoE LoRA（#48632）、LlavaNextVideo 中 tower/connector 的 LoRA（#48594）、LoRA 路径上的 fp32 lm_head（#48525）、优化后的 TrtLlmLoRAExperts（#48759）。
  • 多模态：当 TP 不可用时自动回退到 ViT 数据并行（#49046）。
  • 修复：在 torch.compile 下分块预填充的池化分数已修正（#48901）。

原始正文

vllm-project/vllm released v0.26.0

vLLM v0.26.0 Release Notes

Highlights

This release features 411 commits from 212 contributors (61 new)!

  • New Inkling model family with a full support stack: base modeling (#48799), piecewise CUDA graph support (#48822), Hopper FA4 relative attention (#48858), MTP=1 speculative decoding (#48869), LoRA (#48884), and standard ModelOpt NVFP4 quantization (#48990).
  • DeepSeek-V4 performance push across vendors: a specialized routing kernel (2.94% E2E TPOT, #48660), fused_topk_bias (1.5–2x kernel, #47463), and redundant repeat/copy removal (1.8% E2E TPOT, #48137), plus ROCm two-stage compressor for HCA prefill (#47718), sparse decode/prefill optimizations (#48519, #48788, #46275), and DSpark speculative decoding on AMD (#47419) and XPU (#47677).
  • fp32 lm_head for generation models via head_dtype (#48390), extended to the LoRA path (#48525) and given a ROCm torch.mm fast path (#48688), improving accuracy for generation heads.
  • Flexible attention backends: the attention backend can now be selected per KV-cache group (#48012), and sliding-window support is now an explicit backend capability (#48011) — improving support for hybrid models.
  • KV offloading & tiered secondary storage matured substantially: offloading metrics (#45958, #47666, #47679), tier-owned event handling (#46544, #47923), object-store secondary tier with workload identity (#47063, #47274, #48150), DP-replica-aware tiering (#47987), and encoder-cache (EC) connectors including CPU offloading (#42433, #47423).
  • Rust frontend gained multimodal video (#47959) and audio (#48554), a Seed-OSS tool parser (#47741), and a native vllm-bench port (#48107).
  • Transformers 5.13.0 (#47867) with more models migrated to the Transformers modeling backend: Olmo/Olmo2 (#48100), MistralLarge3 (#48153), and HunyuanVL (#47872).

Model Support

  • New models: Inkling family (#48799, #48822, #48858, #48869, #48884, #48990), BertForMaskedLM (#48463), RobertaForTokenClassification / XLMRobertaForTokenClassification (#47991), LongCat-Flash-Lite n-gram embedding (#47857), Cosmos3 Edge Reasoner (#48291) and Cosmos3-Super registration (#48211), TranslateGemma-12b-it (#41599).
  • Transformers backend migrations: Olmo/Olmo2 (#48100), MistralLarge3 to AutoWeightsLoader (#48153), HunyuanVL native transformers processor for transformers 5.13 (#47872).
  • GLM5.2: migrate MoE sequence-parallel support to the non-torch-compiled path (#47881).
  • LoRA: FlashInfer MoE LoRA for BF16 models (#48632), LoRA for tower/connector in LlavaNextVideo (#48594), fp32 lm_head on the LoRA path (#48525), optimized TrtLlmLoRAExperts (#48759).
  • Multimodal: automatic fallback to ViT data parallelism when TP is unavailable (#49046).
  • Fixes: correct pooling scores for chunked prefill under torch.compile (#48901).

Engine Core

  • fp32 lm_head for generation models via head_dtype (#48390); lower memory for capturing large CUDA graph sizes (#48483); opt-in persistence and reuse of the memory-profiling result across boots (#47388); improved InstantTensor loading (#46868).
  • Attention: select a different attention backend per KV-cache group (#48012); sliding-window as an explicit backend capability (#48011); KV-cache layout refactor packing K/V into the content dim across backends (#44455); MRV2 virtual-batch PCP for MLA (#46570).
  • Speculative decoding: runtime draft weight update (#46725), hybrid (SWA + full attention) DFlash drafters (#47914), SWA support for qwen-eagle3 (#47568), Gemma4-12B DSpark draft model (#47216), DSv4 DSpark on AMD (#47419), separate kv_cache_dtype for speculative_config (#48787).
  • KV offloading: basic offloading metrics (#45958), split CPU cache usage into read/write gauges (#47666) and tiering-lookup-delay into sync/async histograms (#47679), tier-owned event handling and BlockStored events (#46544, #47923), object-store secondary tier with workload identity (#47063, #47274, #48150), DP-replica-aware tiering (#47987), blocks_per_chunk config for heterogeneous KV groups (#48878), P2P default host/port env vars (#47636).
  • Caching: partial prefix-cache hit for hybrid models (#46384), selective hybrid cache retention (#47782), report prefix-cache-reused blocks in full report mode (#45261).
  • Reasoning: optimize TPOT for thinking budget when used with speculative decoding (#46662).
  • RLHF: stateful trainer-send abstractions (#48042).
  • Fixes: host memory leak from undrained new_block_ids (#44490), DSv3.2 + MTP + sequence-parallel accuracy (#48036).

Hardware & Performance

  • DeepSeek-V4: specialized routing kernel (2.94% E2E TPOT, #48660), fused_topk_bias 1.5–2x (#47463), redundant repeat/copy removal (1.8% TPOT, #48137).
  • MoE router GEMMs: BF16x3 router GEMM (#47973), FP32 router GEMV (#48335), generic CuteDSL LL BF16 router GEMM (#42562); TRTLLM BF16 MoE modular kernel (#45182); write FlashInfer combine into final output (#47156).
  • Qwen: fuse more RMSNorm + all-reduce in Qwen3.5 (#46998), replace MoE all-reduce with reduce-scatter (#47006), Qwen3.5 H20 optimization (#48350), expand Triton warmup coverage (#47546).
  • MLA: dense MHA path for short sparse-MLA sequences (#47327); MiniMax-M3 long-context decode indexer on sm100 (#48582).
  • Kernels: CUDA kernel for ReLUSquaredActivation / relu^2 (#39058), Helion kernel lazy registration (#48264), vectorize _copy_mamba_state_block to uint64 (#48110), stop upcasting logits to fp32 in the sampler (#48641).
  • ROCm: fp32 head_dtype torch.mm fast path (#48688), DSv4 two-stage compressor kernel (#47718), sparse decode/prefill optimizations (#48519, #48788, #46275), DSv3.2 sparse MLA KV-split heuristic (#46832) and MTP CUDA-graph mode (#45149), MXFP8 GEMM for MiniMax-M3 (#46117), AITER sparse paged attention + spec decode for MiniMax-M3 (#47287, #47984), MiniMax-M2 fused QK-norm + all-reduce via AITER (#44849), HybridW4A16 linear kernel (#40977), Qwen3-30B-A3B QK-Norm+RoPE+KV runtime fusion (#42749).
  • XPU: batch-invariant kernels (#41934), HND KV layout support (#47975), DSpark spec decode for DSv4 (#47677), nightly/release image publishing (#47880, #48126).
  • CPU: DFlash speculative decoding for GDN models on CPU (#46090), s390x NUMA topology (#40714), native macOS arm64 CPU wheel builds (#48289); POWER VSX math function optimization (#47321) and IBM Power docker builds using prebuilt wheels (#46017).
  • Distributed fusion: FlashInfer MNNVL all-reduce RMS quant fusion (#48064).
  • Build/autotune: arm64 Blackwell SM10x/SM110 image builds (#48041); skip CuTeDSL fp4_gemm autotuning by default (#48268).

Large Scale Serving & Distributed

  • Decode Context Parallel (DCP): hybrid attention support (#40996), DCP + Eagle for Tokenspeed MLA backends (#48180).
  • PD disaggregation: NIXL pipeline-parallel prefill in push mode (#45880).
  • Encoder-cache connectors: EC transfer params (#42433) and CPU-offloading EC connector (#47423).

Quantization

  • Humming w[2-7]a[4,8] weight-only inference with compressed-tensors (#46390); int4 quantization for the emulation MoE backend (#48451); INT2 XPU weight-only quant linear (#47521).
  • NVFP4/MXFP4: nvfp4_per_token online MoE quantization (#48538), CuTe-DSL FlashInfer MXFP4 quantization (#48417); bounded peak memory when repacking FP4 MoE weights for Marlin (#47851) and for NVFP4 MoE weight loading (#46276).
  • MLA: kv_cache_dtype_skip_layers support (#47309).
  • ROCm: HybridW4A16 linear kernel (#40977).

API & Frontend

  • Rust frontend: multimodal video (#47959) and audio (#48554), Seed-OSS tool parser (#47741), native vllm-bench port (#48107), continue_final_message handling with renderer sentinel (#47844).
  • OpenAI compatibility: bad_words in /v1/completions (#46793), expose logprob_token_ids on Python OpenAI endpoints (#43463), include_reasoning param for non-Harmony models (#44301), populate num_cache_creation_tokens on Messages responses (#48535).
  • Endpoint plugins framework (#47454); /abort_requests on the RLHF dev API router (#47173); Deepstream video decoding backend (#42424); overlap preprocessing and computation for pooling models in offline inference (#47699).
  • UX: human-readable integers for more CLI args (#47608), CuTeDSL compilation progress bar (#48881), expanded GPU profiler config scope/annotations (#37524), log worker exit code when a process dies unexpectedly (#38641).
  • Stability/correctness: handle grammar compilation failures without crashing the engine (#47312), fix logprobs token-string collision from SentencePiece spaces (#48674).

Security

  • Replace diskcache to eliminate pickle deserialization (#44549).
  • Fix a concurrent sparse-invariant race that bypassed CVE remediation (#48583).
  • Add resource-bounds validation to derender endpoints (#47260); sanitize server file paths from validation error responses (#46415); bound the completion prompt list to prevent unbounded engine fan-out (#47845); guard lm-format-enforcer regex compilation with a timeout (#47595).

Dependencies

  • Transformers 5.13.0 (#47867), FlashInfer 0.6.14 (#47669), NIXL 1.3.1 (#47559), tpu-inference v0.24.0 (#47835), nvidia-cutlass-dsl 4.6.0 (#47442), vllm_xpu_kernels v0.1.11.1 (#48942).
  • FlashAttention 3 pinned to the torch stable-ABI commit (#47995); ABI-stable FlashMLA build (#48174).

Deprecations & Removals

  • Models removed: TeleChat (#47989), Persimmon and Fuyu (#48096).

New Contributors

Contributors

@mgoin, @yewentao256, @NickLucche, @njhill, @LucasWilkinson, @micah-wil, @khluu, @AndreasKaratzas, @WoosukKwon, @aoshen02, @BugenZhao, @vanshbhatia-amd, @jperezdealgaba, @jeejeelee, @hmellor, @tlrmchlsmth, @MatthewBonanni, @Change72, @chaojun-zhang, @gau-nernst, @gnovack, @ZJY0516, @reidliu41, @Yejing-Lai, @taneem-ibrahim, @Srinivasoo7, @Sunt-ing, @AmeenP, @zhenwei-intel, @LopezCastroRoberto, @matteso1, @yzong-rh, @stefankoncarevic, @Isotr0py, @djramic, @charlifu, @peizhang56, @giuseppegrossi, @drakosha, @NickCao, @benchislett, @Rohan138, @hickeyma, @muhammadfawaz1, @rasmith, @zixi-qi, @music-dino, @BWAAEEEK, @xianbaoqian, @wendyliu235, @atalman, @joerowell, @ErenAta16, @AlejandroParedesLT, @gcanlin, @liranschour, @omerpaz95, @Alex-ai-future, @tanpinsiang, @Fangzhou-Ai, @KKothuri, @zxd1997066, @akii96, @nemanjaudovic, @Etelis, @afierka-intel, @DaoyuanLi2816, @HDCharles, @tahsintunan, @xiaohongchen1991, @sagearc, @mikekg, @kliuae, @qli88, @arpera, @yushangdi, @edwinlim0919, @tjtanaa, @ariG23498, @lucifer1004, @netanel-haber, @kl527, @Rukhaiya2004, @ronensc, @guan404ming, @shaunkotek, @liulanze, @pierDipi, @eldarkurtic, @simon-mo, @robinguo23, @CienetStingLin, @RishabhSaini, @Sahil170595, @jasonlizhengjian, @jacklin78911-collab, @walterbm, @amd-ethany, @zqzten, @nicklasfrahm, @hongxiayang, @alexeldeib, @ManaEstras, @sungbin1015, @aoright, @Saddss, @vivek8123, @voipmonitor, @shawntsai, @almayne, @ilmarkov, @cleonard530, @kjiang249, @chaunceyjiang, @bigPYJ1151, @tsvikas, @deng451e, @tvirolai-amd, @zhewenl, @zihaomu, @ap9272, @staugust, @Yancey0623, @GongLei-HW, @albertoperdomo2, @guoriyue, @ViranjanPagar, @Functionhx, @XuZhou26, @MynameFelix, @larryli2-amd, @ashwing, @thisisjimmyfb, @robertgshaw2-redhat, @mayuyuace, @ibondarenko1, @zhejiangxiaomai, @vx120, @hugo-cen, @tanish-malekar, @zzt93, @guybd, @R3hankhan123, @mmangkad, @omera-nv, @yma11, @Gavin-Morris-04, @pavanimajety, @shanjiaz, @wenpengw-nv, @atalhens, @langzhao-netizen, @emricksini-h, @Zhenzhong1, @DanBlanaru, @mgehre-amd, @mwoodson, @wangxiyuan, @adsridhar, @hnt2601, @gangula-karthik, @tomerg-nvidia, @adhi29, @rishitdholakia13, @divakar-amd, @chaeminlim-mb, @joanvelja, @russellb, @janeyx99, @aarushjain29, @wjabbour, @mahadrehmann, @krishy91, @tzielinski-habana, @avalliappan-nvidia, @ruikangliu, @majian4work, @maxyanghu, @brijrajk, @lengrongfu, @Josephasafg, @elvircrn, @xiaguan, @ricky-chaoju, @iyastreb, @ovidiusm, @tuukkjs, @noooop, @samnordmann, @AndyDai-nv, @xiao-llm, @DiegoCao, @yuvalluria, @jhu960213, @woosebastian, @Debasish-87, @esmeetu, @hao-aaron, @zhangj1an, @wendadawen, @juliendenize, @passtoor-agi, @mosya415, @labAxiaoming, @devalshahamd, @wangxingda, @xuebwang-amd, @fuscof-ibm, @alexxu-roblox, @frida-andersson, @lishunyang12, @izhuhaoran

查看原始发布