vllm

v0.21.0 Breaking

This release includes 2 breaking changes for platform teams planning a safe upgrade.

Published 19d Model Serving & MLOps

View tool

✓ No known CVEs patched

Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

amd blackwell cuda deepseek deepseek-v3 llm

+13 more

gpt-oss inference kimi llama llm-serving model-serving moe openai pytorch qwen qwen3 tpu transformer

Affected surfaces

breaking_upgrade

ReleasePort's take

Moderate signal

editorial:auto 9d

Transformers v4 support is deprecated; vLLM v0.21.0 now requires a C++20 compiler for PyTorch compatibility.

Why it matters: Deprecation of Transformers v4 triggers migration before future removal; the new C++20 compiler requirement mandates updating build environments to maintain PyTorch compatibility in vLLM.

Summary

AI summary

vLLM now requires a C++20 compiler and deprecates Transformers v4, requiring migration.

Changes in this release

Type	Severity	Summary	CVE
Breaking	Medium	C++20 compiler requirement for PyTorch compatibility C++20 compiler requirement for PyTorch compatibility Source: llm_adapter@2026-05-21 Confidence: low	—
Feature
Feature	Medium	KV offload integrates with Hybrid Memory Allocator (HMA) KV offload integrates with Hybrid Memory Allocator (HMA) Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Speculative decoding respects reasoning/thinking budgets Speculative decoding respects reasoning/thinking budgets Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	TOKENSPEED_MLA backend for Blackwell GPUs available TOKENSPEED_MLA backend for Blackwell GPUs available Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	New MiMo-V2.5 architecture support New MiMo-V2.5 architecture support Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Laguna XS.2 architecture support added Laguna XS.2 architecture support added Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Moondream3 architecture support added Moondream3 architecture support added Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Qianfan-OCR model support introduced Qianfan-OCR model support introduced Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Cohere MoE model support enabled Cohere MoE model support enabled Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Cohere Eagle model support enabled Cohere Eagle model support enabled Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	DeepSeek V4 AMD/ROCm support added DeepSeek V4 AMD/ROCm support added Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Speculative decoding for EAGLE Mistral and Gemma4 MTP Speculative decoding for EAGLE Mistral and Gemma4 MTP Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Pipeline parallelism for DeepSeek V4 Pipeline parallelism for DeepSeek V4 Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	Disaggregated serving with bidirectional KV cache transfers Disaggregated serving with bidirectional KV cache transfers Source: llm_adapter@2026-05-21 Confidence: high	—
Feature	Medium	FlashInfer top-k/top-p sampler enabled by default FlashInfer top-k/top-p sampler enabled by default Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	FP8 FlashInfer attention for ViT supported FP8 FlashInfer attention for ViT supported Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	AllPool.forward speed increased 51% AllPool.forward speed increased 51% Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	GPU<->CPU sync eliminated in pooling and attention GPU<->CPU sync eliminated in pooling and attention Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	Multimodal processor skip for text-only inputs Multimodal processor skip for text-only inputs Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	NVFP4 KV cache support added NVFP4 KV cache support added Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	MXFP4 MoE backend introduced MXFP4 MoE backend introduced Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	TurboQuant hybrid model and uniform quantization supported TurboQuant hybrid model and uniform quantization supported Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	Responses API supports streaming tool/function calling with required fields Responses API supports streaming tool/function calling with required fields Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	OpenAI compatibility adds system_fingerprint field OpenAI compatibility adds system_fingerprint field Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	Tool calling includes XGrammar 0.2.0 with structural tags Tool calling includes XGrammar 0.2.0 with structural tags Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	Tokenizer gains Fastokens support Tokenizer gains Fastokens support Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	RLHF APIs expose /start_weight_update and /finish_weight_update endpoints RLHF APIs expose /start_weight_update and /finish_weight_update endpoints Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	ASR engine request abort on cancellation supported ASR engine request abort on cancellation supported Source: llm_adapter@2026-05-21 Confidence: low	—
Feature	Medium	Configuration variable VLLM_SKIP_MODEL_NAME_VALIDATION introduced Configuration variable VLLM_SKIP_MODEL_NAME_VALIDATION introduced Source: llm_adapter@2026-05-21 Confidence: low	—
Deprecation	Medium	Transformers v4 support deprecated Transformers v4 support deprecated Source: llm_adapter@2026-05-21 Confidence: high	—

Full changelog

Highlights

This release features 367 commits from 202 contributors (49 new)!

Transformers v4 deprecated: This release formally deprecates transformers v4 support (#40389). Users should migrate to transformers v5.
C++20 build requirement: vLLM now requires a C++20-compatible compiler for compatibility with PyTorch (#40380). This is a breaking build change.
KV Offload + Hybrid Memory Allocator (HMA): The KV offloading subsystem now integrates with the Hybrid Memory Allocator, including scheduler-side sliding window group support and full HMA enablement (#41228, #41445, #39571).
Speculative decoding with thinking budget: Speculative decoding now respects reasoning/thinking budgets, enabling correct spec decode for reasoning models (#34668).
TOKENSPEED_MLA backend on Blackwell: A new TOKENSPEED_MLA attention backend is available for DeepSeek-R1/Kimi-K25 prefill + decode on Blackwell GPUs (#41778).

Model Support

New architectures: MiMo-V2.5 (#40967), Laguna XS.2 (#41129, #41880), Moondream3 (#32325), Qianfan-OCR (#40136), Cohere MoE (#40817), Cohere Eagle (#42078).
Speculative decoding: EAGLE for Mistral (#41024), Gemma4 MTP (#41745), MTP for MiMo-V2.5 (#41905), Cohere Eagle (#42078).
DeepSeek V4: AMD/ROCm support (#40871), pipeline parallelism (#41694), max reasoning effort (#40982), disaggregated serving fixes (#41957).
Tool calling: Cohere reasoning and tool parsers (#40422), LFM2/2.5 tool parser (#39243).
Gemma3/Gemma4: hidden_act variant support (#40588), pipeline parallelism fix (#40786), MoE fixes (#41206, #41574, #41401), tool parser crash fix (#41991, #42188).
Model Runner V2: Qwen3.5/Mamba hybrid model support (#35520), logprob_token_ids support (#40559).
CUDA graph: ViT CUDA graph support for Qwen2.5-VL (#40830).
Compatibility: Vendor HCXVisionConfig for Transformers v5 (#38447), legacy rope_type checkpoint support (#41734).

Engine Core

KV offloading + HMA: Scheduler-side sliding window groups (#41228), full HMA enablement (#41445), multi-connector HMA (#39571), per-job store completion (#39186), DCP/PCP support in OffloadingConnector (#41549), MooncakeStoreConnector for distributed KV offloading (#40900).
Speculative decoding: Thinking budget support (#34668), independent drafter attention backend selection (#39930), multimodal model support with warning (#41752), per-step allocation elimination (#41043).
Model Runner V2: Rejection sampling acceptance rate fix (#40651), skip metadata rebuild before draft prefill (#40410), rebuild metadata between draft decode steps (#41162), Qwen3.5/Mamba hybrid support (#35520).
Routing: Replace routing replay with device cache and async D2H pipeline (#39917).
Ray: RayExecutorV2 enabled by default (#41421), actor name collision fix for DP > 1 (#40398).
Stability: Two-phase pause to prevent scheduler deadlock (#39366), thread-safe HF tokenizer wrappers (#41181), OOM prevention via max_split_size_mb during model loading (#41268).
IndexCache support for DSA models (#37735).

Hardware & Performance

NVIDIA Blackwell: TOKENSPEED_MLA backend for DSR1/Kimi-K25 (#41778), faster per-token FP8 group quant packed kernel (#41326), FP8 on NVIDIA Thor/SM110 (#39712), CUTLASS scaled mm for non-compatible sizes (#41868).
Performance: FlashInfer top-k/top-p sampler enabled by default (#40376), FP8 FlashInfer attention for ViT (#38065), TurboQuant shared dequant buffers (#40941), AllPool.forward 51% faster (#41163), GPU<->CPU sync elimination in pooling (#41433) and attention (#41434), numpy zero-copy embedding serialization (#41681), multimodal processor skip for text-only (#41246), FlashInfer FP8 async TP fusion (#39505), NVFP4 all-gather GEMM fusion for AsyncTP (#41882), re-enable allreduce+RMS fusion for DP/PP (#41458), DeepSeek bf16→fp32 via torch.mm (#41300), persistent MLA for sparse backend (#41990), configurable safetensors checkpoint prefetch (#41499), fused mhc_post_pre kernel (#41536), 2D-grid W8W8 group quant kernel (#42153), relaxed memory ordering for KV cache swaps (#39306).
AMD ROCm: ROCm 7.2.2 (#41386), DBO (Dynamic Batch Optimization) (#34726), AITER Fused Allreduce+RMSNorm (#37646), Fused Shared Expert (FSE) for Qwen3-Next (#39280), DeepSeek V3.2 TP4 AITER MLA (#41835), GDN linear attention fusion (#40711), eliminate redundant MoE buffer copies in AITER (#41713), CPU offloading support (#40549), DeepEP API update (#39721), cap Triton paged attention block size to fix shared memory OOM (#38502).
CPU: FP8 attention for AMX/AVX-512 (#39445), FP8 W8A16 linear (#41186), FP8 W8A16 MoE (#41314), DNNL AVX2 W8A8 Int8 (#41318), Gated DeltaNet Attention for Qwen 3.5/3.6 (#41025), RISC-V OMP thread auto-binding (#40569).
Intel XPU: Top-k/top-p sample kernel (#39285), out-of-place all-reduce (#41808), LoRA support (#38206).
IBM Power: VSX attention backend (#40451).
FlexAttention: Re-enabled for batch invariant mode (#40842).
MLA: Abstracted MLA prefill backends, eliminated cuDNN dependency (#32623).

Large Scale Serving

Disaggregated serving: Bi-directional KV cache transfers between P and D (#32553), NIXL transfer redesign (#40731), EPLB memory overhead optimization (#40013), NIXL connector bumped to 1.x (#42364), Mooncake KVConnectorStats for transfer observability (#40414), NIXL P-node pre-admission rejection notification (#41269), KV block release for skipped P-ranks (#40449).
DCP: Pack output and LSE in DCP A2A (#41160).
MoE: PluggableLayer interface for out-of-tree MoE runners (#35178).
LoRA: Initial expert parallel (EP) support (#40867), Qwen3.5 LoRA fusion fix (#37912).

Quantization

NVFP4: KV cache support (#40177), Triton dequant/QDQ emulation kernels for Hopper and AMD (#40033), GELU on TRT-LLM NvFP4 fused MoE for Gemma4 (#41050), ModelOpt NVFP4 W4A16 (#41769), NVFP4 all-gather GEMM fusion for AsyncTP (#41882), GLM4-MoE NVFP4 loading fix (#41755).
MXFP4: Humming MXFP4 MoE backend (#41083), FlashInfer CUTLASS MXFP4-MXFP8 MoE fix (#42089).
TurboQuant: Hybrid model and uniform quantization support (#39931).
Compressed tensors: Allow configs with non-explicit ignores (#41965).
FP8: Bias loading fix (#41424), FlashInfer autotune temporarily disabled for correctness (#41524).
DSV4: Improved fused Indexer Q quant kernel (#41428).

API & Frontend

Responses API: Streaming tool/function calling with required (#40700) and named tool/function choice (#41110), resubmitting output items with missing fields (#41355).
OpenAI compatibility: system_fingerprint field in responses (#40537), prompt_embeds content part support (#40720), defer_loading and tool_reference support (#40190), rendered prompt text in chat completion response (#42052), tolerate empty content in forced tool choice (#40148).
Tool calling: XGrammar 0.2.0 with structural tags for strict tool calling + reasoning (#40894), Cohere reasoning/tool parsers (#40422), LFM2/2.5 tool parser (#39243).
Tokenizer: Fastokens support (#41741).
RLHF: Explicit /start_weight_update and /finish_weight_update APIs (#39212).
ASR: Engine request abort on cancellation (#41266).
Configuration: VLLM_SKIP_MODEL_NAME_VALIDATION env var (#34676), configurable model weights loading tracking (#41086), Triton JIT compilation monitor (#40137).

Build & Dependencies

Breaking: C++20 required for PyTorch compatibility (#40380).
Breaking: Transformers v4 deprecated (#40389).
Docker image size reduced by ~2.5 GB via deferred FlashInfer cubin download (#41134).
CUDA 13.0 wheels switched to PyTorch manylinux_2_28 base (#41416).
DeepGEMM bundled wheel built per-Python for CPython compatibility (#41516).
Container image provenance metadata embedded (#40653).
tpu-inference upgraded to v0.19.0 (#41844).
NIXL connector bumped to 1.x (#42364).
ROCm 7.2.2 (#41386).

Contributors

@AndreasKaratzas, @haosdent, @khluu, @yewentao256, @stecasta, @mgoin, @Isotr0py, @hmellor, @chaunceyjiang, @jeejeelee, @noooop, @MatthewBonanni, @njhill, @zyongye, @yzong-rh, @ronensc, @NickLucche, @chaojun-zhang, @dzhengAP, @chfeng-cs, @TheEpicDolphin, @esmeetu, @wzhao18, @ZJY0516, @juliendenize, @kylesayrs, @fadara01, @Etelis, @tianmu-li, @arpera, @ekagra-ranjan, @orozery, @wxsIcey, @jikunshang, @izhuhaoran, @rasmith, @russellb, @Lucaskabela, @Harry-Chen, @alec-flowers, @pmaybank, @Terrencezzj, @hickeyma, @Baekpica, @itej89, @fxmarty-amd, @WoosukKwon, @juhi10071998, @sychen52, @baonudesifeizhai, @vllmellm, @johncalesp, @the-david-oy, @lucianommartins, @bittoby, @Dao007forever, @lyd1992, @yuwenzho, @lesj0610, @sfeng33, @micah-wil, @akii96, @yma11, @SoluMilken, @mmangkad, @SiluPanda, @ojhaanshika, @zhandaz, @bhoomit, @simon-mo, @msanft, @angelayi, @anthonsu, @artem-spector, @zhangxin81, @benoittgt, @joerowell, @yangrz7, @chelnnexy, @liangel-02, @walterbm, @rishitdholakia13, @SKRohit, @BugenZhao, @JaredforReal, @amd-lalithnc, @frgossen, @h-avsha, @DarkLight1337, @danisereb, @laithsakka, @Bortlesboat, @wangluochao902, @Rohan138, @hao-aaron, @puririshi98, @roikoren755, @heachary, @UranusSeven, @dsingal0, @ChenxiQ, @snadampal, @ilmarkov, @wendyliu235, @lequytra, @JisoLya, @LuisRobaina, @sniper35, @eicherseiji, @Yuyi-Ao, @raviguptaamd, @sungsooha, @ganyi1996ppo, @andylolu2, @FredericOdermatt, @ProExpertProg, @rbrugaro-amd, @mcsantiago, @hnt2601, @jinzhen-lin, @taneem-ibrahim, @tomeras91, @alex-jw-brooks, @Aktsvigun, @HanFa, @netanel-haber, @JasonKeyiL, @gshtras, @joa-stdn, @Seven-Streams, @JartX, @xuechendi, @BowenBao, @Akashcodes732, @jeffreywang-anyscale, @czhu-cohere, @zhewenl, @marvinzh, @Lidang-Jiang, @gcanlin, @whx-sjtu, @S1ro1, @liulanze, @Dhruvilbhatt, @laviier, @wi-adam, @aaab8b, @yuankaichen-amd, @ZhanqiuHu, @QwertyJack, @viktorpusTT, @divakar-amd, @starkwj, @benchislett, @jcyang43, @JLiu4Coding, @xy3xy3, @hongxiayang, @amd-mghanimi, @wenyili, @bigPYJ1151, @s-yanev, @AlonKejzman, @noobHappylife, @TomerBN-Nvidia, @MeganEFlynn, @liuzijing2014, @jbuchananr, @lokashrinav, @ssam18, @dllehr-amd, @gmagogsfm, @tpopp, @tjtanaa, @simondanielsson, @zhenwei-intel, @HiroakiMikami, @nholmber, @SumanthRH, @LucasWilkinson, @maeehart, @rishaps, @r-barnes, @gau-nernst, @Kermit-C, @tdoublep, @aoshen02, @Naveassaf, @wangxingran222, @cvan20191, @AbhiOnGithub, @abdulrahman-cohere, @jmamou, @Flink-ddd, @bnellnm, @hqhq1025, @gnovack, @wangxiyuan, @princepride, @jiahanc, @LCAIZJ, @ovidiusm

New Contributors

@abdulrahman-cohere made their first contribution in https://github.com/vllm-project/vllm/pull/41266
@AbhiOnGithub made their first contribution in https://github.com/vllm-project/vllm/pull/42180
@Aktsvigun made their first contribution in https://github.com/vllm-project/vllm/pull/40788
@amd-mghanimi made their first contribution in https://github.com/vllm-project/vllm/pull/41713
@Baekpica made their first contribution in https://github.com/vllm-project/vllm/pull/41206
@benoittgt made their first contribution in https://github.com/vllm-project/vllm/pull/41134
@bittoby made their first contribution in https://github.com/vllm-project/vllm/pull/41690
@chelnnexy made their first contribution in https://github.com/vllm-project/vllm/pull/40754
@ChenxiQ made their first contribution in https://github.com/vllm-project/vllm/pull/40956
@chfeng-cs made their first contribution in https://github.com/vllm-project/vllm/pull/42066
@cvan20191 made their first contribution in https://github.com/vllm-project/vllm/pull/40951
@dzhengAP made their first contribution in https://github.com/vllm-project/vllm/pull/41423
@ghphotoframe made their first contribution in https://github.com/vllm-project/vllm/pull/40859
@HiroakiMikami made their first contribution in https://github.com/vllm-project/vllm/pull/40588
@itej89 made their first contribution in https://github.com/vllm-project/vllm/pull/39721
@JasonKeyiL made their first contribution in https://github.com/vllm-project/vllm/pull/41068
@jbuchananr made their first contribution in https://github.com/vllm-project/vllm/pull/39243
@JisoLya made their first contribution in https://github.com/vllm-project/vllm/pull/41363
@JLiu4Coding made their first contribution in https://github.com/vllm-project/vllm/pull/41832
@juhi10071998 made their first contribution in https://github.com/vllm-project/vllm/pull/41050
@Kermit-C made their first contribution in https://github.com/vllm-project/vllm/pull/42076
@lequytra made their first contribution in https://github.com/vllm-project/vllm/pull/41401
@Lidang-Jiang made their first contribution in https://github.com/vllm-project/vllm/pull/38099
@liulanze made their first contribution in https://github.com/vllm-project/vllm/pull/41571
@lokashrinav made their first contribution in https://github.com/vllm-project/vllm/pull/41681
@LuisRobaina made their first contribution in https://github.com/vllm-project/vllm/pull/40720
@maeehart made their first contribution in https://github.com/vllm-project/vllm/pull/42061
@marvinzh made their first contribution in https://github.com/vllm-project/vllm/pull/40136
@mcsantiago made their first contribution in https://github.com/vllm-project/vllm/pull/41492
@MeganEFlynn made their first contribution in https://github.com/vllm-project/vllm/pull/41880
@nholmber made their first contribution in https://github.com/vllm-project/vllm/pull/39280
@pmaybank made their first contribution in https://github.com/vllm-project/vllm/pull/41012
@raviguptaamd made their first contribution in https://github.com/vllm-project/vllm/pull/34726
@s-yanev made their first contribution in https://github.com/vllm-project/vllm/pull/41755
@S1ro1 made their first contribution in https://github.com/vllm-project/vllm/pull/39213
@Seven-Streams made their first contribution in https://github.com/vllm-project/vllm/pull/40894
@SiluPanda made their first contribution in https://github.com/vllm-project/vllm/pull/40907
@SKRohit made their first contribution in https://github.com/vllm-project/vllm/pull/40786
@snadampal made their first contribution in https://github.com/vllm-project/vllm/pull/32553
@sniper35 made their first contribution in https://github.com/vllm-project/vllm/pull/32325
@ssam18 made their first contribution in https://github.com/vllm-project/vllm/pull/41486
@the-david-oy made their first contribution in https://github.com/vllm-project/vllm/pull/40737
@wangluochao902 made their first contribution in https://github.com/vllm-project/vllm/pull/41043
@wenyili made their first contribution in https://github.com/vllm-project/vllm/pull/41901
@wi-adam made their first contribution in https://github.com/vllm-project/vllm/pull/40749
@xy3xy3 made their first contribution in https://github.com/vllm-project/vllm/pull/40820
@yangrz7 made their first contribution in https://github.com/vllm-project/vllm/pull/40449
@yuankaichen-amd made their first contribution in https://github.com/vllm-project/vllm/pull/40390
@zhangxin81 made their first contribution in https://github.com/vllm-project/vllm/pull/39904