vLLM 0.10.1 release notes

Highlights

v0.10.1 release includes 727 commits, 245 committers (105 new contributors).

NOTE: This release deprecates V0 FA3 support and as a result FP8 kv-cache in V0 may have issues

Model Support

New model families: GPT-OSS with comprehensive tool calling and streaming support (#22327, #22330, #22332, #22335, #22339, #22340, #22342), Command-A-Vision (#22660), mBART (#22883), and SmolLM3 using Transformers backend (#22665).
Vision-language models: Official Eagle multimodal support with Llama4 backend (#20788), Step3 vision-language models (#21998), Gemma3n multimodal (#20495), MiniCPM-V 4.0 (#22166), HyperCLOVAX-SEED-Vision-Instruct-3B (#20931), Emu3 with Transformers backend (#21319), Intern-S1 (#21628), and Prithvi in online serving mode (#21518).
Enhanced existing models: NemotronH support (#22349), Ernie 4.5 Base 0.3B model name change (#21735), GLM-4.5 series improvements (#22215), Granite models with fused MoE configurations (#21332) and quantized checkpoint loading (#22925), Ultravox support for Llama 4 and Gemma 3 backends (#17818), Mamba1 and Jamba model support in V1 (without CUDA graphs) (#21249)
Advanced model capabilities: Qwen3 EPLB (#20815) and dual-chunk attention support (#21924), Qwen native Eagle3 target support (#22333).
Architecture expansions: Encoder-only models without KV-cache enabling BERT-style architectures (#21270), expanded tensor parallelism support in Transformers backend (#22651), tensor parallelism for Deepseek_vl2 vision transformer (#21494), and tensor/pipeline parallelism with Mamba2 kernel for PLaMo2 (#19674).
V1 engine compatibility: Extended support for additional pooling models (#21747) and Step3VisionEncoder distributed processing option (#22697).

Engine Core

CUDA graph performance: Full CUDA graph support with separate attention routines, adding FA2 and FlashInfer compatibility (#20059), plus 6% end-to-end throughput improvement from Cutlass MLA (#22763).
Attention system advances: Multiple attention metadata builders per KV cache specification (#21588), tree attention backend for v1 engine (experimental) (#20401), FlexAttention encoder-only support (#22273), upgraded FlashAttention 3 with attention sink support (#22313), and multiple attention groups for KV sharing patterns (#22672).
Speculative decoding optimizations: N-gram speculative decoding with single KMP token proposal algorithm (#22437), explicit EAGLE3 interface for enhanced compatibility (#22642).
Default behavior improvements: Pooling models now default to chunked prefill and prefix caching (#20930), disabled chunked local attention by default for Llama4 for better performance (#21761).
Extensibility and configuration: Model loader plugin system (#21067), custom operations support for FusedMoe (#22509), rate limiting with bucket algorithm for proxy server (#22643), torch.compile support for bailing MoE (#21664).
Performance optimizations: Improved startup time by disabling C++ compilation of symbolic shapes (#20836), enhanced headless models for pooling in Transformers backend (#21767).

Hardware & Performance

NVIDIA Blackwell (SM100) optimizations: CutlassMLA as default backend (#21626), FlashInfer MoE per-tensor scale FP8 backend (#21458), SM90 CUTLASS FP8 GEMM with kernel tuning and swap AB support (#20396).
NVIDIA RTX 5090/RTX PRO 6000 (SM120) support: Block FP8 quantization (#22131) and CUTLASS NVFP4 4-bit weights/activations support (#21309).
AMD ROCm platform enhancements: Flash Attention backend for Qwen-VL models (#22069), AITER HIP block quantization kernels (#21242), reduced device-to-host transfers (#22683), and optimized kernel performance for small batch sizes 1-4 (#21350).
Attention and compute optimizations: FlashAttention 3 attention sinks performance boost (#22478), Triton-based multi-dimensional RoPE replacing PyTorch implementation (#22375), async tensor parallelism for scaled matrix multiplication (#20155), optimized FlashInfer metadata building (#21137).
Memory and throughput improvements: Mamba2 reduced device-to-device copy overhead (#21075), fused Triton kernels for RMSNorm (#20839, #22184), improved multimodal hasher performance for repeated image prompts (#22825), multithreaded async multimodal loading (#22710).
Parallelization and MoE optimizations: Guided decoding throughput improvements (#21862), balanced expert sharding for MoE models (#21497), expanded fused kernel support for topk softmax (#22211), fused MoE for nomic-embed-text-v2-moe (#18321).
Hardware compatibility and kernels: ARM CPU build fixes for systems without BF16 support (#21848), Machete memory-bound performance improvements (#21556), FlashInfer TRT-LLM prefill attention kernel support (#22095), optimized reshape_and_cache_flash CUDA kernel (#22036), CPU transfer support in NixlConnector (#18293).
Specialized CUDA kernels: GPT-OSS activation functions (#22538), RLHF weight loading acceleration (#21164).

Quantization

Advanced quantization techniques: MXFP4 and bias support for Marlin kernel (#22428), NVFP4 GEMM FlashInfer backends (#22346), compressed-tensors mixed-precision model loading (#22468), FlashInfer MoE support for NVFP4 (#21639).
Hardware-optimized quantization: Dynamic 4-bit quantization with Kleidiai kernels for CPU inference (#17112), TensorRT-LLM FP4 quantization optimized for MoE low-latency inference (#21331).
Expanded model quantization support: BitsAndBytes quantization for InternS1 (#21953) and additional MoE models (#21370, #21548), Gemma3n quantization compatibility (#21974), calibration-free RTN quantization for MoE models (#20766), ModelOpt Qwen3 NVFP4 support (#20101).
Performance and compatibility improvements: CUDA kernel optimization for Int8 per-token group quantization (#21476), non-contiguous tensor support in FP8 quantization (#21961), automatic detection of ModelOpt quantization formats (#22073).
Breaking change: Removed AQLM quantization support (#22943) - users should migrate to alternative quantization methods.

API & Frontend

OpenAI API compatibility: Unix domain socket support for local communication (#18097), improved error response format matching upstream specification (#22099), aligned tool_choice="required" behavior with OpenAI when tools list is empty (#21052).
New API capabilities: Dedicated LLM.reward interface for reward models (#21720), chunked processing for long inputs in embedding models (#22280), AsyncLLM proper response handling for aborted requests (#22283).
Configuration and environment: Multiple API keys support for enhanced authentication (#18548), custom vLLM tuned configuration paths (#22791), environment variable control for logging statistics (#22905), multimodal cache size (#22441), and DeepGEMM E8M0 scaling behavior (#21968).
CLI and tooling improvements: V1 API support for run-batch command (#21541), custom process naming for better monitoring (#21445), improved help display showing available choices (#21760), optional memory profiling skip for multimodal models (#22950), enhanced logging of non-default arguments (#21680).
Tool and parser support: HermesToolParser for models without special tokens (#16890), multi-turn conversation benchmarking tool (#20267).
Distributed serving enhancements: Enhanced hybrid distributed serving with multiple API servers in load balancing mode (#21510), request_id support for external load balancers (#21009).
User experience enhancements: Improved error messaging for multimodal items (#22114), per-request pooling control via PoolingParams (#20538).

Dependencies

FlashInfer updates: Updated to v0.2.8 for improved performance (#21385), moved to optional dependency install with pip install vllm[flashinfer] for flexible installation (#21959).
Mamba SSM restructuring: Updated to version 2.2.5 (#21421), removed from core requirements to reduce installation complexity (#22541).
Docker and deployment: Docker-aware precompiled wheel support for easier containerized deployment (#21127, #22106).
Python package updates: OpenAI Python dependency updated to latest version for API compatibility (#22316).
Dependency optimizations: Removed xformers requirement for Mistral-format Pixtral and Mistral3 models (#21154), deprecation warnings added for old DeepGEMM version (#22194).

V0 Deprecation

Important: As part of the ongoing V0 engine cleanup, several breaking changes have been introduced:

CLI flag updates: Replaced --task with --runner and --convert options (#21470), deprecated --disable-log-requests in favor of --enable-log-requests for clearer semantics (#21739), renamed --expand-tools-even-if-tool-choice-none to --exclude-tools-when-tool-choice-none for consistency (#20544).
API cleanup: Removed previously deprecated arguments and methods as part of ongoing V0 engine codebase cleanup (#21907).

What's Changed

Deduplicate Transformers backend code using inheritance by @hmellor in https://github.com/vllm-project/vllm/pull/21461
[Bugfix][ROCm] Fix for warp_size uses on host by @gshtras in https://github.com/vllm-project/vllm/pull/21205
[TPU][Bugfix] fix moe layer by @yaochengji in https://github.com/vllm-project/vllm/pull/21340
[v1][Core] Clean up usages of SpecializedManager by @zhouwfang in https://github.com/vllm-project/vllm/pull/21407
[Misc] Fix duplicate FusedMoEConfig debug messages by @njhill in https://github.com/vllm-project/vllm/pull/21455
[Core] Support model loader plugins by @22quinn in https://github.com/vllm-project/vllm/pull/21067
remove GLM-4 quantization wrong Code by @zRzRzRzRzRzRzR in https://github.com/vllm-project/vllm/pull/21435
Replace --expand-tools-even-if-tool-choice-none with --exclude-tools-when-tool-choice-none for v0.10.0 by @okdshin in https://github.com/vllm-project/vllm/pull/20544
[Misc] Improve comment for DPEngineCoreActor._set_cuda_visible_devices() by @ruisearch42 in https://github.com/vllm-project/vllm/pull/21501
[Feat] Allow custom naming of vLLM processes by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/21445
bump flashinfer to v0.2.8 by @cjackal in https://github.com/vllm-project/vllm/pull/21385
[Attention] Optimize FlashInfer MetadataBuilder Build call by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/21137
[Model] Officially support Emu3 with Transformers backend by @hmellor in https://github.com/vllm-project/vllm/pull/21319
[Bugfix] Fix CUDA arch flags for MoE permute by @minosfuture in https://github.com/vllm-project/vllm/pull/21426
[Fix] Update mamba_ssm to 2.2.5 by @elvischenv in https://github.com/vllm-project/vllm/pull/21421
[Docs] Update Tensorizer usage documentation by @sangstar in https://github.com/vllm-project/vllm/pull/21190
[Docs] Rewrite Distributed Inference and Serving guide by @crypdick in https://github.com/vllm-project/vllm/pull/20593
[Bug] Fix Compressed Tensor NVFP4 cutlass_fp4_group_mm illegal memory access by @yewentao256 in https://github.com/vllm-project/vllm/pull/21465
Update flashinfer CUTLASS MoE Kernel by @wenscarl in https://github.com/vllm-project/vllm/pull/21408
[XPU] Conditionally import CUDA-specific passes to avoid import errors on xpu platform by @chaojun-zhang in https://github.com/vllm-project/vllm/pull/21036
[P/D] Move FakeNixlWrapper to test dir by @ruisearch42 in https://github.com/vllm-project/vllm/pull/21328
[P/D] Support CPU Transfer in NixlConnector by @juncgu in https://github.com/vllm-project/vllm/pull/18293
[Docs][minor] Fix broken gh-file link in distributed serving docs by @crypdick in https://github.com/vllm-project/vllm/pull/21543
[Docs] Add Expert Parallelism Initial Documentation by @simon-mo in https://github.com/vllm-project/vllm/pull/21373
update flashinfer to v0.2.9rc1 by @weireweire in https://github.com/vllm-project/vllm/pull/21485
[TPU][TEST] HF_HUB_DISABLE_XET=1 the test 3. by @QiliangCui in https://github.com/vllm-project/vllm/pull/21539
[MoE] More balanced expert sharding by @WoosukKwon in https://github.com/vllm-project/vllm/pull/21497
[Frontend] run-batch supports V1 by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21541
[Docs] Fix site_url for RunLLM by @hmellor in https://github.com/vllm-project/vllm/pull/21564
[Bug] Fix DeepGemm Init Error by @yewentao256 in https://github.com/vllm-project/vllm/pull/21554
Fix GLM-4 PP Missing Layer When using with PP. by @zRzRzRzRzRzRzR in https://github.com/vllm-project/vllm/pull/21531
[Kernel] adding fused_moe configs for upcoming granite4 by @bringlein in https://github.com/vllm-project/vllm/pull/21332
[Bugfix] DeepGemm utils : Fix hardcoded type-cast by @varun-sundar-rabindranath in https://github.com/vllm-project/vllm/pull/21517
[DP] Support api-server-count > 0 in hybrid DP LB mode by @njhill in https://github.com/vllm-project/vllm/pull/21510
[TPU][Test] Temporarily suspend this MoE model in test_basic.py. by @QiliangCui in https://github.com/vllm-project/vllm/pull/21560
[Docs] Add requirements/common.txt to run unit tests by @zhouwfang in https://github.com/vllm-project/vllm/pull/21572
Integrate TensorSchema with shape validation for Phi3VImagePixelInputs by @bbeckca in https://github.com/vllm-project/vllm/pull/21232
[CI] Update CODEOWNERS for CPU and Intel GPU by @bigPYJ1151 in https://github.com/vllm-project/vllm/pull/21582
[Bugfix] fix modelscope snapshot_download serialization by @andyxning in https://github.com/vllm-project/vllm/pull/21536
[Model] Support tensor parallel for timm ViT in Deepseek_vl2 by @wzqd in https://github.com/vllm-project/vllm/pull/21494
[Model] Fix a check for None but the return value was empty list in Gemma3 MM vision_embeddings by @hfan in https://github.com/vllm-project/vllm/pull/21479
[Misc][Tools] make max-model-len a parameter in auto_tune script by @yaochengji in https://github.com/vllm-project/vllm/pull/21321
[CI/Build] fix cpu_extension for apple silicon by @ignaciosica in https://github.com/vllm-project/vllm/pull/21195
[Misc] Removed undefined cmake variables MOE_PERMUTE_ARCHS by @chenyang78 in https://github.com/vllm-project/vllm/pull/21262
[TPU][Bugfix] fix OOM issue in CI test by @yaochengji in https://github.com/vllm-project/vllm/pull/21550
[Tests] Harden DP tests by @njhill in https://github.com/vllm-project/vllm/pull/21508
Add H20-3e fused MoE kernel tuning configs for Qwen3-Coder-480B-A35B-Instruct by @Xu-Wenqing in https://github.com/vllm-project/vllm/pull/21598
[Bugfix] GGUF: fix AttributeError: 'PosixPath' object has no attribute 'startswith' by @kebe7jun in https://github.com/vllm-project/vllm/pull/21579
[Quantization] Enable BNB support for more MoE models by @jeejeelee in https://github.com/vllm-project/vllm/pull/21370
[V1] Get supported tasks from model runner instead of model config by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21585
[Bugfix][Logprobs] Fix logprobs op to support more backend by @MengqingCao in https://github.com/vllm-project/vllm/pull/21591
[Model] Fix Ernie4.5MoE e_score_correction_bias parameter by @xyxinyang in https://github.com/vllm-project/vllm/pull/21586
[MODEL] New model support for naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B by @bigshanedogg in https://github.com/vllm-project/vllm/pull/20931
[Frontend] Add request_id to the Request object so they can be controlled better via external load balancers by @kouroshHakha in https://github.com/vllm-project/vllm/pull/21009
[Model] Replace Mamba2 RMSNorm Gated with Fused Triton Kernel by @cyang49 in https://github.com/vllm-project/vllm/pull/20839
[ROCm][AITER] Enable fp8 kv cache on rocm aiter backend. by @fsx950223 in https://github.com/vllm-project/vllm/pull/20295
[Kernel] Improve machete memory bound perf by @czhu-cohere in https://github.com/vllm-project/vllm/pull/21556
Add support for Prithvi in Online serving mode by @mgazz in https://github.com/vllm-project/vllm/pull/21518
[CI] Unifying Dockerfiles for ARM and X86 Builds by @kebe7jun in https://github.com/vllm-project/vllm/pull/21343
[Docs] add auto-round quantization readme by @wenhuach21 in https://github.com/vllm-project/vllm/pull/21600
[TPU][Test] Rollback PR-21550. by @QiliangCui in https://github.com/vllm-project/vllm/pull/21619
Add Unsloth to RLHF.md by @danielhanchen in https://github.com/vllm-project/vllm/pull/21636
[Perf] Cuda Kernel for Int8 Per Token Group Quant by @yewentao256 in https://github.com/vllm-project/vllm/pull/21476
Add interleaved RoPE test for Llama4 (Maverick) by @sarckk in https://github.com/vllm-project/vllm/pull/21478
[Bugfix] Fix sync_and_slice_intermediate_tensors by @ruisearch42 in https://github.com/vllm-project/vllm/pull/21537
[Bugfix] Always set RAY_ADDRESS for Ray actor before spawn by @ruisearch42 in https://github.com/vllm-project/vllm/pull/21540
[TPU] Update ptxla nightly version to 20250724 by @yaochengji in https://github.com/vllm-project/vllm/pull/21555
[Feature] Add support for MoE models in the calibration-free RTN-based quantization by @sakogan in https://github.com/vllm-project/vllm/pull/20766
[Model] Ultravox: Support Llama 4 and Gemma 3 backends by @farzadab in https://github.com/vllm-project/vllm/pull/17818
[Docs] add offline serving multi-modal video input expamle Qwen2.5-VL by @david6666666 in https://github.com/vllm-project/vllm/pull/21530
Correctly kill vLLM processes after finishing serving benchmarks by @huydhn in https://github.com/vllm-project/vllm/pull/21641
[Bugfix] Fix isinstance check for tensor types in _load_prompt_embeds to use dtype comparison by @Mitix-EPI in https://github.com/vllm-project/vllm/pull/21612
[TPU][Test] Divide TPU v1 Test into 2 parts. by @QiliangCui in https://github.com/vllm-project/vllm/pull/21431
Support Intern-S1 by @lvhan028 in https://github.com/vllm-project/vllm/pull/21628
[Misc] remove unused try-except in pooling config check by @reidliu41 in https://github.com/vllm-project/vllm/pull/21618
[Take 2] Correctly kill vLLM processes after benchmarks by @huydhn in https://github.com/vllm-project/vllm/pull/21646
Migrate AriaImagePixelInputs to TensorSchema for shape validation by @bbeckca in https://github.com/vllm-project/vllm/pull/21620
Migrate AyaVisionImagePixelInputs to TensorSchema for shape validation by @bbeckca in https://github.com/vllm-project/vllm/pull/21622
[Bugfix] Investigate Qwen2-VL failing test by @Isotr0py in https://github.com/vllm-project/vllm/pull/21527
Support encoder-only models without KV-Cache by @maxdebayser in https://github.com/vllm-project/vllm/pull/21270
[Bug] Fix has_flashinfer_moe Import Error when it is not installed by @yewentao256 in https://github.com/vllm-project/vllm/pull/21634
[Misc] Improve memory profiling debug message by @yeqcharlotte in https://github.com/vllm-project/vllm/pull/21429
[BugFix] Fix shared storage connector load kv only load attention layer by @david6666666 in https://github.com/vllm-project/vllm/pull/21428
[Refactor] Remove moe_align_block_size_triton by @yewentao256 in https://github.com/vllm-project/vllm/pull/21335
[Bugfix][Apple Silicon] fix missing symbols when build from source on Mac with Apple Silicon by @zhouyeju in https://github.com/vllm-project/vllm/pull/21380
[CI/Build][Doc] Move existing benchmark scripts in CI/document/example to vllm bench CLI by @yeqcharlotte in https://github.com/vllm-project/vllm/pull/21355
[NVIDIA] Explicitly disable shuffled weights for flashinfer blockscale moe fp8 kernels by @kaixih in https://github.com/vllm-project/vllm/pull/21411
Remove xformers requirement for Mistral-format Pixtral and Mistral3 by @wenchen76 in https://github.com/vllm-project/vllm/pull/21154
support torch.compile for bailing moe by @jinzhen-lin in https://github.com/vllm-project/vllm/pull/21664
Migrate Blip2ImagePixelInputs and Blip2ImageEmbeddingInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21656
Migrate DeepseekVL2ImageInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21658
Migrate FuyuImagePatchInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21662
Migrate ChameleonImagePixelInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21657
[VLM] Support HF format Phi-4-MM model by @Isotr0py in https://github.com/vllm-project/vllm/pull/17121
Handle non-serializable objects in vllm bench by @huydhn in https://github.com/vllm-project/vllm/pull/21665
[CI/Build][Doc] Clean up more docs that point to old bench scripts by @yeqcharlotte in https://github.com/vllm-project/vllm/pull/21667
Refactor: Remove numpy dependency from LoggingStatLogger by @skyloevil in https://github.com/vllm-project/vllm/pull/20529
[Misc] add default value for file pattern arg by @andyxning in https://github.com/vllm-project/vllm/pull/21659
Migrate Florence2ImagePixelInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21663
[VLM] Add video support for Intern-S1 by @Isotr0py in https://github.com/vllm-project/vllm/pull/21671
[Refactor] Refactor MOE NVFP4 Code Base: ModelOpt + Compressed Tensor by @yewentao256 in https://github.com/vllm-project/vllm/pull/21631
Fix CUDA permute/unpermute for use with DeepGemm Moe by @CalebDu in https://github.com/vllm-project/vllm/pull/17934
[Misc] Refactor vllm config str by @andyxning in https://github.com/vllm-project/vllm/pull/21666
[Attention] Make CutlassMLA the default backend for SM100 (blackwell) by @alexm-redhat in https://github.com/vllm-project/vllm/pull/21626
[Deprecation][2/N] Replace --task with --runner and --convert by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21470
Fix typo for limit-mm-per-prompt in docs by @joa-stdn in https://github.com/vllm-project/vllm/pull/21697
Fix GLM tool parser by @zRzRzRzRzRzRzR in https://github.com/vllm-project/vllm/pull/21668
[Misc] Add fused_moe configs for Qwen3-Coder-480B-A35B-Instruct-FP8 by @jeejeelee in https://github.com/vllm-project/vllm/pull/21700
[V1] Exception Handling when Loading KV Cache from Remote Store by @liuyumoye in https://github.com/vllm-project/vllm/pull/21534
[Model] Support TP/PP/mamba2 kernel for PLaMo2 by @Alnusjaponica in https://github.com/vllm-project/vllm/pull/19674
[FEAT] [ROCm] [AITER]: Add AITER HIP block quant kernel by @tjtanaa in https://github.com/vllm-project/vllm/pull/21242
Migrate Gemma3ImagePixelInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21676
Migrate Glm4vImageInputs, Glm4vVideoInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21678
Migrate GLMVImagePixelInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21679
Migrate GraniteSpeechAudioInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21682
Migrate Idefics3ImagePixelInputs and Idefics3ImageEmbeddingInputs to … by @bbeckca in https://github.com/vllm-project/vllm/pull/21683
[Bugfix] [issue-21565] Fix the incompatibility issue with stream and named function calling when Thinking is disabled by @hsliuustc0106 in https://github.com/vllm-project/vllm/pull/21573
[bugfix] fix profile impact benchmark results by @lengrongfu in https://github.com/vllm-project/vllm/pull/21507
[Bugfix] Fix shape checking for Fuyu by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21709
[Bugfix] fix max-file-size type from str to int by @andyxning in https://github.com/vllm-project/vllm/pull/21675
[BugFix] Fix ChunkedLocalAttention when the hybrid kv-cache is disabled by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/21707
[v1][mamba] Added mamba_type into MambaSpec by @Josephasafg in https://github.com/vllm-project/vllm/pull/21715
Migrate KeyeImageInputs and KeyeVideoInputs to TensorSchema by @bbeckca in https://github.com/vllm-project/vllm/pull/21686
[Model] Prioritize Transformers fallback over suffix matching by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21719
[feature] add log non default args in LLM by @lengrongfu in https://github.com/vllm-project/vllm/pull/21680
[Bugfix] Fix Ernie4_5_MoeForCausalLM shared experts by @jeejeelee in https://github.com/vllm-project/vllm/pull/21717
[Bugfix] Fix environment variable setting in CPU Dockerfile by @bigPYJ1151 in https://github.com/vllm-project/vllm/pull/21730
[Bugfix] Fix glm4.1v video_grid_thw tensor shape scheme by @Isotr0py in https://github.com/vllm-project/vllm/pull/21744
[PD] let p2p nccl toy proxy handle /chat/completions by @chaunceyjiang in https://github.com/vllm-project/vllm/pull/21734
[Ernie 4.5] Name Change for Base 0.3B Model by @vasqu in https://github.com/vllm-project/vllm/pull/21735
[Bugfix] Improve JSON extraction in LlamaToolParser by @key4ng in https://github.com/vllm-project/vllm/pull/19024
[Docs] Add revision date to rendered docs by @hmellor in https://github.com/vllm-project/vllm/pull/21752
[Bugfix]check health for engine core process exiting unexpectedly by @wuhang2014 in https://github.com/vllm-project/vllm/pull/21728
[Bugfix][CI/Build] Update peft version in test requirement by @Isotr0py in https://github.com/vllm-project/vllm/pull/21729
[Logs] Change flashinfer sampler logs to once by @mgoin in https://github.com/vllm-project/vllm/pull/21759
[Misc] Reduce logs for model resolution by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21765
[Bugfix] Mistral crashes on tool with no description by @HugoMichard in https://github.com/vllm-project/vllm/pull/21167
[CI/Build] Fix plugin tests by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21758
[XPU] IPEX-optimized Punica Wrapper on XPU by @chaojun-zhang in https://github.com/vllm-project/vllm/pull/21703
[Bugfix] Fix granite speech shape validation by @DarkLight1337 in https://github.com/vllm-project/vllm/pull/21762
[P/D] Log warnings related to prefill KV expiry by @njhill in https://github.com/vllm-project/vllm/pull/21753
Use metavar to list the choices for a CLI arg when custom values are also accepted by @hmellor in https://github.com/vllm-project/vllm/pull/21760
update flashinfer to v0.2.9rc2 by @weireweire in https://github.com/vllm-project/vllm/pull/21701
[AMD][BugFix] Fix omission of wvSplitK kernel for small batch sizes (1-4) due to torch.compile by @rasmith in https://github.com/vllm-project/vllm/pull/21350
[Bug] Enforce contiguous input for dynamic_scaled_fp8_quant and static_scaled_fp8_quant by @yewentao256 in https://github.com/vllm-project/vllm/pull/21773
[AMD][CI/Build] Fix the AMD issue caused by inappropriate of symbol exposure by @houseroad in https://github.com/vllm-project/vllm/pull/21647
Revert "[V1] Exception Handling when Loading KV Cache from Remote Store" by @KuntaiDu in https://github.com/vllm-project/vllm/pull/21778
[Bugfix] DeepGEMM is not enabled on B200 due to _lazy_init() by @smarterclayton in https://github.com/vllm-project/vllm/pull/21472
[Feat]: Add support for Dynamic Quant 4 bit CPU kleidiai kernels by @nikhil-arm in https://github.com/vllm-project/vllm/pull/17112
[Perf] Disable chunked local attention by default with llama4 by @LucasWilkinson in https://github.com/vllm-project/vllm/pull/21761
[Kernel] SM90 CUTLASS FP8 GEMM: add support for swap AB + kernel tuning by @LyrisZhong in https://github.com/vllm-project/vllm/pull/20396
[Docs] Minimize spacing for supported_hardware.md table by @mgoin in https://github.com/vllm-project/vllm/pull/21779
[Refactor] Merge Compressed Tensor FP8 CompressedTensorsW8A8Fp8MoEMethod and CompressedTensorsW8A8Fp8MoECutlassMethod by @yewentao256 in https://github.com/vllm-project/vllm/pull/21775
[CI] Parallelize Kernels MoE Test by @mgoin in https://github.com/vllm-project/vllm/pull/21764
skip fusedmoe layer for

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