vLLM 0.9.0 release notes

Highlights

This release features 649 commits, from 215 contributors (82 new contributors!)

vLLM has upgraded to PyTorch 2.7! (#16859) This is a breaking change for environment dependency.
- The default wheel has been upgraded from CUDA 12.4 to CUDA 12.8. We will distribute CUDA 12.6 wheel on GitHub artifact.
- As a general rule of thumb, our CUDA version policy follow PyTorch's CUDA version policy.
Enhanced NVIDIA Blackwell support. vLLM now ships with initial set of optimized kernels on NVIDIA Blackwell with both attention and mlp.
- You can use our docker image or install FlashInfer nightly wheel pip install https://download.pytorch.org/whl/cu128/flashinfer/flashinfer_python-0.2.5%2Bcu128torch2.7-cp38-abi3-linux_x86_64.whl then set VLLM_ATTENTION_BACKEND=FLASHINFER for better performance.
- Upgraded support for the new FlashInfer main branch. (#15777)
- Please checkout https://github.com/vllm-project/vllm/issues/18153 for the full roadmap
Initial DP, EP, PD support for large scale inference
- EP:
  - Permute and unpermute kernel for moe optimization (#14568)
  - Modularize fused experts and integrate PPLX kernels (#15956)
  - Refactor pplx init logic to make it modular (prepare for deepep) (#18200)
  - Add ep group and all2all interface (#18077)
- DP:
  - Decouple engine process management and comms (#15977)
- PD:
  - NIXL Integration (#17751)
  - Local attention optimization for NIXL (#18170)
  - Support multiple kv connectors (#17564)
Migrate docs from Sphinx to MkDocs (#18145, #18610, #18614, #18616. #18622, #18626, #18627, #18635, #18637, #18657, #18663, #18666, #18713)

Notable Changes

Removal of CUDA 12.4 support due to PyTorch upgrade to 2.7.
Change top_k to be disabled with 0 (still accept -1 for now) (#17773)
The seed is now set to 0 by default for V1 Engine, meaning that different vLLM runs now yield the same outputs even if temperature > 0. This does not modify the random state in user code since workers are run in separate processes unless VLLM_USE_V1_MULTIPROCESSING=0. (#17929, #18741)

Model Enhancements

Support MiMo-7B (#17433), MiniMax-VL-01 (#16328), Ovis 1.6 (#17861), Ovis 2 (#15826), GraniteMoeHybrid 4.0 (#17497), FalconH1* (#18406), LlamaGuard4 (#17315)
- Please install the development version of transformers (from source) to use Falcon-H1.
Embedding models: nomic-embed-text-v2-moe (#17785), new class of gte models (#17986)
Progress in Hybrid Memory Allocator (#17394, #17479, #17474, #17483, #17193, #17946, #17945, #17999, #18001, #18593)
DeepSeek: perf enhancement by moving more calls into cuda-graph region(#17484, #17668), Function Call (#17784), MTP in V1 (#18435)
Qwen2.5-1M: Implements dual-chunk-flash-attn backend for dual chunk attention with sparse attention support (#11844)
Qwen2.5-VL speed enhancement via rotary_emb optimization (#17973)
InternVL models with Qwen2.5 backbone now support video inputs (#18499)

Performance, Production and Scaling

Support full cuda graph in v1 (#16072)
Pipeline Parallelism: MultiprocExecutor support (#14219), torchrun (#17827)
Support sequence parallelism combined with pipeline parallelism (#18243)
Async tensor parallelism using compilation pass (#17882)
Perf: Use small max_num_batched_tokens for A100 (#17885)
Fast Model Loading: Tensorizer support for V1 and LoRA (#17926)
Multi-modality: Automatically cast multi-modal input dtype before transferring device (#18756)

Security

Prevent side-channel attacks via cache salting (#17045)
Fix image hash collision in certain edge cases (#17378)
Add VLLM_ALLOW_INSECURE_SERIALIZATION env var (#17490)
Migrate to REGEX Library to prevent catastrophic backtracking (#18454, #18750)

Features

CLI: deprecated=True (#17426)
Frontend: progress bar for adding requests (#17525), chat_template_kwargs in LLM.chat (#17356), /classify endpoint (#17032), truncation control for embedding models (#14776), cached_tokens in response usage (#18149)
LoRA: default local directory LoRA resolver plugin. (#16855)
Metrics: kv event publishing (#16750), API for accessing in-memory Prometheus metrics (#17010)
Quantization: nvidia/DeepSeek-R1-FP4 (#16362), Quark MXFP4 format (#16943), AutoRound (#17850), torchao models with AOPerModuleConfig (#17826), CUDA Graph support for V1 GGUF support (#18646)
Reasoning: deprecate --enable-reasoning (#17452)
Spec Decode: EAGLE share input embedding (#17326), torch.compile & cudagraph to EAGLE (#17211), EAGLE3 (#17504), log accumulated metrics(#17913), Medusa (#17956)
Structured Outputs: Thinking compatibility (#16577), Spec Decoding (#14702), Qwen3 reasoning parser (#17466), tool_choice: required for Xgrammar (#17845), Structural Tag with Guidance backend (#17333)
Transformers backend: named parameters (#16868), interleaved sliding window attention (#18494)

Hardwares

NVIDIA: cutlass support for blackwell fp8 blockwise gemm (#14383)
TPU: Multi-LoRA implementation(#14238), default max-num-batched-tokens (#17508), V1 backend by default (#17673), top-logprobs (#17072)
Neuron: NeuronxDistributedInference support (#15970), Speculative Decoding, Dynamic on-device sampling (#16357), Mistral Model (#18222), Multi-LoRA (#18284)
AMD: Enable FP8 KV cache on V1 (#17870), Tuned fused moe config for Qwen3 MoE on MI300X (#17535, #17530), AITER biased group topk (#17955), Block-Scaled GEMM (#14968), MLA (#17523), Radeon GPU use Custom Paged Attention (#17004), reduce the number of environment variables in command line (#17229)
Extensibility: Make PiecewiseBackend pluggable and extendable (#18076)

Documentation

Update quickstart and install for cu128 using --torch-backend=auto (#18505)
NVIDIA TensorRT Model Optimizer (#17561)
Usage of Qwen3 thinking (#18291)

Developer Facing

Benchmark: Add single turn MTBench to Serving Bench (#17202)
Usability: Decrease import time of vllm.multimodal (#18031)
Code Format: Code formatting using ruff format (#17656, #18068, #18400)
Readability:
- Configuration and arguments unification is now complete! (#17130, #17453, #17562)
- Update deprecated type hinting from Python 3.7 (#18056, #18130, #18132, #18129, #18073, #18072, #18126, #18128, #18057, #18058)
Process:
- Propose a deprecation policy for the project (#17063)
Testing: expanding torch nightly tests (#18004)

What's Changed

Support loading transformers models with named parameters by @wuisawesome in https://github.com/vllm-project/vllm/pull/16868
Add tuned triton fused_moe configs for Qwen3Moe by @mgoin in https://github.com/vllm-project/vllm/pull/17328
[Benchmark] Add single turn MTBench to Serving Bench by @ekagra-ranjan in https://github.com/vllm-project/vllm/pull/17202
[Optim] Compute multimodal hash only once per item by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17314
implement Structural Tag with Guidance backend by @mmoskal in https://github.com/vllm-project/vllm/pull/17333
[V1][Spec Decode] Make Eagle model arch config driven by @ekagra-ranjan in https://github.com/vllm-project/vllm/pull/17323
[model] make llama4 compatible with pure dense layers by @luccafong in https://github.com/vllm-project/vllm/pull/17315
[Bugfix] Fix numel() downcast in fused_layernorm_dynamic_per_token_quant.cu by @r-barnes in https://github.com/vllm-project/vllm/pull/17316
Ignore '<string>' filepath by @zou3519 in https://github.com/vllm-project/vllm/pull/17330
[Bugfix] Add contiguous call inside rope kernel wrapper by @timzsu in https://github.com/vllm-project/vllm/pull/17091
[Misc] Add a Jinja template to support Mistral3 function calling by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17195
[Model] support MiniMax-VL-01 model by @qscqesze in https://github.com/vllm-project/vllm/pull/16328
[Misc] Move config fields to MultiModalConfig by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17343
[Misc]Use a platform independent interface to obtain the device attributes by @ponix-j in https://github.com/vllm-project/vllm/pull/17100
[Fix] Documentation spacing in compilation config help text by @Zerohertz in https://github.com/vllm-project/vllm/pull/17342
[Build][Bugfix] Restrict setuptools version to <80 by @gshtras in https://github.com/vllm-project/vllm/pull/17320
[Model] Ignore rotary embed load for Cohere model by @ekagra-ranjan in https://github.com/vllm-project/vllm/pull/17319
Update docs requirements by @hmellor in https://github.com/vllm-project/vllm/pull/17379
[Doc] Fix QWen3MOE info by @jeejeelee in https://github.com/vllm-project/vllm/pull/17381
[Bugfix] Clean up MiniMax-VL and fix processing by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17354
pre-commit autoupdate by @hmellor in https://github.com/vllm-project/vllm/pull/17380
[Frontend] Support chat_template_kwargs in LLM.chat by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17356
Transformers backend tweaks by @hmellor in https://github.com/vllm-project/vllm/pull/17365
Fix: Spelling of inference by @a2q1p in https://github.com/vllm-project/vllm/pull/17387
Improve literal dataclass field conversion to argparse argument by @hmellor in https://github.com/vllm-project/vllm/pull/17391
[V1] Remove num_input_tokens from attn_metadata by @heheda12345 in https://github.com/vllm-project/vllm/pull/17193
[Bugfix] add qwen3 reasoning-parser fix content is None when disable … by @mofanke in https://github.com/vllm-project/vllm/pull/17369
fix gemma3 results all zero by @mayuyuace in https://github.com/vllm-project/vllm/pull/17364
[Misc][ROCm] Exclude cutlass_mla_decode for ROCm build by @tywuAMD in https://github.com/vllm-project/vllm/pull/17289
Enabling multi-group kernel tests. by @Alexei-V-Ivanov-AMD in https://github.com/vllm-project/vllm/pull/17115
[Docs] Propose a deprecation policy for the project by @russellb in https://github.com/vllm-project/vllm/pull/17063
[Doc][Typo] Fixing label in new model requests link in overview.md by @casinca in https://github.com/vllm-project/vllm/pull/17400
[TPU][V1][CI] Replace python3 setup.py develop with standard pip install --e on TPU by @NickLucche in https://github.com/vllm-project/vllm/pull/17374
[CI] Uses Python 3.11 for TPU by @aarnphm in https://github.com/vllm-project/vllm/pull/17359
[CI/Build] Add retry mechanism for add-apt-repository by @reidliu41 in https://github.com/vllm-project/vllm/pull/17107
[Bugfix] Fix Minicpm-O-int4 GPTQ model inference by @Isotr0py in https://github.com/vllm-project/vllm/pull/17397
Simplify (and fix) passing of guided decoding backend options by @hmellor in https://github.com/vllm-project/vllm/pull/17008
Remove Falcon3 2x7B from CI by @hmellor in https://github.com/vllm-project/vllm/pull/17404
Fix: Python package installation for opentelmetry by @dilipgb in https://github.com/vllm-project/vllm/pull/17049
[V1][Spec Decode] Apply torch.compile & cudagraph to EAGLE by @luyuzhe111 in https://github.com/vllm-project/vllm/pull/17211
Remove Bamba 9B from CI by @hmellor in https://github.com/vllm-project/vllm/pull/17407
[V1][Feature] Enable Speculative Decoding with Structured Outputs by @benchislett in https://github.com/vllm-project/vllm/pull/14702
[release] Always git fetch all to get latest tag on TPU release by @khluu in https://github.com/vllm-project/vllm/pull/17322
Truncation control for embedding models by @gmarinho2 in https://github.com/vllm-project/vllm/pull/14776
Update PyTorch to 2.7.0 by @huydhn in https://github.com/vllm-project/vllm/pull/16859
Improve configs - ModelConfig by @hmellor in https://github.com/vllm-project/vllm/pull/17130
Fix call to logger.info_once by @hmellor in https://github.com/vllm-project/vllm/pull/17416
Fix some speculative decode tests with tl.dot by @huydhn in https://github.com/vllm-project/vllm/pull/17371
Support LoRA for Mistral3 by @mgoin in https://github.com/vllm-project/vllm/pull/17428
[Intel GPU] [CI]Fix XPU ci, setuptools >=80.0 have build issue by @jikunshang in https://github.com/vllm-project/vllm/pull/17298
[Hardware][Intel GPU] Upgrade to torch 2.7 by @jikunshang in https://github.com/vllm-project/vllm/pull/17444
[Bugfix] Fix AttributeError: 'State' object has no attribute 'engine_client' by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17434
[MODEL ADDITION] Ovis2 Model Addition by @mlinmg in https://github.com/vllm-project/vllm/pull/15826
Make the _apply_rotary_emb compatible with dynamo by @houseroad in https://github.com/vllm-project/vllm/pull/17435
[Misc] Remove deprecated files by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17447
[V1][Bugfix]: vllm v1 verison metric num_gpu_blocks is None by @lengrongfu in https://github.com/vllm-project/vllm/pull/15755
[TPU][V1][CI] Update regression test baseline for v6 CI by @NickLucche in https://github.com/vllm-project/vllm/pull/17064
[Core] Prevent side-channel attacks via cache salting by @dr75 in https://github.com/vllm-project/vllm/pull/17045
[V1][Metrics] add support for kv event publishing by @alec-flowers in https://github.com/vllm-project/vllm/pull/16750
[Feature] The Qwen3 reasoning parser supports guided decoding by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17466
[Docs] Add command for running mypy tests from CI by @russellb in https://github.com/vllm-project/vllm/pull/17475
[Fix] Support passing args to logger by @aarnphm in https://github.com/vllm-project/vllm/pull/17425
[Bugfix] Fixed mistral tokenizer path when pointing to file by @psav in https://github.com/vllm-project/vllm/pull/17457
[V1] Allow turning off pickle fallback in vllm.v1.serial_utils by @russellb in https://github.com/vllm-project/vllm/pull/17427
[Docs] Update optimization.md doc by @mgoin in https://github.com/vllm-project/vllm/pull/17482
[BugFix] Fix authorization of openai_transcription_client.py by @hhy3 in https://github.com/vllm-project/vllm/pull/17321
[Bugfix][ROCm] Restrict ray version due to a breaking release by @gshtras in https://github.com/vllm-project/vllm/pull/17480
[doc] add install tips by @reidliu41 in https://github.com/vllm-project/vllm/pull/17373
doc: fix bug report Github template formatting by @davidxia in https://github.com/vllm-project/vllm/pull/17486
[v1][Spec Decode] Make sliding window compatible with eagle prefix caching by @heheda12345 in https://github.com/vllm-project/vllm/pull/17398
Bump Compressed Tensors version to 0.9.4 by @rahul-tuli in https://github.com/vllm-project/vllm/pull/17478
[Misc] Rename Audios -> Audio in Qwen2audio Processing by @alex-jw-brooks in https://github.com/vllm-project/vllm/pull/17507
[CI][TPU] Skip Multimodal test by @lsy323 in https://github.com/vllm-project/vllm/pull/17488
[Bugfix][ROCm] Fix import error on ROCm by @gshtras in https://github.com/vllm-project/vllm/pull/17495
[Bugfix] Temporarily disable gptq_bitblas on ROCm by @nlzy in https://github.com/vllm-project/vllm/pull/17411
[CI][TPU] Skip structured outputs+spec decode tests on TPU by @mgoin in https://github.com/vllm-project/vllm/pull/17510
[CI][Bugfix] Fix failing V1 Test due to missing 'cache_salt' arg by @mgoin in https://github.com/vllm-project/vllm/pull/17500
[CI/Build] Reorganize models tests by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17459
FIxing the AMD test failures caused by PR#16457 by @Alexei-V-Ivanov-AMD in https://github.com/vllm-project/vllm/pull/17511
[Build] Require setuptools >= 77.0.3 for PEP 639 by @russellb in https://github.com/vllm-project/vllm/pull/17389
[ROCm] Effort to reduce the number of environment variables in command line by @hongxiayang in https://github.com/vllm-project/vllm/pull/17229
[BugFix] fix speculative decoding memory leak when speculation is disabled by @noyoshi in https://github.com/vllm-project/vllm/pull/15506
[BugFix] Fix mla cpu - missing 3 required positional arguments by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/17494
Avoid overwriting vllm_compile_cache.py by @youngkent in https://github.com/vllm-project/vllm/pull/17418
[Core] Enable IPv6 with vllm.utils.make_zmq_socket() by @russellb in https://github.com/vllm-project/vllm/pull/16506
[Misc] Optimize the Qwen3_ReasoningParser extract_reasoning_content by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17515
Improve configs - ObservabilityConfig by @hmellor in https://github.com/vllm-project/vllm/pull/17453
[Bugfix][Benchmarks] Allow benchmark of deepspeed-mii backend to select a model by @tishizaki in https://github.com/vllm-project/vllm/pull/17285
[Frontend] Show progress bar for adding requests by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17525
[Misc] Clean up test docstrings and names by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17521
[FEAT] [ROCm]: Add Qwen/Qwen3-30B-A3B-FP8 fused moe config for MI300X by @tjtanaa in https://github.com/vllm-project/vllm/pull/17530
Fix more broken speculative decode tests by @huydhn in https://github.com/vllm-project/vllm/pull/17450
[doc] add streamlit integration by @reidliu41 in https://github.com/vllm-project/vllm/pull/17522
[FEAT] [ROCm]: Add Qwen/Qwen3-235B-A22B-FP8 TP4 triton fused moe config by @tjtanaa in https://github.com/vllm-project/vllm/pull/17535
[Feature][Frontend]: Deprecate --enable-reasoning by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17452
[ROCm] remove unsupported archs from rocm triton flash-attention supported list by @hongxiayang in https://github.com/vllm-project/vllm/pull/17536
[torch.compile] Add torch inductor pass for fusing silu_and_mul with subsequent scaled_fp8_quant operations by @SageMoore in https://github.com/vllm-project/vllm/pull/10867
[Misc] refactor example - cpu_offload_lmcache by @reidliu41 in https://github.com/vllm-project/vllm/pull/17460
[CI/Build] Remove awscli dependency by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17532
Move the last arguments in arg_utils.py to be in their final groups by @hmellor in https://github.com/vllm-project/vllm/pull/17531
[Model] Refactor Ovis2 to support original tokenizer by @Isotr0py in https://github.com/vllm-project/vllm/pull/17537
[ROCm] update installation guide to include build aiter from source instructions by @hongxiayang in https://github.com/vllm-project/vllm/pull/17542
[Misc]add configurable cuda graph size by @CXIAAAAA in https://github.com/vllm-project/vllm/pull/17201
[Bugfix] Fix lint error by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17547
[ROCM] Add gfx950 to the custom attention archs by @jpvillam-amd in https://github.com/vllm-project/vllm/pull/16034
Remove duplicate code from dbrx.py by @sstamenk in https://github.com/vllm-project/vllm/pull/17550
[Bug]change the position of cuda_graph_sizes in dataclasses by @CXIAAAAA in https://github.com/vllm-project/vllm/pull/17548
[Misc][Tools][Benchmark] Publish script to auto tune server parameters by @Chenyaaang in https://github.com/vllm-project/vllm/pull/17207
[V1][Spec Decode] Apply torch.compile & cudagraph to EAGLE3 by @zixi-qi in https://github.com/vllm-project/vllm/pull/17504
[Bugfix] Disable gptq_bitblas for <SM80 to fix GPTQ on V100/T4 by @mgoin in https://github.com/vllm-project/vllm/pull/17541
[Doc] note that not all unit tests pass on CPU platforms by @davidxia in https://github.com/vllm-project/vllm/pull/17554
[Attention] MLA move o_proj q_proj into cuda-graph region by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/17484
[CI] Actually run tests/kv_transfer/test_disagg.py in CI by @mgoin in https://github.com/vllm-project/vllm/pull/17555
Check if bitblas is installed during support check by @mgoin in https://github.com/vllm-project/vllm/pull/17572
[Misc] Continue refactoring model tests by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17573
Fix PixtralHF missing spatial_merge_size by @mgoin in https://github.com/vllm-project/vllm/pull/17571
Add pt_load_map_location to allow loading to cuda by @jerryzh168 in https://github.com/vllm-project/vllm/pull/16869
[Bugifx] Remove TritonPlaceholder from sys.modules by @Isotr0py in https://github.com/vllm-project/vllm/pull/17317
[Core] [Bugfix] Add Input Embeddings by @qthequartermasterman in https://github.com/vllm-project/vllm/pull/15428
[BugFix] Fix Memory Leak by @robertgshaw2-redhat in https://github.com/vllm-project/vllm/pull/17567
[Misc] Rename assets for testing by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17575
add more pytorch related tests for torch nightly by @yangw-dev in https://github.com/vllm-project/vllm/pull/17422
[doc] add the print result by @reidliu41 in https://github.com/vllm-project/vllm/pull/17584
Automatically tell users that dict args must be valid JSON in CLI by @hmellor in https://github.com/vllm-project/vllm/pull/17577
[Security] Fix image hash collision by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17378
Support W8A8 INT8 MoE for compressed-tensors by @mgoin in https://github.com/vllm-project/vllm/pull/16745
[doc] miss result by @reidliu41 in https://github.com/vllm-project/vllm/pull/17589
[Misc] Clean up input processing by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17582
[Bugfix] fix tmp_out and exp_sums dimensions by @hliuca in https://github.com/vllm-project/vllm/pull/17438
[BugFix][Attention] Fix sliding window attention in V1 giving incorrect results by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/17574
permute/unpermute kernel for moe optimization by @CalebDu in https://github.com/vllm-project/vllm/pull/14568
Add NVIDIA TensorRT Model Optimizer in vLLM documentation by @Edwardf0t1 in https://github.com/vllm-project/vllm/pull/17561
[Hardware][AMD] Improve OAM device ID + llama4 Maverick MOE tuning by @xw285cornell in https://github.com/vllm-project/vllm/pull/16263
[easy] Print number of needed GPUs in skip message by @zou3519 in https://github.com/vllm-project/vllm/pull/17594
fix typo in logging by @ehartford in https://github.com/vllm-project/vllm/pull/17605
[release] Add command to clean up Docker containers/images in TPU release machine by @khluu in https://github.com/vllm-project/vllm/pull/17606
[Neuron][Build] Require setuptools >= 77.0.3 for PEP 639 by @liangfu in https://github.com/vllm-project/vllm/pull/17603
Update test requirements to CUDA 12.8 by @22quinn in https://github.com/vllm-project/vllm/pull/17576
[Quantizaton] [AMD] Add support for running DeepSeek int8 w8a8 MoE on ROCm by @rasmith in https://github.com/vllm-project/vllm/pull/17558
[Frontend][TPU] Add TPU default max-num-batched-tokens based on device name by @Chenyaaang in https://github.com/vllm-project/vllm/pull/17508
[Build/CI] Upgrade CUTLASS to 3.9.1 by @tlrmchlsmth in https://github.com/vllm-project/vllm/pull/17602
[Bugfix][ROCm] Using device_type because on ROCm the API is still torch.cuda by @gshtras in https://github.com/vllm-project/vllm/pull/17601
[Core] Gate prompt_embeds behind a feature flag by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17607
[Bugfix] Fix broken Qwen2.5-omni tests by @Isotr0py in https://github.com/vllm-project/vllm/pull/17613
[Misc] V0 fallback for --enable-prompt-embeds by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17615
Add full API docs and improve the UX of navigating them by @hmellor in https://github.com/vllm-project/vllm/pull/17485
[Bugfix] Prioritize dtype in root config before checking text config by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17629
[Bugfix][Easy] Fix whitespace in shm_broadcast.py logging by @tlrmchlsmth in https://github.com/vllm-project/vllm/pull/17635
[Bugfix] fix KeyError on top logprobs are special tokens by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/17637
[Build/CI] Upgrade CUTLASS to 3.9.2 by @tlrmchlsmth in https://github.com/vllm-project/vllm/pull/17641
[Kernel] some optimizations for dense marlin and moe marlin by @jinzhen-lin in https://github.com/vllm-project/vllm/pull/16850
[Doc] Fix broken cuda installation doc rendering by @Isotr0py in https://github.com/vllm-project/vllm/pull/17654
Use git-path commit in hook by @thomasjpfan in https://github.com/vllm-project/vllm/pull/17616
[Benchmarks] Remove invalid option under V1 engine by @russellb in https://github.com/vllm-project/vllm/pull/17651
[BugFix] Increase timeout for startup failure test by @njhill in https://github.com/vllm-project/vllm/pull/17642
[TPU] Enable gemma3-27b with TP>1 on multi-chips. by @vanbasten23 in https://github.com/vllm-project/vllm/pull/17335
[TPU][V1] Add support for top-logprobs by @NickLucche in https://github.com/vllm-project/vllm/pull/17072
[Bugfix] LoRA - Retire unused maxnreg LoRA kernel argument by @varun-sundar-rabindranath in https://github.com/vllm-project/vllm/pull/17677
Update nm to rht in doc links + refine fp8 doc by @mgoin in https://github.com/vllm-project/vllm/pull/17678
[Model] Add GraniteMoeHybrid 4.0 model by @s3woz in https://github.com/vllm-project/vllm/pull/17497
[easy] Fix logspam on PiecewiseBackend errors by @zou3519 in https://github.com/vllm-project/vllm/pull/17138
[Bugfix] Fixed prompt length for random dataset by @Xarbirus in https://github.com/vllm-project/vllm/pull/17408
[Doc] Update notes for H2O-VL and Gemma3 by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/17219
[Misc] Fix ScalarType float4 naming by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/17690
Fix dockerfilegraph pre-commit hook by @hmellor in https://github.com/vllm-project/vllm/pull/17698
[Bugfix] Fix triton import with local TritonPlaceholder by @MengqingCao in https://github.com/vllm-project/vllm/pull/17446
[V1] Enable TPU V1 backend by default by @mgoin in https://github.com/vllm-project/vllm/pull/17673
[V1][PP] Support PP for MultiprocExecutor by @bigPYJ1151 in https://github.com/vllm-project/vllm/pull/14219
[v1] AttentionMetadata for each layer by @heheda12345 in https://github.com/vllm-project/vllm/pull/17394
[Feat] Add deprecated=True to CLI args by @aarnphm in https://github.com/vllm-project/vllm/pull/17426
[Docs] Use gh-file to add links to tool_calling.md by @windsonsea in https://github.com/vllm-project/vllm/pull/17709
[v1] Introduce KVCacheBlocks as interface between Scheduler and KVCacheManager by @heheda12345 in https://github.com/vllm-project/vllm/pull/17479
[doc] Add RAG Integration example by @reidliu41 in https://github.com/vllm-project/vllm/pull/17692
[Bugfix] Fix modality limits in vision language example by @Dar

These notes run past the length kept in the archive. The rest is on the publisher’s page.