vllm-project/vllm-omni v0.30.0rc1 on GitHub

Highlights

This release features 262 merged changes from 126 contributors, including 45 new contributors.

vLLM-Omni v0.30.0rc1 focuses on three major advances: 1) a unified full-duplex framework and scalable serving, 2) interactive video generation and improved cross-stage data transfer, and 3) expanded speech, image, and vision-language-action model support. The release rebases onto vLLM 0.30.0, advances streaming execution, and improves quantization and memory efficiency across multimodal pipelines.

Key Improvements

Unified full-duplex execution around engine-owned sessions, introducing dedicated duplex orchestration, an in-process client, and AURA integration through the shared Realtime API. (#7413, #7633)
Advanced interactive and streaming video generation, adding live camera control for LingBot World, worker-side MP4 encoding, asynchronous chunk transfers, and MiniMax-H3 latent editing and long-video continuation. (#7198, #7048, #7406, #7465, #7838)
Broadened model coverage with Tencent AuK/AuK-Flash, Breeze-TTS-2, Anima, and π0.5, alongside Gepard-1.0 online speech serving and direct FastH3 8-Step V2 checkpoint loading. (#7385, #7084, #4083, #6950, #7499, #7610)

Core Architecture & Runtime

Rebased the project onto vLLM 0.30.0. (#7820)
Introduced DuplexOmni, DuplexOmniEngine, and DuplexOrchestrator, moving session ownership and execution into the engine layer. Added InlineDuplexClient for in-process use and a shared model plugin interface. (#7413)
Added an optional NIXL connector for cross-stage tensor and structured-payload transport, plus native Mooncake AR-to-DiT KV transfer with scheduler-owned GPU pages and opt-in HunyuanImage3 prefetch. (#6093, #7166, #7637)
Preserved structured configurations through runtime startup and refactored Omni prefix caching into separate manager and controller components. (#6849, #6654)
Added the initial world-model rollout serving API, with session creation, stepping, reset, closure, and committed-step tracking. This establishes a serving interface for external RL integrations. (#4770)

Model Support

Added Tencent AuK and AuK-Flash through a two-stage encoder–diffusion pipeline for instruction-driven speech generation and audio editing, including the distilled four-step Flash checkpoint. (#7385)
Added Breeze-TTS-2, with a two-stage autoregressive talker and codec pipeline, incremental 24 kHz audio, voice design, reference-based cloning, and voice direction. (#7084)
Added Anima native text-to-image generation, including local single-file safetensors checkpoint loading. (#4083)
Added π0.5 (Pi0.5) vision-language-action inference through the OpenPI realtime robot API, generating action chunks from camera images, language instructions, and robot state. (#6950)
Extended Gepard-1.0 to /v1/audio/speech with streaming and concurrent-request isolation, and added direct FastH3 8-Step V2 checkpoint loading. (#7499, #7610)

Audio, Speech & Realtime Serving

Integrated AURA’s ASR → Thinker → Talker → Code2Wav pipeline into the unified full-duplex runtime, supporting video-only inputs and shared session, interruption, and resume controls. (#7633)
Enabled Qwen3-TTS Model Runner V2 and made it the experimental default in the bundled CUDA profile. Enabled event-driven orchestration by default for Qwen3-TTS. (#7781, #7930, #7088)
Improved CosyVoice3 flow batching and incremental vocoder streaming, and accelerated MOSS-TTS on Ascend with incremental depth-transformer KV caching, NPUGraph codec decoding, and disaggregated deployment profiles. (#4876, #7521, #6967, #7280, #7052)

Diffusion, Image & Video Generation

Added mid-stream camera interaction for LingBot World 2, plus Ulysses sequence parallelism and streaming VAE decoding distributed across Ulysses ranks. (#7198, #6841, #7651)
Advanced the streaming video output pipeline with session-owned VAE decoding, bounded worker-side encoding, and asynchronous device-to-host chunk transfers through reusable pinned buffers. (#6533, #7018, #7048, #7406)
Expanded MiniMax-H3 with latent-mask editing, long-video latent continuation with driving audio, and VAE encoders in disaggregated deployments. (#7465, #7838, #6939)
Migrated MammothModa2’s DiT stage to the shared diffusion runtime and enabled Cache-DiT acceleration. Added SeaCache support for Cosmos3. (#7134, #7291, #6922)

Quantization & Memory Efficiency

Added Cosmos3 mixed W8A8/W8A16 and W4A4/W4A16 denoising, online FP8 support for the Boogu-Image MLLM and LingBot World, and FP8 KV caching for MammothModa2’s AR stage. (#6560, #7342, #7549, #7436)
Improved HSDP startup with rank-0 shared weight loading, accelerated LoRA delta computation, and reduced MiniMax-H3 video VAE memory peaks. (#7005, #7590, #7241)
Expanded Wan2.2 quantized execution on Ascend with MXFP4 options, W4A8 fallback, and MindIE dense and sparse quantized attention. (#7210, #7561)

Serving, Frontend & API Behavior

Added opt-in multi-process API serving through --api-server-count, allowing supported local EngineCore pipelines to share GPU stage engines across frontend workers. (#6923)
Expanded ComfyUI workflows for MiniMax-H3 reference-based generation, upscaling, and latent editing, and added a music-generation node with MiniMax Music 3 support. (#7483, #7472, #7898, #7516)
Added Speech API streaming metrics, duplex performance metrics, and a LingBot World realtime streaming benchmark. (#6853, #7714, #7645)

Breaking Changes

Removed first-party Dynin-Omni and dots.tts support, including their model registrations, deployment configurations, and examples. (#7655)
Updated the duplex session protocol. /v1/duplex is now an alias for /v1/realtime?duplex=1; native_duplex and client-selected IDs for new sessions have been removed. New session IDs are server-assigned. (#7413, #7647)
Unknown diffusion configuration fields now raise errors instead of being silently discarded. Supported legacy aliases emit deprecation warnings, and conflicting alias/canonical values are rejected. (#5172)

Note

Qwen3-TTS Model Runner V2 remains experimental. The bundled CUDA profile defaults to V2; other supported backends retain V1. Set model_runner: v1 to opt out. (#7930)
AuK support is offline and non-streaming in this release; an OpenAI Speech API adapter is not included. Breeze-TTS-2 currently supports cfg_scale=1.0 only. (#7385, #7084)
Multi-process API serving remains opt-in and excludes diffusion and remote/headless stages. Native Mooncake KV prefetch is also opt-in, with initial support scoped to HunyuanImage3 on CUDA using Mooncake TCP. (#6923, #7637)

What's Changed

[Refactor] P0.2: Migrate API server helpers out of api_server by @herotai214 in #5453
[CI] Stabilize Qwen3-Omni Server VAD E2E by @LHXuuu in #7356
[CI/Build] Diff-aware source_file_dependencies for CUDA/NPU pipelines by @yenuo26 in #6597
[Core][Diffusion] Add a typed pre-D2H video media contract by @NancyFyong in #6615
[Bugfix] Bound HWR domain initialization lock waits by @hsliuustc0106 in #7128
[Bugfix] Escalate diffusion worker shutdown and retain survivors by @hsliuustc0106 in #7126
[Misc] Add standalone safetensors retention diagnostic by @hsliuustc0106 in #7145
[CI] Isolate layerwise offload memory measurements by @andyluo7 in #6938
[Model] Add Cosmos3 mixed W8A8/W8A16 and W4A4/W4A16 denoising by @wkutak in #6560
[Test] Use public render_jinja_template in MiniCPM-o native template test by @tlysanhuo in #7362
[Bugfix] Fix video prewarm cache retention and cancel-restart delay by @psv666 in #7363
Cosmos3 action policy improvements by @MaciejBalaNV in #6460
[BugFix][CI] Restore diff-aware source filtering for post-merge L3 by @yenuo26 in #7371
[Bugfix] Fail when a diffusion LoRA adapter binds no layer by @Hiro208 in #7349
[Bugfix] Fix host-memory leak on aborted /v1/images/generations (#6462) by @zhang-keliang in #6561
[Refactor] Declare model-local KV held outside the paged manager by @linyueqian in #6171
[Realtime] Emit current (non-beta) OpenAI audio/transcript event names by @NickCao in #7339
[Bugfix][Core] Clean up failed HWR atomic metadata writes by @BANANASJIM in #6956
[Bugfix] Keep MiniMax-H3 reference audio budgets separate by @david6666666 in #7281
[Bugfix] Fix Helios USP: per-component split for correct sequence parallelism by @yancaocn in #6930
[Perf][Diffusion] Optimize HSDP startup via Rank-0 shared weight loading and accelerated LoRA delta computation by @SamitHuang in #7005
[Example] Migrate HunyuanImage-3.0 to model_extras + shared task examples by @suyanli220 in #5559
[Model] Avoid scalar synchronizations in GLM-Image preparation by @yuweih205 in #7172
[Model][ERNIE-Image] Delay AdaLN modulation broadcast by @yuweih205 in #7171
[Kernel][MiniMax-H3] Run Q/K RMSNorm-RoPE in one launch by @yuweih205 in #7167
[CI][ROCm] Align AMD image with vLLM 0.29 by @andyluo7 in #7395
[Bugfix] Add embed_multimodal to MiniCPM-o 4.5 omni LLM class by @Hiro208 in #7384
[Model] Add LingBot World Ulysses sequence parallelism by @wtz2333 in #6841
[Feature][TTS] Add Speech API streaming metrics by @gxxx-hum in #6853
[Bugfix][Model] Fix FLUX.2 Klein multi-image edit metadata by @kuafou in #7430
[BugFix] Fix leftovers of the legacy OpenAI realtime API event names by @NickCao in #7426
[Model] Add Tencent AuK speech generation and editing (encoder + diffusion pipeline) by @linyueqian in #7385
[XPU][Docker] Align XPU image and CI with vLLM v0.29.0 by @Joshna-Medisetty in #7441
[Bugfix] Add explicit error when using CFGP with distilled Cosmos3 models by @MaciejBalaNV in #7427
[Perf][Diffusion] Run MammothModa2 DiT attention through the shared attention layer by @MrlixiangWE in #7094
[Bugfix] Give model CLI flags typed owners in the Omni config by @Hiro208 in #7390
[Bugfix] Require a model for vllm serve --omni (fixes #4158) by @abinggo in #4167
[Bugfix] Send a downstream terminal chunk when a parked stage ends by @psv666 in #6889
[NPU] upgrade to v0.29.0 by @FayeSpica in #7433
[Bugfix][Model][Lance] Support decoded video frames in video editing by @junpengw67-max in #5128
[Refactor][Diffusion] Remove model-specific names from LoRA and ModelOpt loader defaults by @congw729 in #5907
Optimize CosyVoice3 Stage1 flow batching by @gerayking in #4876
[3/N] Encode streamed video on the worker with bounded batching by @specture724 in #7018
[Kernel][Boogu-Image] Fuse Q/K RMSNorm + interleaved RoPE via fused_qk_norm_rope by @Holworth in #6982
[Bugfix][Frontend] Honor output_compression on the image generations route by @hsliuustc0106 in #7447
[Core] Add NIXL omni connector by @yuanwu2017 in #6093
[AR-Diffusion] Add session-owned streaming VAE decode by @linzhenpl07 in #6533
[Perf][Frontend] Chunk in-memory image file responses by @kuafou in #7459
[Doc][NPU] Keep AllGather for MiniMax-H3 INT8 DLO by @KrystalRay in #7444
[P0][RFC #3747] RL rollout serving: session management + world_model_env by @Srinivasoo7 in #4770
[Bugfix][NPU] Materialize causal mask in FlashAttention dense path by @Oliver7th in #7324
[CI/Build] Align CUDA release image with vLLM 0.29 by @LiquidGunay in #7445
[Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #7490
[Frontend] Add ComfyUI FastH3 node, fix t2va aspect ratio and dropped audio by @princepride in #7456
[CI/Build] Add experimental AMD MI300 nightly lane by @andyluo7 in #6978
[CI/Build] Share nightly CUDA YAML between H100/L4 and B200 by @yenuo26 in #7028
[Performance][MammothModa2] Enable and validate FP8 KV cache for AR stage by @JiahengX in #7436
[Bugfix] Forward Qwen3-Omni streaming video sampling parameters by @psv666 in #7325
[Bugfix][MiniCPMO45] Return bare tensor in Thinker forward to fix repetition loop (#7497) by @BeatSeat in #7517
fix(qwen-image): drop the shadowed dead _get_qwen_prompt_embeds duplicate by @Anai-Guo in #7461
[Bugfix][Frontend] Preserve duplex chat audio format and duration by @NolenLiang in #7316
[Model] Add Breeze-TTS-2 two-stage AR TTS support (talker + codec, streaming) by @lzwnoname in #7084
[Qwen2.5-Omni] Avoid redundant embedding computation by @nodeeeeee in #7477
[Bugfix][Qwen3-Omni] Fix synchronous Thinker pipeline parallelism by @dshah1333 in #7345
[Model][MiniMax-H3] Multi-GPU serving fixes and layer-wise offload component selection by @heyuanliu-intel in #7047
[Bugfix] Allow Fish Speech phoneme-control tokens through normalize_fish_speech_text by @pujitha24 in #7043
[Bugfix][Hardware][Ascend] Fix Qwen3-TTS code predictor dtype and nested graph replay by @yadongtan in #6639
[BugFix] Fix importing renamed module vllm.entrypoints.openai.engine.… by @NickCao in #7531
[Tests] Skip global GPU cleanup for parallel online diffusion tests by @NickCao in #7534
[Misc][HiggsAudioV3] Diagnose skipped reference audio substitution by @LOGO127 in #7098
[Model] Stop MiniCPM-o Thinker at TTS boundaries by @vuuihc in #7463
[Performance] Optimize MammothModa2 AR → DiT hidden-state transfer by @kunkunblueberry in #7102
[Hardware][Ascend][Model] Support MOSS-TTS-Nano on Ascend A2 by @Big2Wheel in #7192
[Kernel] Fuse Ming streaming ISTFT overlap-add and normalization by @yashkgp in #7338
[Bugfix][MiniCPM-o] Attach a unit's whole frame group and size the vision slot budget like Stage0 by @twu3202 in #7271
[Model] Add VoxCPM2 startup LoRA adapter support by @Bezdarnost in #7108
[Bugfix][Qwen3-TTS] Gate async-chunk Code2Wav on non_streaming_mode (#4371) by @smartDream-chao in #6898
[Frontend][Benchmark] Add duplex performance metrics for OmniInteract and Omni-DuplexEval benchmark by @ZacheryAU in #7242
[Bugfix] Don't shutdown Engine on py_generator=True by @alex-jw-brooks in #6334
[Bugfix] Use current OpenAI audio event names in duplex client metrics by @linyueqian in #7541
[Refactor][Diffusion] Complete generic backend plan consumption by @specture724 in #7313
[Bugfix] Restore the async PP sampled-token broadcast in the AR runner by @dshah1333 in #7393
[CI][ROCm] Add non-blocking engine and model executor GPU coverage by @andyluo7 in #7397
[Frontend] Add MiniMax H3 video upscale workflow (WF-07) by @xuexueligao in #7472
[Lingbot World] Bound image condition storage and extend temporal RoPE by @wtz2333 in #6838
[Diffusion][Perf] Add Qwen-Image QK RoPE Triton path by @dongbo910220 in #5931
[Tests] Test image edit input limits in tiny model framework by @NickCao in #7434
[Model] Add Gepard-1.0 OpenAI /v1/audio/speech serving by @mjZhaoElaine in #7499
[Perf][Cosmos3] Add SeaCache support for Cosmos3 by @yzhautouskay in #6922
[CI][ROCm] Add non-blocking entrypoint GPU coverage by @andyluo7 in #7398
[Bugfix][Core] Keep sample-rate snapshots bounded in delta output by @LOGO127 in #7448
[NPU][CosyVoice3] Add CPU STFT fallback by @peterDengcx in #7520
[CI][Bugfix] Update Qwen3-TTS CI Validation Message by @FayeSpica in #7276
[Bugfix] Route Music3 float32 attention to SDPA by @liuyao0322 in #7354
[Voxtral] Raise BadRequest for ref_audio by @clodaghwalsh17 in #7012
[Bugfix] Fix Qwen3-Omni streaming crops in mixed-length batches by @dshah1333 in #7340
[Bugfix] Tolerate a bounded TTS tail in the audio-text similarity gate by @tlysanhuo in #7358
[Bugfix][Wan2.2] Apply UniPC flow shift once by @HAAZZZEEEE in #7576
[CI/Build][Hardware][Ascend] Add torchmetrics to the NPU CI image by @GodHu777777 in #7570
[Bugfix] Align video frame consumption reporting with prompt sampling by @psv666 in #7329
[Bugfix] Per-key accumulation strategy for Qwen3-TTS codec frame outputs by @linyueqian in #7608
[Bugfix] Keep async video jobs queued until inference and abort on delete by @ZhengWG in #6759
[Feature][Qwen3-TTS] Make Code2Wav honor the stage dtype by @gxxx-hum in #6059
[Diffusion] MAGI-2: regional compile coverage by @cr-gao in #7174
[Bugfix] Skip generic text-only warmup for DreamZero by @tzhouam in #7548
[LingBot World] Reuse the text encode for a prompt already encoded by @linzhenpl07 in #6845
[Core][Frontend] Unified Full-duplex Framework by @chickeyton in #7413
[Attention] [P1] Keep H3 VSA policy in the model and share sparse primitives by @lishunyang12 in #7535
[Bugfix][MiniCPM-o] Stabilize code2wav CFM CUDA graph caching with reference audio normalization and mel-frame bucketing by @y-null in #7416
[Frontend] Add MiniMax-H3 references and Ref2VA workflow by @LinzeShi in #7483
[feat] Extend disaggregated MiniMax-H3 stage with VAE encoders by @asukaqaq-s in #6939
[Frontend] Add MiniMax H3 ComfyUI text-to-video workflow (WF-01) by @zhuhu00 in #7423
[Boogu-Image] Fuse QKV/FFN projections + switch to diffusion RMSNorm by @Bounty-hunter in #6649
[Diffusion] Fuse Qwen-Image-Edit select01 modulation for CUDA by @dongbo910220 in #5921
[Bugfix][Omni-DuplexEval] Load local Hugging Face dataset layouts in duplex eval loader by @RyanYun09 in #7331
[Bugfix] Fail a duplex session that outgrows max_model_len instead of killing the EngineCore by @twu3202 in #7277
[Perf][CosyVoice3] Bounded-window incremental HiFT vocoder streaming by @timzsu in #7521
[Bugfix] Preserve benchmark metric sample counts in results by @psv666 in #7624
[Bugfix][Qwen-Image] Restore RotaryEmbedding CUDA RoPE for Diffusers e2e by @NumberWan in #7513
[CI][MiniCPM-o] Lower the Daily-Omni accuracy gate to 0.77 by @y-null in #7657
[Bugfix] Reject failed diffusion benchmark warmups by @yuweih205 in #7092
[Misc] Remove dead deploy keys from personaplex.yaml by @THUqliu in #7454
[Bugfix][Qwen3-Omni] Include deferred residual in captured Thinker states by @dshah1333 in #7304
[BugFix][NPU] Fix Fish Speech S2 Pro NPU support by @KrystalRay in #7546
[Bugfix][Frontend] Route synthetic aborts through final stage by @CarrotSwordsman in #7006
Preserve audio boundary embeddings in Qwen2.5-Omni interleaved input by @nodeeeeee in #7510
[BugFix][TTS] Fix MOSS Realtime serving and generation by @akshatvishu in #5661
[Bugfix] Fix Breeze-TTS-2 repeated silence and audio EOS handling by @linyueqian in #7667
[Frontend] Add ComfyUI Generate Music node with initial MiniMax Music 3 support by @FayeSpica in #7516
[Bugfix] Fix Qwen3-TTS word timestamps with async chunking by @Bezdarnost in #7544
[Frontend] Allow explicit turn-based serving by @LinzeShi in #7675
[Doc] Scope Spark FP8 guidance to the validated revision by @bojiang-li in #7536
[CI][bugfix] Fix MiniMax H3 FP8 quality test remote-code loading by @david6666666 in #7617
[Bugfix] Update ComfyUI serving-mode mock after helper rename by @LinzeShi in #7694
[Hardware][Ascend] Add Wan2.2 MXFP4 UOS, Smooth and W4A8 fallback by @HAAZZZEEEE in #7210
[Core] Retain structured configs through runtime startup by @maithilijoshi20 in #6849
[Bugfix][NPU] Fix talker_mtp ACL graph capture missing force_uniform_decode by @Wallbreazzz in #7008
[Frontend] Scale Omni serving across API processes by @Sy0307 in #6923
[Model][PersonaPlex] Add paced multi-session validation by @LOGO127 in #7428
[CI] Fix Structured Diffusion Config Wan Test by @alex-jw-brooks in #7707
[Diffusion][Model] Add direct FastH3 8-Step V2 checkpoint loading by @david6666666 in #7610
[Model][PersonaPlex] add per-row RingKV offsets and active rows by @MrDongsls in #7670
[Core] Give the OpenAI Realtime wire codec a reusable home (RFC #6592 P0a) by @psv666 in #7640
[Bugfix][MiniCPM-o] Size the duplex HD slice reservation from the frame by @twu3202 in #7654
[Voxtral] Follow up to raise BadRequest for ref_audio by @clodaghwalsh17 in #7649
[Doc][NPU] Refresh MiniMax-H3 Ascend recipes for 0.28-era configurations by @brandneway in #7720
[Model] Add π0.5 (Pi0.5) VLA model support by @chenchaoxu7575 in #6950
fix(minicpmo45): deadline-align native duplex silence continuation by @Tiagosf00 in #7059
[Qwen3-Omni] Remove redundant text embedding in DeepStack path by @howard-shan in #7573
[Feature][TTS] MOSS-TTS: place the reference-audio encoder on the code2wav stage's GPU by @gcanlin in #7724
[Frontend] Add a shared realtime web UI for MiniCPM-o and Qwen3-Omni (#7222) by @amy-why-3459 in #7585
[4/N] Unify worker-side MP4 encoding and wire Wan into it by @specture724 in #7048
[Kernel] Fuse LongCat paired Q/K RoPE by @dongbo910220 in #7500
[Bugfix][MAGI-2] Fix layerwise offload OOM by pointing offload block attrs to 'block' (#7523) by @BeatSeat in #7540
[Diffusion] Fuse ERNIE-Image Q/K RoPE for eager execution by @dongbo910220 in #7502
[Misc] treewide: remove shm_threshold_bytes from deploy config by @NickCao in #7522
[XPU][HunyuanImage3] Fix HunyuanImage-3.0 text-to-image on XPU by @Joshna-Medisetty in #7674
[Bugfix] Fix Parallel State Initialization For MoE + Diffusion by @alex-jw-brooks in #7676
[5/N] Schedule chunk transfers as their own stage by @specture724 in #7406
[Hardware][Ascend] Integrate MindIE dense and sparse quantized attention for Wan2.2 T2V by @De-cs in #7561
[CI/Build][ROCm] Stabilize shared AMD test lanes by @andyluo7 in #7706
[AR-Diffusion] Bound KV metadata for long rollouts by @wtz2333 in #7498
[Bugfix][MiniMax-H3] Use Hopper-safe modulation precision by @david6666666 in #7693
[Bugfix][Ascend NPU][Diffusion] Load over-wide unquantized fallback weights straight into host memory under DLOFix/mxfp online offload after quant by @brandneway in #7009
[Test][Ascend] Share HunyuanVideo-1.5 E2E across GPU and NPU by @zengchuang-hw in #7547
[Perf][Diffusion] Activate NPU/GPU matmul for LoRA delta computation when weights are on CPU by @holykie in #7590
[CI/Build][MiniCPM-o] Fix the Nightly tests by @chickeyton in #7758
[BugFix][TTS] Restore MOSS v1 codec streaming by @akshatvishu in #6420
[Bugfix][Core] Fix CLI parallel flags losing to nested deploy parallel_config by @zwhzzz0821 in #7786
[CI/Build] Retarget skipped HiDream and MammothModa2 E2E tests to current issues by @yenuo26 in #7755
[CI][Qwen3-Omni] Compare structured-config sampling defaults against vLLM normalization by @linyueqian in #7742
[CI][Qwen3-Omni] Keep the deploy-YAML literal in the structured-config sampling check by @linyueqian in #7803
[Bugfix][Core] Apply Realtime item truncation once per command by @NolenLiang in #7800
[XPU][Bugfix] Make AR-Diffusion KV preallocation device-agnostic by @Joshna-Medisetty in #7777
[Model] Migrate MammothModa2 DiT to the shared diffusion runtime by @Levius-Fubuki in #7134
[XPU][CI] Drop the USTC PyPI mirror from the XPU image build by @Joshna-Medisetty in #7808
[CI/Build] Restrict ERNIE fused RoPE tests to NVIDIA CUDA by @andyluo7 in #7771
[Refactor][Diffusion] Resolve distributed layerwise topology in the plan by @specture724 in #7326
[Core] Refactor omni prefix cache into Manager/Controller by @ZhengWG in #6654
[Bugfix] Use monotonic clocks for test server startup timeouts by @psv666 in #7628
[Refactor] Reject unknown diffusion config fields by @TaffyOfficial in #5172
[Model] Add circlestone-labs/Anima by @akshatvishu in #4083
[Bugfix]Include Reference Audio Cache Salt for HiggsAudio v2 by @sphinxkkkbc in #7826
[Bugfix][Qwen-Image] Keep AutoRound W4A16 on CPU when offloading; Fixes #7555 by @NumberWan in #7579
[Bugfix][Test] Assert stable-audio CPU offload savings on max_memory_allocated by @IneshReddy249 in #6826
[Bugfix] Register statistics for dynamically added replicas by @leegangtoe in #7298
fix(e2e): re-point HunyuanImage3 offline test to text_to_image.py by @chethanuk in #7525
[CI/Build][ROCm] Cache AMD test images in registry by @andyluo7 in #7712
[Bugfix][RAINFUSION_ATTN] Fix padding-mask rejection on packed [real, pad] layouts and reuse query tail as output padding by @brandneway in #7235
[Perf][MiniMax-H3] Cut video VAE encode/decode memory peaks for low-memory serving by @brandneway in #7241
[Hardware][Ascend] Enable VoxCPM2 eager optimizations by @Big2Wheel in #7207
[Bugfix] Avoid multimodal cache collisions in multistage image editing by @princepride in #7817
[Model][Frontend] MiniMax-H3: Add latent-mask editing by @xiaoyu-xyz in #7465
[Bugfix] Stabilize NIXL ownership test claim queries by @zengchuang-hw in #7870
[Misc] Remove Dynin-Omni and dots.tts support by @yenuo26 in #7655
[Bugfix] Honor request-level seed in VoxCPM2 CFM noise by @tlysanhuo in #7866
[Benchmark][MiniCPM-o] Port Video-MME dataset support to Omni bench s… by @amy-why-3459 in #6987
[Refactor][Diffusion] Move single-model layers into their model directories by @congw729 in #5908
[Hardware][Ascend] Use incremental KV-cache decode for MOSS-TTS depth transformer by @jingchengtian in #6967
[Bugfix] Avoid waveform-list GC scans and release completed MOSS batches by @gcanlin in #7885
[Bugfix] Replace ffmpeg subprocess with PyAV to avoid HEVC decoder hang (#7364) by @RyanYun09 in #7504
[Bugfix] Fix Qwen3-Omni realtime playback interruption and conversation history by @psv666 in #7791
[Feat] Add native Mooncake KV transfer from AR to DiT by @asukaqaq-s in #7166
[Bugfix] Route text-only chat as per-request comprehension in HunyuanImage3 AR sampler by @MrlixiangWE in #6111
[Model] Add MiniMax-H3 long-video latent continuation with driving audio by @princepride in #7838
[Core][Diffusion] Camera interaction for diffusion streaming generation (LingBot World 2 as example) by @fhfuih in #7198
[Performance] Capture MOSS-TTS codec streaming decode with NPUGraph by @Wallbreazzz in #7280
feat(comfyui): add MiniMax H3 first and last frame inputs by @avraichur96 in #7449
[Model] Add online FP8 support for Boogu-Image MLLM by @Dmaner in #7342
[Core][Model] Enable Qwen3-TTS MRV2 and optimize the TTS pipeline by @Sy0307 in #7781
[LingBot World] Eleven bit-identical removals of repeated work in the realtime path by @tzhouam in #7648
[Bugfix] Honor the requested output canvas in BAGEL img2img by @nussejzz in #7287
[Bugfix][MiniMax-H3] Fix Hopper keyframe and modulation precision by @david6666666 in #7913
[Bugfix][Engine] Preserve text_encoder_tp_size in OmniEngineArgs by @CarrotSwordsman in #7652
[Feature] Add experimental LingBot World last-step KV reuse by @wtz2333 in #7816
[LingBot World] Shard the streaming VAE decode across the Ulysses ranks by @tzhouam in #7651
[Test] add stability test case for High-priority model by @zhumingjue138 in #7571
[Model] Enable online FP8 linears for LingBot World by @tsinghua-code in #7549
[Frontend] Add MiniMax-H3 latent editing workflows (WF-05) by @FayeSpica in #7898
[Benchmark] Add LingBot-World realtime streaming benchmark by @tzhouam in #7645
[Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #7911
[CI/Build] Promote LingBot-Video dense smoke to ready/merge and wire L4 expansion by @yenuo26 in #7892
[BugFix] Fix diffusion TTS voice error responses by @akshatvishu in #7798
[Bugfix] Preserve NextStep rectangular latent dimensions by @0z5a in #7853
[Bugfix] Isolate weekly CPU tests from inherited attention env and hf_api mocks by @yenuo26 in #7565
[Bugfix] Refresh duplex client defaults after session updates by @wuli666 in #7784
[Model] Integrate AURA into Full-Duplex Runtime by @NumberWan in #7633
[Core][Diffusion] Add pause_generation(mode="keep") for diffusion stages by @cr-gao in #7685
[Model] Default Qwen3-TTS to experimental Model Runner V2 by @Sy0307 in #7930
[Perf][Qwen3-TTS] Enable event-driven orchestration by default by @Sy0307 in #7088
[Bugfix] Keep --no-guardrails out of diffusion stage config by @Levius-Fubuki in #7917
[tts][feature][NPU] Add 2-NPU and 3-NPU disaggregated deploy profiles for MOSS-TTS-Local by @jingchengtian in #7052
[Model] MammothModa2: enable Cache-DiT acceleration for the DiT stage by @shihongzhi in #7291
[Bugfix] Add opt-in MiniCPM-o 4.5 Stage-0 sliding window by @0z5a in #7631
[Misc] Deprecate VLLM_OMNI_VOXCPM_CODE_PATH by @armaanamatya in #7779
[CI][ROCm] Shorten the blocking CosyVoice test path by @andyluo7 in #7748
[Misc] Refactor static stage metadata lookups by @NickCao in #7602
[Bugfix] Catch asyncio.TimeoutError so duplex serving starts on Python 3.10 by @twu3202 in #7412
[Frontend] Reject built-in name collisions in voice upload and add VLLM_OMNI_SPEAKER_REGISTRATION_POLICY by @hoseung2 in #6848
[Frontend] Remove obsolete experimental duplex client parameters by @natureofnature in #7647
[Bugfix] Reject stray <|AUDIO|> placeholder with use_audio_in_video in Qwen2.5-Omni by @LiRunGuo in #7959
[Bugfix][Model] Fix CoVo-Audio prompt processing and dummy loading by @jeffaa729 in #7909
[Bugfix][TTS] Fix Audio8 voice clone reference resolution by @liuyihua95 in #7815
[Core] Optimize Qwen2.5-Omni Token2Wav buffer loading by @akshatvishu in #7061
[Bugfix] Fix GLM-TTS DiT float32 crash with flash-attn backends by @Sworol in #7756
[CI][Perf] migrate diffusion DFX benches to vllm bench serve --omni by @yenuo26 in #7737
[Doc][skip ci] Add design doc for scheduler-managed diffusion paged KV cache by @Acerak01-fy in #6891
[Misc] Add portable MiniMax H3 skills by @princepride in #7923
[Model] Reduce MOSS TTS reference audio preparation overhead by @Sy0307 in #7725
[CI] Count VLLM_OMNI_SPEAKER_REGISTRATION_POLICY in the inventory snapshot by @armaanamatya in #7977
[Bugfix] Keep resumable duplex prompts out of async-chunk prewarm by @twu3202 in #7992
[CI/Build] Give diffusion CPU tests more timeout headroom by @cr-gao in #7983
[Bugfix] Stabilize NIXL manager receiver tests by @cr-gao in #7984
[Feat] Add prefetch for Mooncake cross-stage paged KV transfer in HunyuanImage3 by @Acerak01-fy in #7637
[Frontend][Benchmark] Add duplex performance metrics to refactored full-duplex framework by @ZacheryAU in #7714
[CI] Bump the public env var snapshot count after #6848 landed by @linyueqian in #7964
[CI][Voxtral] Resolve failing test_prepare_speech_generation_awaits_voxtral_async by @clodaghwalsh17 in #7601
[Bugfix] Mark Omni chat benchmark streams with error events as failed by @Levius-Fubuki in #8000
[Bugfix] Restore GLM-Image unshifted timesteps after set_timesteps by @cs-fisha in #7953
[Bugfix] Mark image edit benchmark stream errors as failed by @Levius-Fubuki in #8015
[Doc] Correct the SessionClosed admission-slot guarantee (#7636 Issue 22) by @yanbao1217 in #7869
[Rebase] Rebase to vLLM 0.30.0 by @tzhouam in #7820

New Contributors

@yancaocn made their first contribution in #6930
@kuafou made their first contribution in #7430
@junpengw67-max made their first contribution in #5128
@gerayking made their first contribution in #4876
@Holworth made their first contribution in #6982
@yuanwu2017 made their first contribution in #6093
@Srinivasoo7 made their first contribution in #4770
@Oliver7th made their first contribution in #7324
@LiquidGunay made their first contribution in #7445
@JiahengX made their first contribution in #7436
@NolenLiang made their first contribution in #7316
@lzwnoname made their first contribution in #7084
@nodeeeeee made their first contribution in #7477
@yadongtan made their first contribution in #6639
@vuuihc made their first contribution in #7463
@Big2Wheel made their first contribution in #7192
@Bezdarnost made their first contribution in #7108
@xuexueligao made their first contribution in #7472
@yzhautouskay made their first contribution in #6922
@peterDengcx made their first contribution in #7520
@GodHu777777 made their first contribution in #7570
@LinzeShi made their first contribution in #7483
@zhuhu00 made their first contribution in #7423
@RyanYun09 made their first contribution in #7331
@THUqliu made their first contribution in #7454
@chenchaoxu7575 made their first contribution in #6950
@Tiagosf00 made their first contribution in #7059
@howard-shan made their first contribution in #7573
@De-cs made their first contribution in #7561
@holykie made their first contribution in #7590
@Levius-Fubuki made their first contribution in #7134
@leegangtoe made their first contribution in #7298
@chethanuk made their first contribution in #7525
@xiaoyu-xyz made their first contribution in #7465
@avraichur96 made their first contribution in #7449
@tsinghua-code made their first contribution in #7549
@shihongzhi made their first contribution in #7291
@armaanamatya made their first contribution in #7779
@hoseung2 made their first contribution in #6848
@LiRunGuo made their first contribution in #7959
@jeffaa729 made their first contribution in #7909
@liuyihua95 made their first contribution in #7815
@Sworol made their first contribution in #7756
@cs-fisha made their first contribution in #7953
@yanbao1217 made their first contribution in #7869

Full Changelog: v0.29.0rc1...v0.30.0rc1