vllm 0.8.5

v0.8.5

on Python PyPI

Highlights

This release features important multi-modal bug fixes, day 0 support for Qwen3, and xgrammar's structure tag feature for tool calling.

Model Support

• Day 0 support for Qwen3 and Qwen3MoE. This release fixes fp8 weight loading (#17318) and adds tuned MoE configs (#17328).
• Add ModernBERT (#16648)
• Add Granite Speech Support (#16246)
• Add PLaMo2 (#14323)
• Add Kimi-VL model support (#16387)
• Add Qwen2.5-Omni model support (thinker only) (#15130)
• Snowflake Arctic Embed (Family) (#16649)
• Accuracy fixes for Llama4 Int4 (#16801), chat template for Llama 4 models (#16428), enhanced AMD support (#16674, #16847)

V1 Engine

• Add structural_tag support using xgrammar (#17085)
• Disaggregated serving:
  • KV Connector API V1 (#15960)
  • Adding LMCache KV connector for v1 (#16625)
• Clean up: Remove Sampler from Model Code (#17084)
• MLA: Simplification to batch P/D reordering (#16673)
• Move usage stats to worker and start logging TPU hardware (#16211)
• Support FlashInfer Attention (#16684)
• Faster incremental detokenization (#15137)
• EAGLE-3 Support (#16937)

Features

• Validate urls object for multimodal content parts (#16990)
• Prototype support sequence parallelism using compilation pass (#16155)
• Add sampling params to v1/audio/transcriptions endpoint (#16591)
• Enable vLLM to Dynamically Load LoRA from a Remote Server (#10546)
• Add vllm bench [latency, throughput] CLI commands (#16508)

Performance

• Attention:
  • FA3 decode perf improvement - single mma warp group support for head dim 128 (#16864)
  • Update to lastest FA3 code (#13111)
  • Support Cutlass MLA for Blackwell GPUs (#16032)
• MoE:
  • Add expert_map support to Cutlass FP8 MOE (#16861)
  • Add fp8_w8a8 fused MoE kernel tuning configs for DeepSeek V3/R1 on NVIDIA H20 (#16753)
• Support Microsoft Runtime Kernel Lib for our Low Precision Computation - BitBLAS (#6036)
• Optimize rotary_emb implementation to use Triton operator for improved performance (#16457)

Hardwares

• TPU:
  • Enable structured decoding on TPU V1 (#16499)
  • Capture multimodal encoder during model compilation (#15051)
  • Enable Top-P (#16843)
• AMD:
  • AITER Fused MOE V1 Support (#16752)
  • Integrate Paged Attention Kernel from AITER (#15001)
  • Support AITER MLA (#15893)
  • Upstream prefix prefill speed up for vLLM V1 (#13305)
  • Adding fp8 and variable length sequence support to Triton FAv2 kernel (#12591)
  • Add skinny gemms for unquantized linear on ROCm (#15830)
  • Follow-ups for Skinny Gemms on ROCm. (#17011)

Documentation

• Add open-webui example (#16747)
• Document Matryoshka Representation Learning support (#16770)
• Add a security guide (#17230)
• Add example to run DeepSeek with Ray Serve LLM (#17134)
• Benchmarks for audio models (#16505)

Security and Dependency Updates

• Don't bind tcp zmq socket to all interfaces (#17197)
• Use safe serialization and fix zmq setup for mooncake pipe (#17192)
• Bump Transformers to 4.51.3 (#17116)

Build and testing

• Add property-based testing for vLLM endpoints using an API defined by an OpenAPI 3.1 schema (#16721)

What's Changed

• Improve configs - SchedulerConfig by @hmellor in #16533
• [Misc] remove warning if triton>=3.2.0 by @DefTruth in #16553
• [Misc] refactor examples by @reidliu41 in #16563
• [Misc] Update usage with mooncake lib for kv transfer by @ShangmingCai in #16523
• [fix]: Dockerfile.ppc64le fixes for opencv-python and hf-xet by @Shafi-Hussain in #16048
• [Bugfix] Multi-modal caches not acting like LRU caches by @DarkLight1337 in #16593
• [TPU][V1] Fix exponential padding when max-num-batched-tokens is not a power of 2 by @NickLucche in #16596
• Fix triton install condition on CPU by @hmellor in #16600
• s390x: Fix PyArrow build and add CPU test script for Buildkite CI by @Nash-123 in #16036
• [Model][VLM] Add Kimi-VL model support by @courage17340 in #16387
• [Hardware][TPU] Add torchvision to tpu dependency file by @lsy323 in #16616
• [DOC][TPU] Add core idea about avoiding recompilation after warmup by @yaochengji in #16614
• config check sleep mode support oot platforms by @celestialli in #16562
• [Core][Bugfix] Fix Offline MM Beam Search by @alex-jw-brooks in #16390
• [Kernel] moe wna16 marlin kernel by @jinzhen-lin in #14447
• [BugFix]: Update minimum pyzmq version by @taneem-ibrahim in #16549
• [Bugfix] Fix tests/kernels/test_mamba_ssm_ssd.py by @tlrmchlsmth in #16623
• [Bugfix] Fix broken GritLM model and tests (missing pooling_metadata) by @pooyadavoodi in #16631
• Add vllm bench [latency, throughput] CLI commands by @mgoin in #16508
• Fix vLLM x torch.compile config caching by @zou3519 in #16491
• [Misc] refactor argument parsing in examples by @reidliu41 in #16635
• [CI/Build] Fix LoRA OOM by @jeejeelee in #16624
• Add "/server_info" endpoint in api_server to retrieve the vllm_config.  by @Cangxihui in #16572
• [Kernel] Remove redundant Exp calculations by @DefTruth in #16123
• [Misc] Update compressed-tensors WNA16 to support zero-points by @dsikka in #14211
• [Misc] Enable vLLM to Dynamically Load LoRA from a Remote Server by @angkywilliam in #10546
• [Model] Add PLaMo2 by @Alnusjaponica in #14323
• [Bugfix] fix gpu docker image mis benchmarks dir by @lengrongfu in #16628
• [Misc] Modify LRUCache touch by @jeejeelee in #16689
• Disable remote caching when calling compile_fx by @zou3519 in #16611
• [Feature] add model aware kv ops helper by @billishyahao in #16020
• [ROCM] Bind triton version to 3.2 in requirements-built.txt by @SageMoore in #16664
• [V1][Structured Output] Move xgrammar related utils to backend_xgrammar.py by @shen-shanshan in #16578
• [CI] Cleanup additional_dependencies: [toml] for pre-commit yapf hook by @yankay in #16405
• [Misc] refactor examples series by @reidliu41 in #16708
• [Doc] Improve OOM troubleshooting by @DarkLight1337 in #16704
• [Bugfix][Kernel] fix potential cuda graph broken for merge_attn_states kernel by @DefTruth in #16693
• [Model] support modernbert by @xsank in #16648
• [Hardware] Add processor inputs to platform validation by @joerunde in #16680
• Improve error for structured output backend selection by @hmellor in #16717
• [Misc] Remove redundant comment by @jianzs in #16703
• Help user create custom model for Transformers backend remote code models by @hmellor in #16719
• [V1][Performance] Implement custom serializaton for MultiModalKwargs [Rebased] by @p88h in #16432
• [V1][Spec Dec Bug Fix] Respect Spec Dec Method Specification by @luyuzhe111 in #16636
• Adding vllm buildkite job for IBM Power by @AaruniAggarwal in #16679
• [V1][Frontend] Improve Shutdown And Logs by @robertgshaw2-redhat in #11737
• [rocm][V0] fix selection logic for custom PA in V0 by @divakar-amd in #16426
• [Bugfix] Update Florence-2 tokenizer to make grounding tasks work by @Isotr0py in #16734
• [Bugfix] Revert max_prompt_len validation for decoder-only models. by @davidheineman in #16741
• [V1] Remove log noise when idle by @russellb in #16735
• [Ray] Improve documentation on batch inference by @richardliaw in #16609
• [misc] ignore marlin_moe_wna16 local gen codes by @DefTruth in #16760
• [Doc] Add more tips to avoid OOM by @DarkLight1337 in #16765
• [doc] add open-webui example by @reidliu41 in #16747
• [Bugfix] Fix GLM4 model by @intervitens in #16618
• [Doc] Fix a 404 link in installation/cpu.md by @windsonsea in #16773
• [Misc] refactor examples series - lmcache by @reidliu41 in #16758
• Improve configs - TokenizerPoolConfig + DeviceConfig by @hmellor in #16603
• fix: hyperlink by @reidliu41 in #16778
• [Doc] Make sure to update vLLM when installing latest code by @DarkLight1337 in #16781
• [Doc] Document Matryoshka Representation Learning support by @noooop in #16770
• [Doc] Changed explanation of generation_tokens_total and prompt_tokens_total counter type metrics to avoid confusion by @insukim1994 in #16784
• [V1][Perf] Faster incremental detokenization by @njhill in #15137
• [Bugfix]Fix index out of range error in api server log by @WangErXiao in #16787
• [Kernel] Add fp8_w8a8 fused MoE kernel tuning configs for DeepSeek V3/R1 on NVIDIA H20 by @Ximingwang-09 in #16753
• [Model] use AutoWeightsLoader for olmoe,opt,orion,persimmon,phi3_small by @lengrongfu in #16548
• [TPU][V1] Fix padding recompilation when max-num-batched-tokens is not even by @NickLucche in #16726
• [V1][TPU] Enable Top K by @NickLucche in #15489
• [ROCM] enable aiter fused moe kernel for llama4 bf16 checkpoints by @sijiac in #16674
• [V1][Metrics] Fix http metrics middleware by @markmc in #15894
• [MLA] Simplification to batch P/D reordering by @njhill in #16673
• [P/D][V1] KV Connector API V1 by @ApostaC in #15960
• [Attention] Update to lastest FA3 code by @LucasWilkinson in #13111
• Add property-based testing for vLLM endpoints using an API defined by an OpenAPI 3.1 schema by @tarukumar in #16721
• [Doc] Improve help examples for --compilation-config by @DarkLight1337 in #16729
• [Misc] Update outdated note: LMCache now supports chunked prefill by @chaunceyjiang in #16697
• [V1][Structured Output] Minor modification to _validate_structured_output() by @shen-shanshan in #16748
• Add hardware print to TPU V1 test by @mgoin in #16792
• [BugFix] Accuracy fix for llama4 int4 - improperly casted scales by @LucasWilkinson in #16801
• Improve configs - MultiModalConfig + PoolerConfig + DecodingConfig by @hmellor in #16789
• [Misc] add collect_env to cli and docker image by @lengrongfu in #16759
• [ROCm] [Attention] Cleanup ROCm output passing by @ProExpertProg in #16431
• [Bugfix] fix pp for llama4 by @luccafong in #16746
• [Doc] add podman setup instructions for official image by @nathan-weinberg in #16796
• [Docs] Fix a link and grammar issue in production-stack.md by @windsonsea in #16809
• [Model] use AutoWeightsLoader for BigCode, GPT-J by @jonghyunchoe in #16823
• [Misc] Clean up Kimi-VL by @DarkLight1337 in #16833
• Fix nullable_kvs fallback by @hmellor in #16837
• [New Model]: Snowflake Arctic Embed (Family) by @noooop in #16649
• [Misc] refactor examples series - Chat Completion Client With Tools by @reidliu41 in #16829
• [Doc] Updated Llama section in tool calling docs to have llama 3.2 config info by @jmho in #16857
• publish neuron docker image by @omrishiv in #16733
• [Model][VLM] Add Qwen2.5-Omni model support (thinker only) by @fyabc in #15130
• [rocm][MI300] llama4 maverick fp8 moe config tp8 by @divakar-amd in #16847
• [Frontend] Add sampling params to v1/audio/transcriptions endpoint by @NickLucche in #16591
• [Misc] Benchmarks for audio models by @NickLucche in #16505
• [V1][Misc] stop update prefix cache stats when logs_stats is disabled by @vie-serendipity in #16460
• [Model] Refactor Phi-4-multimodal to use merged processor and support V1 by @Isotr0py in #15477
• [Model] Qwen2.5-Omni Cleanup by @ywang96 in #16872
• [VLM] Clean up models by @DarkLight1337 in #16873
• [doc] update hyperlink by @reidliu41 in #16877
• Log how much time loading a compiled artifact takes by @zou3519 in #16848
• Serialize tensors using int8 views by @p88h in #16866
• Improve configs - CacheConfig by @hmellor in #16835
• [easy] Pass compile_fx only the config patches by @zou3519 in #16845
• [Bugfix] Fix v1/spec_decode/test_ngram.py by @zixi-qi in #16895
• [CI/CD][V1] Add spec decode tests to CI by @WoosukKwon in #16900
• [Bugfix] Fix distributed bug in Qwen2.5-VL & Qwen2.5-Omni by @fyabc in #16907
• [Doc] Split dummy_processor_inputs() in Multimodal Docs by @alex-jw-brooks in #16915
• Restore buffers when wake up from level 2 sleep (#16564) by @fingertap in #16889
• [Misc] fix collect_env version parse by @wangxiyuan in #15267
• [Misc] Refactor platform to get device specific stream and event by @shen-shanshan in #14411
• [Bugfix] Fix GLM rotary_dim issue and support v1 by @Isotr0py in #16912
• Raise error for data-parallel with benchmark_throughput by @kartikx in #16737
• [XPU][Bugfix] minor fix for XPU by @yma11 in #15591
• [doc] install required python3-dev apt package by @davidxia in #16888
• [Doc] mention how to install in CPU editable mode by @davidxia in #16923
• [Core] Speed up decode by remove synchronizing operation in sampler by @chanh in #16436
• [V1][Spec Decode] Handle draft tokens beyond max_model_len by @WoosukKwon in #16087
• [TPU][V1] Implicitly adjust page size when there's SMEM OOM by @yaochengji in #16871
• Update Qwen1.5-MoE-W4A16-compressed-tensors.yaml by @mgoin in #16946
• [TPU][V1] Capture multimodal encoder during model compilation by @NickLucche in #15051
• [V1] V1 FlashInfer Attention by @mgoin in #16684
• [TPU][V1] Enable Top-P by @NickLucche in #16843
• [Doc] Remove unnecessary V1 flag by @DarkLight1337 in #16924
• [BugFix][Spec Decode] No in-place update to draft probs by @WoosukKwon in #16952
• [Bugfix]: fix issue with n>1 sampling on v1 requests overriding each other by @jeffrey-dot-li in #16863
• [ROCm] Add aiter tkw1 kernel for Llama4 fp8 by @kliuae in #16727
• [Misc] Remove the chunked prefill warning for LoRA by @jeejeelee in #16925
• [Kernel] Add expert_map support to Cutlass FP8 MOE by @varun-sundar-rabindranath in #16861
• [V1] Remove additional_config check by @wangxiyuan in #16710
• [Performance][ROCm] Add skinny gemms for unquantized linear on ROCm by @charlifu in #15830
• Support S3 Sharded loading with RunAI Model Streamer by @omer-dayan in #16317
• [Bugfix] Fix f-string for Python 3.9-3.11 by @DarkLight1337 in #16962
• [Doc] Update ai_accelerator/hpu-gaudi.inc.md by @windsonsea in #16956
• [Perf] Optimize _update_states for GPU model runner by @SnowCharmQ in #16910
• [Bugfix] Fix the issue where llm.generate cannot be called repeatedly after setting GuidedDecodingParams by @chaunceyjiang in #16767
• [Model] Use autoweightloader for mamba by @sfeng33 in #16950
• [V1] Remove pre-allocation for KV cache by @WoosukKwon in #16941
• [Kernel] Support Microsoft Runtime Kernel Lib for our Low Precision Computation - BitBLAS by @LeiWang1999 in #6036
• [BugFix] Fix incremental detokenization perf issue by @njhill in #16963
• [Doc] Improve documentation for multimodal CLI args by @DarkLight1337 in #16960
• [FEAT][ROCm] Integrate Paged Attention Kernel from AITER by @vllmellm in #15001
• [Misc] refactor example series by @reidliu41 in #16972
• [Bugfix] Fix distributed bug again in Qwen2.5-VL & Qwen2.5-Omni by @fyabc in #16974
• Improve configs - SpeculativeConfig by @hmellor in #16971
• [BugFix] Pass in correct VLLM config in FlashInfer backend (#13207) by @timzsu in #16973
• [Misc] Add S3 environment variables for better support of MinIO. by @chaunceyjiang in #16977
• [frontend] enhance tool_calls type check by @reidliu41 in #16882
• [FEAT][ROCm]: Support AITER MLA by @vllmellm in #15893
• Add assertion for no objects while hashing hf_config by @zou3519 in #16930
• Fencing Kernels Tests for enabling on AMD by @Alexei-V-Ivanov-AMD in #16929
• [BugFix] Remove default multiproc executor collective_rpc timeout by @njhill in #17000
• [Core][V1][TPU] Enable structured decoding on TPU V1 by @Chenyaaang in #16499
• [Bugfix] validate urls object for multimodal content parts by @gcalmettes in #16990
• add Dockerfile build vllm against torch nightly by @yangw-dev in #16936
• [Kernel][ROCM] Upstream prefix prefill speed up for vLLM V1 by @maleksan85 in #13305
• [V1][DP] More robust DP/EP dummy request coordination by @njhill in #16277
• [BugFix] Revert ROCm Custom Paged Attention Env Flag Check by @vllmellm in #17022
• Revert "[Misc] Add S3 environment variables for better support of MinIO." by @chaunceyjiang in #17021
• [misc] tune some env vars for GB200 by @youkaichao in #16992
• [INTEL-HPU][v0] Port delayed sampling to upstream by @xuechendi in #16949
• [doc] add download path tips by @reidliu41 in #17013
• [Bugfix] Triton FA function takes no keyword arguments by @vllmellm in #16902
• [V1] Avoid socket errors during shutdown when requests are in in-flight by @njhill in #16807
• [BugFix] llama4 fa3 fix - RuntimeError: scheduler_metadata must have shape (metadata_size) by @LucasWilkinson in #16998
• [Misc] Improve readability of get_open_port function. by @gitover22 in #17024
• [Bugfix] Fix AssertionError: skip_special_tokens=False is not supported for Mistral tokenizers by @chaunceyjiang in #16964
• [CI] Run v1/test_serial_utils.py in CI by @russellb in #16996
• Mistral-format support for compressed-tensors by @mgoin in #16803
• Categorize tests/kernels/ based on kernel type by @mgoin in #16799
• [Doc] Add top anchor and a note to quantization/bitblas.md by @windsonsea in #17042
• Ensure that pid passed to kill_process_tree is int for mypy by @hmellor in #17051
• [CI] Update structured-output label automation by @russellb in #17055
• Improve Transformers backend model loading QoL by @hmellor in #17039
• CacheConfig.block_size should always be int when used by @hmellor in #17052
• Use @property and private field for data_parallel_rank_local by @hmellor in #17053
• [Frontend] Support guidance:no-additional-properties for compatibility with xgrammar by @tjohnson31415 in #15949
• [BugFix][V1] Fix int32 token index overflow when preparing input ids by @sarckk in #16806
• [V1][Spec Decode] Always use argmax for sampling draft tokens by @WoosukKwon in #16899
• [CI/Build] workaround for CI build failure by @csy1204 in #17070
• [Quantization]add prefix for commandA quantized model by @CXIAAAAA in #17017
• [Minor] Use larger batch sizes for A100/B100/B200/MI300x by @WoosukKwon in #17073
• [Bugfix] Enable V1 usage stats by @mgoin in #16986
• More informative error when using Transformers backend by @hmellor in #16988
• Addendum Fix to support FIPS enabled machines with MD5 hashing by @sydarb in #17043
• [Bugfix][Core] add seq_id_to_seq_group clearing to avoid memory leak when s… by @zhangyuygss in #16472
• [V1] Update structured output by @reidliu41 in #16812
• [doc] update to hyperlink by @reidliu41 in #17096
• Add docs for runai_streamer_sharded by @omer-dayan in #17093
• [Chore] Remove Sampler from Model Code by @WoosukKwon in #17084
• Disable enforce_eager for V1 TPU sampler and structured output tests by @mgoin in #17016
• Simplify TokenizerGroup by @hmellor in #16790
• Fix OOT registration test by @hmellor in #17099
• [V1][PP] Optimization: continue scheduling prefill chunks by @ruisearch42 in #17080
• [Misc] Remove OLMo2 config copy by @Isotr0py in #17066
• Improve static type checking in LoRAModelRunnerMixin by @hmellor in #17104
• [V1][Structured Output] Clear xgrammar compiler object when engine core shut down to avoid nanobind leaked warning by @shen-shanshan in #16954
• [Frontend] Using matryoshka_dimensions control the allowed output dimensions. by @noooop in #16970
• Add missing rocm_skinny_gemms kernel test to CI by @mgoin in #17060
• [Misc] refactor example series - structured outputs by @reidliu41 in #17040
• [V1][Spec Decoding] Add num_drafts and num_accepted_tokens_per_position metrics by @markmc in #16665
• [CI] Add automation for the tool-calling github label by @russellb in #17118
• Updating builkite job for IBM Power by @AaruniAggarwal in #17111
• existing torch installation pip command fix for docs by @atilla00 in #17059
• Molmo Requirements by @Eyshika in #17026
• Add :markdownhelp: to EngineArgs docs so markdown docstrings render properly by @hmellor in #17124
• Improve configs - LoRAConfig + PromptAdapterConfig by @hmellor in #16980
• [Docs] Generate correct github links for decorated functions by @russellb in #17125
• Add collective_rpc to llm engine by @yinghai in #16999
• Add chat template for Llama 4 models by @maxdebayser in #16428
• [Misc] Add example to run DeepSeek with Ray Serve LLM by @ruisearch42 in #17134
• Better error message for missing mistral params.json by @mgoin in #17132
• Use custom address for listening socket by @jglaser in #15988
• [FEAT] [ROCm]: AITER Fused MOE V1 Support by @vllmellm in #16752
• [Attention] FA3 decode perf improvement - single mma warp group support for head dim 128 by @LucasWilkinson in #16864
• fix float16 support for kimi-vl by @zhouzaida in #17156
• [Doc] V1 : Update LoRA status by @varun-sundar-rabindranath in #17133
• [Docs] Fix True->true in supported_models.md by @mgoin in #17141
• Move missed SchedulerConfig args into scheduler config group in EngineArgs by @hmellor in #17131
• [Misc] Clean up redundant code in uniproc_executor.py by @lifuhuang in #16762
• [Bugfix][Misc] Use TritonPlaceholderModule to defensively import triton by @MengqingCao in #15099
• [Misc] Benchmark Serving Script Support Appending Results by @LucasWilkinson in #17028
• [Perf]Optimize rotary_emb implementation to use Triton operator for improved inference performance by @cynthieye in #16457
• [Bugfix] remove fallback in guided_json (int range, patterns) by @csy1204 in #16725
• [Quantization][FP8] Add support for FP8 models with input_scale for output projection and QK quantization by @rasmith in #15734
• [Doc] Add headings to improve gptqmodel.md by @windsonsea in #17164
• Only turn on FastIncrementalDetokenizer when tokenizers >= 0.21.1 by @houseroad in #17158
• [Doc] Add two links to disagg_prefill.md by @windsonsea in #17168
• [Doc] Move todo out of beam search docstring by @alex-jw-brooks in #17183
• [Bugfix] Fix mistral model tests by @DarkLight1337 in #17181
• [Bugfix] Fix Mistral ChatCompletionRequest Body Exception by @JasmondL in #16769
• Bump Transformers to 4.51.3 by @hmellor in #17116
• Use Transformers helper get_text_config() instead of checking for text_config by @hmellor in #17105
• [doc] update wrong hf model links by @reidliu41 in #17184
• [Misc] Inline Molmo requirements by @DarkLight1337 in #17190
• [Security] Use safe serialization and fix zmq setup for mooncake pipe by @russellb in #17192
• [V1] Move usage stats to worker and start logging TPU hardware by @dyli-google in #16211
• [Bugfix] Fix hybrid model tests by @DarkLight1337 in #17182
• Fix Python packaging edge cases by @tiran in #17159
• [BugFix][Frontend] Fix LLM.chat() tokenization by @njhill in #16081
• [V1][Spec Decode] EAGLE-3 Support by @benchislett in #16937
• [Misc] Refine ray_serve_deepseek example by @ruisearch42 in #17204
• [Bugfix] gemma[2,3] interleaved attention when sliding window is disabled by @heheda12345 in #17180
• [AMD][FP8][BugFix] Remove V1 check in arg_utils.py for FP8 since it is not necessary by @rasmith in #17215
• [v1] [P/D] Adding LMCache KV connector for v1 by @ApostaC in #16625
• [Bugfix] [pytorch] Patch AOTAutogradCache._get_shape_env by @jamesjwu in #17142
• [MISC][AMD] Add unused annotation to rocm kernel file by @houseroad in #17097
• [doc] add Anything LLM integration by @reidliu41 in #17216
• [Minor][Spec Decode] Add use_eagle to SpeculativeConfig by @WoosukKwon in #17213
• [Doc] Minor fix for the vLLM TPU setup page by @yarongmu-google in #17206
• [Minor][Models] Fix Return Types of Llama & Eagle by @WoosukKwon in #17220
• Allocate kv_cache with stride order by @wenscarl in #16605
• [ROCm][Misc] Follow-ups for Skinny Gemms on ROCm. by @charlifu in #17011
• [V1][Metrics] Allow V1 AsyncLLM to use custom logger by @liuzijing2014 in #14661
• [BugFix] Avoid race conditions in zero-copy tensor transmission by @njhill in #17203
• [CI/test] Fix Eagle Correctness Test by @WoosukKwon in #17209
• [Core] Remove prompt string from engine core data structures by @njhill in #17214
• [Bugfix] Fix missing int type for -n in multi-image example by @Isotr0py in #17223
• [Bugfix] Fix standard models tests by @DarkLight1337 in #17217
• [Hardware][Intel-Gaudi] Update hpu-extension and update bucketing system for HPU device by @adobrzyn in #17186
• [V1] Add structural_tag support using xgrammar by @russellb in #17085
• [BUGFIX] use random for NONE_HASH only when PYTHONHASHSEED not set by @andyxning in #17088
• [Chore] added stubs for vllm_flash_attn during development mode by @aarnphm in #17228
• [Docs] Update structured output doc for V1 by @russellb in #17135
• [Bugfix] fix error due to an uninitialized tokenizer when using skip_tokenizer_init with num_scheduler_steps by @junstar92 in #9276
• Disable the torch.compile cache checks when VLLM_DISABLE_COMPILE_CACHE=1 by @houseroad in #16573
• [MISC] rename interval to max_recent_requests by @andyxning in #14285
• [Bugfix] Fix Qwen2.5-Omni M-RoPE position ids generation by @imkero in #16878
• [Minor] Fix lint error in main branch by @WoosukKwon in #17233
• [CI/Build] remove -t for run-lm-eval-gsm-hf-baseline.sh by @reidliu41 in #16271
• Update test_flash_attn.py by @ShuaibinLi in #17102
• [Kernel][Triton][FP8] Adding fp8 and variable length sequence support to Triton FAv2 kernel by @rasmith in #12591
• [Misc] Make cached tokenizer pickle-compatible by @DarkLight1337 in #17048
• [Bugfix] Fix QWen2 VL multimodal mapping by @jeejeelee in #17240
• [Bugfix] Get a specific type of layer from forward context by @heheda12345 in #17222
• [MISC] Use string annotation types for class definitions by @jianzs in #17244
• [Misc] Change buckets of histogram_iteration_tokens to [1, 8, 16, 32, 64, 128, 256, 512, 1024, 2048, 4096, 8096] to represent number of tokens by @sfc-gh-zhwang in #17033
• [Bugfix] Fix Lora Name Parsing by @alex-jw-brooks in #17196
• [NVIDIA] Support Cutlass MLA for Blackwell GPUs by @kaixih in #16032
• [Feature] support sequence parallelism using compilation pass by @cascade812 in #16155
• [doc] Add feature status legend by @reidliu41 in #17257
• [Metrics] Fix minor inconsistencies in bucket progression by @DarkLight1337 in #17262
• [V1][Spec Decode] Make eagle compatible with prefix caching. by @LiuXiaoxuanPKU in #17137
• [BugFix] Fix vllm_flash_attn install issues by @LucasWilkinson in #17267
• [Bugfix] Fix missing ARG in Dockerfile for arm64 platforms by @lkm-schulz in #17261
• [Bugfix] Fix cutlass dispatch for fp8/int8 to properly invoke M<=16 c… by @Ther-LF in #16751
• [Bugfix] Fix Mistral3 spatial merge error by @mgoin in #17270
• [Doc] Fix wrong github link in LMCache examples by @KuntaiDu in #17274
• [Doc] small fix by @reidliu41 in #17277
• [Misc] Validate stop_token_ids contents by @njhill in #17268
• [Minor][Models] Pass partial_rotary_factor parameter to rope by @Eviannn in #17266
• [Core] Remove legacy input mapper/processor from V0 by @DarkLight1337 in #15686
• [Model] Add Granite Speech Support by @alex-jw-brooks in #16246
• Update tpu_worker.py 's typo by @idouba in #17288
• Add missing class docstring for PromptAdapterConfig by @hmellor in #17302
• [Bugfix] Add missing get_language_model to new MLLMs by @DarkLight1337 in #17300
• [doc] update wrong model id by @reidliu41 in #17287
• [Misc] Minor typo/grammar in platforms/interface.py by @NickLucche in #17307
• [Misc] Clean up Qwen2.5-Omni code by @DarkLight1337 in #17301
• [Docs] Add a security guide by @russellb in #17230
• Improve conversion from dataclass configs to argparse arguments by @hmellor in #17303
• Make name of compressed-tensors quant method consistent across vLLM by @hmellor in #17255
• Explicitly explain quant method override ordering and ensure all overrides are ordered by @hmellor in #17256
• [Security] Don't bind tcp zmq socket to all interfaces by @russellb in #17197
• [Chore] cleanup license indicators in light of SPDX by @aarnphm in #17259
• [BugFix] Fix cascade attention - RuntimeError: scheduler_metadata must have shape (metadata_size) by @LucasWilkinson in #17283
• [Bugfix] Fix moe weight losing all extra attrs after process_weights_after_loading. by @charlifu in #16854
• [Model] Qwen3 Dense FP8 Compat Fixes by @simon-mo in #17318

New Contributors

• @Nash-123 made their first contribution in #16036
• @celestialli made their first contribution in #16562
• @taneem-ibrahim made their first contribution in #16549
• @Cangxihui made their first contribution in #16572
• @angkywilliam made their first contribution in #10546
• @Alnusjaponica made their first contribution in #14323
• @xsank made their first contribution in #16648
• @jianzs made their first contribution in #16703
• @p88h made their first contribution in #16432
• @AaruniAggarwal made their first contribution in #16679
• @davidheineman made their first contribution in #16741
• @richardliaw made their first contribution in #16609
• @intervitens made their first contribution in #16618
• @windsonsea made their first contribution in #16773
• @insukim1994 made their first contribution in #16784
• @Ximingwang-09 made their first contribution in #16753
• @sijiac made their first contribution in #16674
• @tarukumar made their first contribution in #16721
• @nathan-weinberg made their first contribution in #16796
• @jmho made their first contribution in #16857
• @vie-serendipity made their first contribution in #16460
• @zixi-qi made their first contribution in #16895
• @fingertap made their first contribution in #16889
• @kartikx made their first contribution in #16737
• @davidxia made their first contribution in #16888
• @chanh made their first contribution in #16436
• @jeffrey-dot-li made their first contribution in #16863
• @sfeng33 made their first contribution in #16950
• @LeiWang1999 made their first contribution in #6036
• @timzsu made their first contribution in #16973
• @yangw-dev made their first contribution in #16936
• @gitover22 made their first contribution in #17024
• @csy1204 made their first contribution in #17070
• @sydarb made their first contribution in #17043
• @zhangyuygss made their first contribution in #16472
• @atilla00 made their first contribution in #17059
• @Eyshika made their first contribution in #17026
• @yinghai made their first contribution in #16999
• @jglaser made their first contribution in #15988
• @zhouzaida made their first contribution in #17156
• @lifuhuang made their first contribution in #16762
• @JasmondL made their first contribution in #16769
• @tiran made their first contribution in #17159
• @jamesjwu made their first contribution in #17142
• @wenscarl made their first contribution in #16605
• @liuzijing2014 made their first contribution in #14661
• @adobrzyn made their first contribution in #17186
• @andyxning made their first contribution in #17088
• @junstar92 made their first contribution in #9276
• @ShuaibinLi made their first contribution in #17102
• @cascade812 made their first contribution in #16155
• @lkm-schulz made their first contribution in #17261
• @Ther-LF made their first contribution in #16751
• @Eviannn made their first contribution in #17266
• @idouba made their first contribution in #17288

Full Changelog: v0.8.4...v0.8.5