v0.30.0

## Highlights

This release features 504 merged changes from 183 contributors, including 94 new contributors.

vLLM-Omni `v0.30.0` is led by three major advances: 1) a **unified full-duplex serving framework**, 2) **native cross-stage KV and multimodal payload transfer**, and 3) **interactive world-model and streaming video generation**, 4) **realtime minimax-h3 4-step inference in both Nvidia Blackwell and Ascend NPU 950 series**. The release rebases onto vLLM 0.30.0, expands model support across speech, image, audiovisual generation, and VLA, and improves batching, memory efficiency, and deployment scalability.

### Key Improvements

* **Unified full-duplex serving around engine-owned sessions**, separating duplex orchestration from turn-based execution and introducing shared model-plugin interfaces. MiniCPM-o 4.5 uses the new framework, with AURA adding multimodal interaction and multi-turn history. **(#7413, #7633)**
* **Extended scheduler-managed diffusion KV into cross-stage transfer and reuse**, adding native Mooncake AR-to-DiT handoff, KV prefetch, and HunyuanImage3 cross-request paged prefix caching. Added NIXL transport for complete AR/encoder-to-diffusion conditioning payloads. **(#7166, #7637, #7719, #6264)**
* **Advanced interactive world-model and video execution**, with LingBot World stepwise generation and mid-stream camera control, session-owned streaming VAE decode, and bounded worker-side video encoding and asynchronous chunk transfer. **(#6844, #7198, #6533, #7048, #7406)**

### Core Architecture & Runtime

* Rebased the project onto **vLLM 0.30.0**, including corresponding Ascend NPU integration updates. **(#7820, #8027)**
* Introduced **`DuplexOmni`, `DuplexOmniEngine`, and `DuplexOrchestrator`**, moving session ownership and lifecycle management into the engine. Eligible models can serve duplex sessions and Chat Completions using the same engine and weights. **(#7413)**
* Added **parallel stage initialization** with device-aware admission and locking, allowing independent stages and replicas to initialize concurrently. Centralized Omni configuration resolution and retained structured configuration through runtime startup. **(#5224, #7292, #5140, #6849)**
* Added **NIXL Omni connectors** and complete conditioning-payload transfer, including text, visual/audio latents, and layout metadata for disaggregated MiniMax H3 execution. **(#6093, #6264)**
* Added scheduler-level **`pause_generation(mode="keep")` for request-level diffusion execution**, preserving queued work and acknowledging device synchronization before returning. **(#7685)**

### Model Support

* Added **MAGI-2 Preview** with native text/image-to-video-and-audio generation, synchronized stereo audio, distributed execution, and Cache-DiT integration. **(#5918)**
* Added **Ming-Image-0.1-Design** and **Design-Layer**, supporting text-to-image generation, single-reference image editing, and RGBA layer decomposition. **(#8021)**
* Added **Anima** native text-to-image inference from local single-file checkpoints, and support for **HunyuanImage-3.0-Instruct-Distil**. **(#4083, #4048)**
* Expanded **SenseNova-U1.5-8B-MoT** support with its distilled eight-step LoRA, and added the **Boogu-Image Turbo** native DMD execution path. **(#6516, #6699)**
* Added **Tencent AuK and AuK-Flash** for instruction-driven speech generation and editing, **Audio8 TTS Preview 0.6B**, and **Breeze-TTS-2** with a two-stage streaming AR/codec pipeline. **(#7385, #6157, #7084)**
* Added an **experimental JoyAI-VL-Interaction native multi-stage speech pipeline** and an **experimental Mage-VL duplex adapter and reference-backed offline integration**. **(#5352, #6537)**
* Added **π0.5 (Pi0.5)**, extending vision-language-action model coverage. **(#6950)**

### Audio, Speech & Realtime Serving

* Moved the shared **MiniCPM-o 4.5 and PersonaPlex duplex infrastructure out of the experimental namespace**, and integrated AURA with the unified runtime for vision-only inputs, session history, and overlapping subsequent turns with draining speech output. **(#6196, #7633)**
* Enabled **event-driven orchestration by default for Qwen3-TTS** and selected **experimental Model Runner V2 in its default CUDA deploy profile**. **(#7088, #7781, #7930)**
* Added **server-side VAD for turn-based omni models** and a shared realtime web UI for MiniCPM-o and Qwen3-Omni. **(#6618, #7585)**
* Improved TTS execution with **OmniVoice variable-length attention and request batching**, optimized MOSS-TTS Local batching and streaming codec execution, and bounded-window incremental CosyVoice3 vocoder streaming. **(#6408, #7202, #7521)**
* Added **Gepard-1.0 Speech API serving**, configurable speech output sample rates, and Speech API streaming metrics. **(#7499, #6553, #6853)**

### Diffusion, Image & Video Generation

* Expanded **MiniMax H3** with the full LightX2V Turbo LoRA matrix, direct FastH3 8-Step V2 checkpoint loading, latent-mask editing, and long-video latent continuation with driving audio. **(#7062, #7610, #7465, #7838)**
* Extended **LingBot World** with Ulysses sequence parallelism, distributed streaming VAE decode, and experimental last-step KV reuse. **(#6841, #7651, #7816)**
* Added **BAGEL step execution and continuous batching**, request-level batching for Boogu-Image, and migrated MammothModa2 DiT onto the shared diffusion runtime with request batching and Cache-DiT acceleration. **(#6359, #6968, #7134, #7476, #7291)**
* Expanded **SANA-Video 2B** with tensor, CFG, and sequence parallelism, Cache-DiT, and CPU offload. Added CFG and sequence parallelism for Boogu-Image. **(#5861, #5940, #5882, #6786, #5718)**
* Improved video output handling with **chunked worker-side MP4 encoding and bounded asynchronous device-to-host transfers**, reducing full-video materialization and producer-side blocking. **(#7018, #7048, #7406)**

### Quantization & Memory Efficiency

* Added **component-selective offload policies** and unified diffusion offload topology resolution. Improved HSDP startup through rank-0 shared weight loading and accelerated LoRA delta computation. **(#5929, #7209, #7005)**
* Added **Cosmos3 mixed W8A8/W8A16 and W4A4/W4A16 denoising**, FP8 support for Boogu-Image and its MLLM, and online FP8 linears for LingBot World. **(#6560, #6925, #7342, #7549)**
* Added **MammothModa2 AR-stage FP8 KV cache**, reduced MiniMax H3 VAE memory peaks, and introduced reusable exact projection caching with MiniMax H3 integration. **(#7436, #7241, #7987)**

### 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 processes. **(#6923)**
* Added an initial **RL rollout serving API** with session creation, stepping, reset, close, and status operations for `world_model_env` integration. **(#4770)**
* Expanded **ComfyUI workflows** with FastH3, MiniMax H3 reference inputs, latent editing, and MiniMax Music 3 generation. **(#7456, #7483, #7898, #7516)**
* Extended benchmarking to **image/video endpoints and duplex interaction**, adding OmniInteract and Omni-DuplexEval performance metrics, accuracy evaluation, and detailed diffusion pipeline timings. **(#4728, #7242, #7317, #6822)**

### Platforms & Hardware Coverage

* Expanded **Ascend NPU** support with Wan2.2 MXFP4 and MindIE quantized attention paths, MOSS-TTS-Nano on Ascend A2, and multi-NPU MOSS-TTS Local deployment profiles. **(#7210, #7561, #7192, #7052)**
* Enabled **HunyuanImage-3.0 text-to-image and the LTX-2 family on Intel XPU**, and made AR-diffusion KV preallocation device-agnostic. **(#7674, #7961, #7777)**
* Expanded **ROCm** engine, executor, and entrypoint coverage, and enabled LTX DiffVAE kernel fusions on **NVIDIA SM100/SM103**. **(#7397, #7398, #7350)**

### Breaking Changes

* **`diffusion_batch_size` has been removed.** Use `max_num_seqs` instead. Unknown diffusion configuration fields now raise validation errors rather than being silently discarded. **(#6484, #5172)**
* **Duplex clients must follow the updated session contract.** `/v1/duplex` is an alias of `/v1/realtime?duplex=1`; obsolete experimental client parameters and the `native_duplex` session parameter have been removed. Use `auto_response` instead of the removed `full_duplex` alias. **(#7413, #7647, #8024)**
* **Realtime audio and transcript events now use current, non-beta OpenAI event names.** Clients consuming the previous names must update their handlers. **(#7339, #7426)**
* First-party support for **Dynin-Omni and dots.tts** has been removed. **(#7655)**

### Note

* Qwen3-TTS's default **Model Runner V2 remains experimental and CUDA-specific**; non-CUDA platforms retain V1. Set `model_runner: v1` to opt out. Its event-driven orchestration default is also model-specific and can be overridden with `VLLM_OMNI_EVENT_DRIVEN_ORCH`. **(#7930, #7088)**
* Multi-process API serving is **opt-in and currently limited to supported local EngineCore pipelines**; diffusion stages and remote/headless stages are not supported in this mode. **(#6923)**
* Native Mooncake KV transfer is **opt-in**. The existing transfer path remains available when `kv_transfer_config` is absent. **(#7166)**
* Mage-VL uses its **Transformers/reference implementation**, while the initial JoyAI native speech profile executes stages sequentially. These integrations should not be interpreted as native vLLM execution for Mage-VL or overlapping streaming execution for JoyAI. **(#6537, #5352)**

## What's Changed
* [Diffusion] Add Boogu Edit-Turbo support docs and tests by @xRay2016 in https://github.com/vllm-project/vllm-omni/pull/6701
* [Config] Remove diffusion_batch_size and use max_num_seqs instead. by @LyxWxj in https://github.com/vllm-project/vllm-omni/pull/6484
* [Core] Validate PipelineConfig in `__post_init__` by @alex-jw-brooks in https://github.com/vllm-project/vllm-omni/pull/6829
* [Frontend] Add configurable output sample rate to Speech API by @AuFlow in https://github.com/vllm-project/vllm-omni/pull/6553
* [Test] Add MiniMax-H3 DLO DP2 variants and Turbo LoRA L3 test by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/6556
* [Doc] Update README and docs for v0.28.0 by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/6858
* [Bugfix][Diffusion] Preserve stage attention backend by @AndyZhou952 in https://github.com/vllm-project/vllm-omni/pull/6645
* docs: add concise API server endpoint guide by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/6114
* [Bugfix][Cosmos3] Avoid logging prompts at INFO level by @rahul-steiger-nv in https://github.com/vllm-project/vllm-omni/pull/6913
* [CI][ROCm] Fix deterministic AMD L3 blockers by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/6884
* [Model] Support SenseNova-U1.5-8B-MoT and its distilled 8-step LoRA by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/6516
* [Bugfix] Prevent build_engine_args_dict from mutating stage_config.engine_args by @Arifuzzamanjoy in https://github.com/vllm-project/vllm-omni/pull/6783
* [Bugfix][Ascend NPU][Boogu] Replace complex RoPE with real-valued operations by @JW-L-7 in https://github.com/vllm-project/vllm-omni/pull/6571
* [Bugfix][Diffusion] Keep the GQA ratio when padding Ulysses heads by @Xenoryn in https://github.com/vllm-project/vllm-omni/pull/5716
* [Hardware][Ascend] Use npu_rotary_mul fused kernel for MOSS-TTS codec RoPE by @jingchengtian in https://github.com/vllm-project/vllm-omni/pull/6908
* [CI] Migrate L4 jobs to l4-k8s and split by GPU count by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/6890
* [Model] Add TP and CFG parallelism to SANA-Video 2B by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/5861
* [Feature][MiniMax-H3] Complete VSA and Ulysses support by @princepride in https://github.com/vllm-project/vllm-omni/pull/6909
* [Doc] Refine TRTLLM attention guidance by @bobboli in https://github.com/vllm-project/vllm-omni/pull/6724
* [Bugfix][Diffusion] Register Sana I2V pipeline as video output by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/6953
* [CI] Select mirror_hardwares presets via MIRROR_HW (H100/B200) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/5543
* [Model] Add Cache-DiT and CPU offload support for SANA-Video 2B by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/5882
* [Bugfix][Qwen3-TTS] Reject task mismatches before EngineCore dispatch by @gxxx-hum in https://github.com/vllm-project/vllm-omni/pull/6113
* [Bugfix][TTS] Isolate concurrent StepAudio2 TRT builds by @BANANASJIM in https://github.com/vllm-project/vllm-omni/pull/6957
* [Perf][Bugfix][OmniVoice] Restore float16 serving, and fuse the generator hot loop by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/6317
* [Bugfix] Fix vLLM serve error response imports by @maithilijoshi20 in https://github.com/vllm-project/vllm-omni/pull/6707
* [Bugfix][Cosmos3] Support multi-chunk transfer with distributed VAE by @rahul-steiger-nv in https://github.com/vllm-project/vllm-omni/pull/6920
* [Feature][Diffusion]: CFG parallelism support for Boogu-Image by @nagisa-kunhah in https://github.com/vllm-project/vllm-omni/pull/6786
* [Core] Optimize diffusion LoRA expand accumulation by @0z5a in https://github.com/vllm-project/vllm-omni/pull/6268
* [Feat] Add support and testcase for HunyuanImage-3.0-Instruct-Distil by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/4048
* [model]Add experimental Mage-VL full-duplex adapter and offline integration by @zyforsure in https://github.com/vllm-project/vllm-omni/pull/6537
* [Bugfix][XPU] Add MoTRMSNorm.forward_xpu delegating to forward_native by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/5568
* [Bugfix][Qwen2.5-Omni] Fix two stale call sites in the speech helpers by @Anai-Guo in https://github.com/vllm-project/vllm-omni/pull/6886
* [Bugfix] Validate --stage-overrides shape in parse_stage_overrides (drop field allowlist) by @MciG-ggg in https://github.com/vllm-project/vllm-omni/pull/6230
* [Bugfix][TTS] Make CosyVoice3 TRT plan publication concurrency-safe by @BANANASJIM in https://github.com/vllm-project/vllm-omni/pull/6955
* [Bugfix] Restore Omni health route ownership by @qujing226 in https://github.com/vllm-project/vllm-omni/pull/6723
* [Bugfix][LTX] Restore distilled two-stage video serving by @mglyn in https://github.com/vllm-project/vllm-omni/pull/6847
* [Doc] Add production diffusion model skill by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/6097
* [Bugfix] Preserve BF16 residual semantics in fused MiniMax-H3 modulation by @bobboli in https://github.com/vllm-project/vllm-omni/pull/6878
* [Model] Add BAGEL step execution and continuous batching by @Sky-Trigger in https://github.com/vllm-project/vllm-omni/pull/6359
* [Perf][NemotronVoiceChat] Fast defaults for all three deploy profiles: 2.8x offline, realtime full-duplex audio by @yuekaizhang in https://github.com/vllm-project/vllm-omni/pull/6354
* [Benchmark] Add local OmniInteract performance cases by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/6817
* [Perf][Model] Optimize LTX video output transport by @mglyn in https://github.com/vllm-project/vllm-omni/pull/7000
* [Diffusion] Enable request-level batching for Boogu-Image (T2I) by @ShengleiFu in https://github.com/vllm-project/vllm-omni/pull/6968
* [Bugfix] Release diffusion worker RPC results on ranks that do not reply by @heyuanliu-intel in https://github.com/vllm-project/vllm-omni/pull/6989
* [Perf][DreamZero] Remove redundant CUDA compile guard by @MikeyDong1 in https://github.com/vllm-project/vllm-omni/pull/6983
* [Bugfix] Fix Qwen3-Omni MoE backend selection by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/7019
* [Bugfix] Update Hub kernel dependency floor by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/7020
* [CI][Bugfix] Normalize spoken units in text comparison by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/6947
* [Bugfix][BAGEL] Restore original denoise step count by @Sky-Trigger in https://github.com/vllm-project/vllm-omni/pull/7049
* [Feat][dots.tts] Add Online Serving Support by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/6235
* [Bugfix] Give exact-shape codec models an unpadded input_ids view by @IneshReddy249 in https://github.com/vllm-project/vllm-omni/pull/6775
* [Bugfix] Fix Higgs Audio v3 voice-clone token validation by @LOGO127 in https://github.com/vllm-project/vllm-omni/pull/7065
* [Perf][CI] Use Whisper on GPU when memory permits by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/5675
* [Feature][Moss-TTS] Reference Audio Preprocessing by @BruceLoveDecimal in https://github.com/vllm-project/vllm-omni/pull/4982
* [Bugfix] Fix fallback loading for IndexTTS external models by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/6137
* [Model] Add sequence parallelism to SANA-Video 2B by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/5940
* [Model] Add native MAGI-2 Preview diffusion support by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/5918
* [Core] Graduate MiniCPM-o 4.5 and PersonaPlex full-duplex serving out of experimental by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/6196
* Parallel stage initialization (admission + SH/EX device locks) (Worker refactor 1/N) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/5224
* [Bugfix] Drop the removed diffusion_batch_size kwarg from the stage-init test by @princepride in https://github.com/vllm-project/vllm-omni/pull/7101
* [Bugfix][Diffusion] Fuse distilled LoRA weights before HSDP sharding by @SamitHuang in https://github.com/vllm-project/vllm-omni/pull/6948
* [diffusion][feature] Add component-selective offload policies by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/5929
* [Diffusion] Implement paged AR↔DiT KV connector ("v1", reusing vllm's native mooncake) by @fhfuih in https://github.com/vllm-project/vllm-omni/pull/6310
* [Doc] Fix markdownlint findings in the serving API reference by @MaxFreedomPollard in https://github.com/vllm-project/vllm-omni/pull/7111
* [Model] Add Boogu-Image Turbo native DMD path by @prettygirlisnotme in https://github.com/vllm-project/vllm-omni/pull/6699
* [CI][Bugfix] Use the matching checkpoint for Qwen3-TTS VoiceDesign tests by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/7054
* [Model] FP8 Support For Boogu-Image by @Dmaner in https://github.com/vllm-project/vllm-omni/pull/6925
* [CI/Build] Stabilize Qwen3-TTS speaker-embedding tests against miss-EOS flake by @mjZhaoElaine in https://github.com/vllm-project/vllm-omni/pull/6924
* [Perf][Diffusion] Skip the Ulysses length all-gather when shards are padded equal by @Xenoryn in https://github.com/vllm-project/vllm-omni/pull/5717
* [Feature][MiniMax-H3] Support the full LightX2V Turbo LoRA matrix by @princepride in https://github.com/vllm-project/vllm-omni/pull/7062
* [Bugfix] Fix Hunyuan Image-3 expert mapping contract by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/7021
* [Model] MiniMax-H3: validate versioned text-conditioning handoff by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/6720
* [Model] Preserve MiniCPM-o Talker full attention until capacity by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/6799
* [Feature] Define the runner preprocess phase contract and regression coverage by @xiaobaixia222 in https://github.com/vllm-project/vllm-omni/pull/7136
* [Bugfix][Diffusion] Restore the TRTLLM attention default for the MiniMax-H3 modular alias by @princepride in https://github.com/vllm-project/vllm-omni/pull/7162
* [Core][Diffusion] Refactor prompt updates into a modality-ready interaction framework by @fhfuih in https://github.com/vllm-project/vllm-omni/pull/6294
* [AR-Diffusion] Allocate K and V separately so the compiled path stops cloning the pool by @linzhenpl07 in https://github.com/vllm-project/vllm-omni/pull/6463
* [BugFix][Qwen-Image-Edit] restore txt_seq_lens RoPE width after #3588 regression by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/5586
* [Feature][LingBot World]stepwise execution by @BruceLoveDecimal in https://github.com/vllm-project/vllm-omni/pull/6844
* [Doc] Add A100 recipe for Wan2.2-TI2V-5B by @lucasruan1618 in https://github.com/vllm-project/vllm-omni/pull/7189
* [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7178
* [Bugfix][Diffusion] Canonicalize cache_backend=None to "none" by @zhang-keliang in https://github.com/vllm-project/vllm-omni/pull/7041
* [Bugfix][Diffusion] Include guidance_scale_2_provided in the request-batch key by @ShengleiFu in https://github.com/vllm-project/vllm-omni/pull/7078
* [Bugfix][Diffusion] Make Cosmos3 transfer image arrays writable by @rahul-steiger-nv in https://github.com/vllm-project/vllm-omni/pull/6915
* [Frontend][Model] Add Cosmos3 control uploads to video API by @FredHuangNV in https://github.com/vllm-project/vllm-omni/pull/7027
* [Bugfix][Cosmos3] Auto-pad non-divisible Ulysses GEN sequences by @rahul-steiger-nv in https://github.com/vllm-project/vllm-omni/pull/6918
* [Model][Perf] Optimize LTX Ulysses sequence parallelism by @mglyn in https://github.com/vllm-project/vllm-omni/pull/7079
* [skip ci][Doc] Module design doc for diffusion runtime by @fhfuih in https://github.com/vllm-project/vllm-omni/pull/6440
* [Bugfix][MiniMax-H3] Stabilize keyframe VAE encoding by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7191
* [Perf][Diffusion] Keep LoRA weights resident across activation cycles by @princepride in https://github.com/vllm-project/vllm-omni/pull/7195
* [Bugfix][NPU] Fix MiniMax-H3 INT8 quantization dispatch by @KrystalRay in https://github.com/vllm-project/vllm-omni/pull/6876
* [Bugfix][MiniCPM-o] Allow ragged audio_feature_lens across a batch by @pujitha24 in https://github.com/vllm-project/vllm-omni/pull/7071
* [Bugfix][Cosmos3] Fix distributed Transfer output envelope by @rahul-steiger-nv in https://github.com/vllm-project/vllm-omni/pull/7205
* [1/N] Stream Wan VAE chunks to the media consumer by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7016
* [Model] Add Boogu-Image Turbo pipeline resolution by @Jerry2423 in https://github.com/vllm-project/vllm-omni/pull/6961
* [Migrate]Move model-specific helpers out of serving_speech by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/6697
* [Refactor] fish_speech: vendor DAC codec modules, drop external fish-speech dependency by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/4871
* [Frontend] Add omni benchmark support for image and video endpoints by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/4728
* [Bugfix][Qwen3-Omni] Read code-predictor RoPE theta from rope_parameters by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7228
* [Refactor 1/N] Centralize Omni config resolution by @fake0fan in https://github.com/vllm-project/vllm-omni/pull/5140
* [Bugfix] Fix Step-Audio2 L4 CI KV cache allocation by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/6839
* [Perf]Boogu-Image Skip dense attention masks by @AbelSara in https://github.com/vllm-project/vllm-omni/pull/6871
* [Bugfix] Keep --disable-log-stats out of stage config resolution by @fake0fan in https://github.com/vllm-project/vllm-omni/pull/7237
* [Bugfix] Bound omni benchmark per-request timeout to 15 min by default by @tlysanhuo in https://github.com/vllm-project/vllm-omni/pull/7130
* [CI/Build] Wait for duplex reaper recovery in retry test by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7225
* [Bugfix] Fix video benchmark timeout and failure accounting by @congw729 in https://github.com/vllm-project/vllm-omni/pull/7259
* [Frontend] Add server-side VAD for turn-based omni models by @LHXuuu in https://github.com/vllm-project/vllm-omni/pull/6618
* [Bugfix][Model] GR00T-N1.7: honor processor_config image pipeline (use_albumentations / crop_fraction / letter_box_transform) to match Isaac-GR00T by @liangmenghuang in https://github.com/vllm-project/vllm-omni/pull/7083
* [Model][JoyAI-VL-Interaction] Add native multi-stage pipeline by @ShuoleiWang in https://github.com/vllm-project/vllm-omni/pull/5352
* [Bugfix] Preserve additional_config for LLM stages by @AbelSara in https://github.com/vllm-project/vllm-omni/pull/7272
* [Bugfix][Examples] Use --profiler-config flag in offline TTS examples by @Asthenia0412 in https://github.com/vllm-project/vllm-omni/pull/6763
* [Bugfix] Skip HWR store-size scans when no limit is configured by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7131
* [CI][ROCm] Route LTX2 Ulysses parity to two-GPU lane by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7234
* [Bugfix][Model] GR00T-N1.7: honor the per-request seed for flow-matching noise by @liangmenghuang in https://github.com/vllm-project/vllm-omni/pull/7253
* Add vLLM-Omni library info to Hugging Face Hub requests by @hmellor in https://github.com/vllm-project/vllm-omni/pull/5381
* [Bugfix][NPU] Limit MiniMax H3 modulation grid size by @KrystalRay in https://github.com/vllm-project/vllm-omni/pull/6794
* [Bugfix] Build the forced-aligner prompt without a chat template (word timestamps one bin late) by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7240
* [Refactor][Diffusion] Resolve offload topology through one plan resolver by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7209
* [Doc] Add AI usage policy for contributions by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7305
* [Bugfix][MiMo-Audio] Align code2wav decode with tokenizer device by @smartDream-chao in https://github.com/vllm-project/vllm-omni/pull/6539
* [Bugfix][MiniCPM-o] Fix the audio_embeds input path by @eval-dev in https://github.com/vllm-project/vllm-omni/pull/5730
* [Feat][OmniVoice]Support Varlen Attn,  Request-Batch and Step-Execution by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/6408
* [Model] Add Audio8 TTS Preview 0.6B (DualAR, 44.1 kHz codec) by @NancyFyong in https://github.com/vllm-project/vllm-omni/pull/6157
* [Bugfix][Frontend] Accept the msgpack-numpy package's numpy markers on the OpenPI endpoint by @ZJLi2013 in https://github.com/vllm-project/vllm-omni/pull/6051
* [Frontend] Opt-in WebSocket TTS split_granularity and session seed by @rk9595 in https://github.com/vllm-project/vllm-omni/pull/7046
* [Bugfix][Frontend] Clear the P0 multimodal cache through the renderer by @ZenAlexa in https://github.com/vllm-project/vllm-omni/pull/7003
* [Bugfix][Frontend] Enforce image pixel limits for video input references by @BANANASJIM in https://github.com/vllm-project/vllm-omni/pull/6963
* [Bugfix][TTS] Isolate shared Higgs v3 reference encode from request cancellation by @EchoHayate in https://github.com/vllm-project/vllm-omni/pull/7076
* [Bugfix][CosyVoice3] Resolve hash snapshot pipeline by @XuTianle0101 in https://github.com/vllm-project/vllm-omni/pull/6896
* [CI] Skip Qwen3-Omni Server VAD multi-turn realtime test (#7279) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7314
* [Bugfix][Magi2] Allow import without an active Triton driver by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7239
* [Core] Split Omni connector model runner mixin by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/6903
* [Bugfix] Make LTX vocoder decoding deterministic by @mglyn in https://github.com/vllm-project/vllm-omni/pull/7231
* [Doc] [Recipe] Add FLUX.1-schnell recipe for RTX 5090 32GB by @Sparks-M in https://github.com/vllm-project/vllm-omni/pull/7299
* [Doc] Qwen3-TTS: add 0.6B on 1x A100 40GB by @chi030303 in https://github.com/vllm-project/vllm-omni/pull/7289
* [Perf][Model] Add optimized LTX-2.5 DiffVAE operators by @mglyn in https://github.com/vllm-project/vllm-omni/pull/7308
* [2/N] Add a minimal temporal chunk callback for MiniMax-H3 by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7017
* [Feature][Diffusion] Expose detailed pipeline timings by @bobboli in https://github.com/vllm-project/vllm-omni/pull/6822
* [Bugfix] Resolve #6931 hub FA3 on torch 2.13 via kernels 0.16.1 by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7185
* [Bugfix][Ascend] fix npu 310/a5 bugs by @zyz111222 in https://github.com/vllm-project/vllm-omni/pull/6685
* [Bugfix][Engine] Group overlapping device stages into one sequential init component by @ZhengWG in https://github.com/vllm-project/vllm-omni/pull/7328
* fix: reserve Qwen3-Omni NVFP4 backend fix by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/7200
* [BugFix] Add field validators for /v1/audio/generate request by @Shaun-Walsh in https://github.com/vllm-project/vllm-omni/pull/4741
* [CI][ROCm] Match CUDA/NPU L2/L3 label routing by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/6966
* [CI/Build] Avoid duplicate stage CLI deploy config by @CarrotSwordsman in https://github.com/vllm-project/vllm-omni/pull/7007
* [CI/Build][ROCm] Normalize SenseNova paged-decode hardware markers by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/6935
* [Model] Skip unused frame packing in Wan2.2 S2V by @yuweih205 in https://github.com/vllm-project/vllm-omni/pull/7155
* [Doc] Add dual DGX Spark MiniMax-H3 results by @bojiang-li in https://github.com/vllm-project/vllm-omni/pull/7343
* [Model] Optimize MOSS-TTS Local batched execution and streaming codec by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7202
* [Bugfix][XPU] Restore N-D output shape for W8A16 FP8 linear by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7301
* [Doc] Document num_outputs_per_prompt for /v1/videos by @Hiro208 in https://github.com/vllm-project/vllm-omni/pull/7341
* [Skills] Add perf-evidence isolation, stage-attribution, and realtime-contract requirements by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/6820
* [Bugfix] Allow LLM replicas on different GPUs to initialize concurrently by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/7292
* [CI/Build] Stabilize LTX2 vocoder autocast test on ROCm by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7336
* [NPU][CI] Add A5 and 310P CI support by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/6875
* [Kernel] Enable LTX DiffVAE fusions on SM100 and SM103 by @mglyn in https://github.com/vllm-project/vllm-omni/pull/7350
* [Bugfix][MiniCPM-o] Align structured chat content with native omni rendering by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7344
* [Rebase] Rebase to vLLM 0.29.0 by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7230
* [Refactor] P0.2: Migrate API server helpers out of api_server by @herotai214 in https://github.com/vllm-project/vllm-omni/pull/5453
* [CI] Stabilize Qwen3-Omni Server VAD E2E by @LHXuuu in https://github.com/vllm-project/vllm-omni/pull/7356
* [CI/Build] Diff-aware source_file_dependencies for CUDA/NPU pipelines by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/6597
* [Core][Diffusion] Add a typed pre-D2H video media contract by @NancyFyong in https://github.com/vllm-project/vllm-omni/pull/6615
* [Bugfix] Bound HWR domain initialization lock waits by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7128
* [Bugfix] Escalate diffusion worker shutdown and retain survivors by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7126
* [Misc] Add standalone safetensors retention diagnostic by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7145
* [CI] Isolate layerwise offload memory measurements by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/6938
* [Model] Add Cosmos3 mixed W8A8/W8A16 and W4A4/W4A16 denoising by @wkutak in https://github.com/vllm-project/vllm-omni/pull/6560
* [Test] Use public render_jinja_template in MiniCPM-o native template test by @tlysanhuo in https://github.com/vllm-project/vllm-omni/pull/7362
* [Bugfix] Fix video prewarm cache retention and cancel-restart delay by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7363
* Cosmos3 action policy improvements by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/6460
* [BugFix][CI] Restore diff-aware source filtering for post-merge L3 by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7371
* [Bugfix] Fail when a diffusion LoRA adapter binds no layer by @Hiro208 in https://github.com/vllm-project/vllm-omni/pull/7349
* [Bugfix] Fix host-memory leak on aborted /v1/images/generations (#6462) by @zhang-keliang in https://github.com/vllm-project/vllm-omni/pull/6561
* [Refactor] Declare model-local KV held outside the paged manager by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/6171
* [Realtime] Emit current (non-beta) OpenAI audio/transcript event names by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7339
* [Bugfix][Core] Clean up failed HWR atomic metadata writes by @BANANASJIM in https://github.com/vllm-project/vllm-omni/pull/6956
* [Bugfix] Keep MiniMax-H3 reference audio budgets separate by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7281
* [Bugfix] Fix Helios USP: per-component split for correct sequence parallelism by @yancaocn in https://github.com/vllm-project/vllm-omni/pull/6930
* [Perf][Diffusion] Optimize HSDP startup via Rank-0 shared weight loading and accelerated LoRA delta computation by @SamitHuang in https://github.com/vllm-project/vllm-omni/pull/7005
* [Example] Migrate HunyuanImage-3.0 to model_extras + shared task examples by @suyanli220 in https://github.com/vllm-project/vllm-omni/pull/5559
* [Model] Avoid scalar synchronizations in GLM-Image preparation by @yuweih205 in https://github.com/vllm-project/vllm-omni/pull/7172
* [Model][ERNIE-Image] Delay AdaLN modulation broadcast by @yuweih205 in https://github.com/vllm-project/vllm-omni/pull/7171
* [Kernel][MiniMax-H3] Run Q/K RMSNorm-RoPE in one launch by @yuweih205 in https://github.com/vllm-project/vllm-omni/pull/7167
* [CI][ROCm] Align AMD image with vLLM 0.29 by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7395
* [Bugfix] Add embed_multimodal to MiniCPM-o 4.5 omni LLM class by @Hiro208 in https://github.com/vllm-project/vllm-omni/pull/7384
* [Model] Add LingBot World Ulysses sequence parallelism by @wtz2333 in https://github.com/vllm-project/vllm-omni/pull/6841
* [Feature][TTS] Add Speech API streaming metrics by @gxxx-hum in https://github.com/vllm-project/vllm-omni/pull/6853
* [Bugfix][Model] Fix FLUX.2 Klein multi-image edit metadata by @kuafou in https://github.com/vllm-project/vllm-omni/pull/7430
* [BugFix] Fix leftovers of the legacy OpenAI realtime API event names by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7426
* [Model] Add Tencent AuK speech generation and editing (encoder + diffusion pipeline) by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7385
* [XPU][Docker] Align XPU image and CI with vLLM v0.29.0 by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7441
* [Bugfix] Add explicit error when using CFGP with distilled Cosmos3 models by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/7427
* [Perf][Diffusion] Run MammothModa2 DiT attention through the shared attention layer by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/7094
* [Bugfix] Give model CLI flags typed owners in the Omni config by @Hiro208 in https://github.com/vllm-project/vllm-omni/pull/7390
* [Bugfix] Require a model for `vllm serve --omni` (fixes #4158) by @abinggo in https://github.com/vllm-project/vllm-omni/pull/4167
* [Bugfix] Send a downstream terminal chunk when a parked stage ends by @psv666 in https://github.com/vllm-project/vllm-omni/pull/6889
* [NPU] upgrade to v0.29.0 by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/7433
* [Bugfix][Model][Lance] Support decoded video frames in video editing by @junpengw67-max in https://github.com/vllm-project/vllm-omni/pull/5128
* [Refactor][Diffusion] Remove model-specific names from LoRA and ModelOpt loader defaults by @congw729 in https://github.com/vllm-project/vllm-omni/pull/5907
* Optimize CosyVoice3 Stage1 flow batching by @gerayking in https://github.com/vllm-project/vllm-omni/pull/4876
* [3/N] Encode streamed video on the worker with bounded batching by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7018
* [Kernel][Boogu-Image] Fuse Q/K RMSNorm + interleaved RoPE via fused_qk_norm_rope by @Holworth in https://github.com/vllm-project/vllm-omni/pull/6982
* [Bugfix][Frontend] Honor output_compression on the image generations route by @hsliuustc0106 in https://github.com/vllm-project/vllm-omni/pull/7447
* [Core] Add NIXL omni connector by @yuanwu2017 in https://github.com/vllm-project/vllm-omni/pull/6093
* [AR-Diffusion] Add session-owned streaming VAE decode by @linzhenpl07 in https://github.com/vllm-project/vllm-omni/pull/6533
* [Perf][Frontend] Chunk in-memory image file responses by @kuafou in https://github.com/vllm-project/vllm-omni/pull/7459
* [Doc][NPU] Keep AllGather for MiniMax-H3 INT8 DLO by @KrystalRay in https://github.com/vllm-project/vllm-omni/pull/7444
* [P0][RFC #3747] RL rollout serving: session management + world_model_env by @Srinivasoo7 in https://github.com/vllm-project/vllm-omni/pull/4770
* [Bugfix][NPU] Materialize causal mask in FlashAttention dense path by @Oliver7th in https://github.com/vllm-project/vllm-omni/pull/7324
* [CI/Build] Align CUDA release image with vLLM 0.29 by @LiquidGunay in https://github.com/vllm-project/vllm-omni/pull/7445
* [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7490
* [Frontend] Add ComfyUI FastH3 node, fix t2va aspect ratio and dropped audio by @princepride in https://github.com/vllm-project/vllm-omni/pull/7456
* [CI/Build] Add experimental AMD MI300 nightly lane by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/6978
* [CI/Build] Share nightly CUDA YAML between H100/L4 and B200 by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7028
* [Performance][MammothModa2] Enable and validate FP8 KV cache for AR stage by @JiahengX in https://github.com/vllm-project/vllm-omni/pull/7436
* [Bugfix] Forward Qwen3-Omni streaming video sampling parameters by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7325
* [Bugfix][MiniCPMO45] Return bare tensor in Thinker forward to fix repetition loop (#7497) by @BeatSeat in https://github.com/vllm-project/vllm-omni/pull/7517
* fix(qwen-image): drop the shadowed dead _get_qwen_prompt_embeds duplicate by @Anai-Guo in https://github.com/vllm-project/vllm-omni/pull/7461
* [Bugfix][Frontend] Preserve duplex chat audio format and duration by @NolenLiang in https://github.com/vllm-project/vllm-omni/pull/7316
* [Model] Add Breeze-TTS-2 two-stage AR TTS support (talker + codec, streaming) by @lzwnoname in https://github.com/vllm-project/vllm-omni/pull/7084
* [Qwen2.5-Omni] Avoid redundant embedding computation by @nodeeeeee in https://github.com/vllm-project/vllm-omni/pull/7477
* [Bugfix][Qwen3-Omni] Fix synchronous Thinker pipeline parallelism by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7345
* [Model][MiniMax-H3] Multi-GPU serving fixes and layer-wise offload component selection by @heyuanliu-intel in https://github.com/vllm-project/vllm-omni/pull/7047
* [Bugfix] Allow Fish Speech phoneme-control tokens through normalize_fish_speech_text by @pujitha24 in https://github.com/vllm-project/vllm-omni/pull/7043
* [Bugfix][Hardware][Ascend] Fix Qwen3-TTS code predictor dtype and nested graph replay by @yadongtan in https://github.com/vllm-project/vllm-omni/pull/6639
* [BugFix] Fix importing renamed module vllm.entrypoints.openai.engine.… by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7531
* [Tests] Skip global GPU cleanup for parallel online diffusion tests by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7534
* [Misc][HiggsAudioV3] Diagnose skipped reference audio substitution by @LOGO127 in https://github.com/vllm-project/vllm-omni/pull/7098
* [Model] Stop MiniCPM-o Thinker at TTS boundaries by @vuuihc in https://github.com/vllm-project/vllm-omni/pull/7463
* [Performance] Optimize MammothModa2 AR → DiT hidden-state transfer by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/7102
* [Hardware][Ascend][Model] Support MOSS-TTS-Nano on Ascend A2 by @Big2Wheel in https://github.com/vllm-project/vllm-omni/pull/7192
* [Kernel] Fuse Ming streaming ISTFT overlap-add and normalization by @yashkgp in https://github.com/vllm-project/vllm-omni/pull/7338
* [Bugfix][MiniCPM-o] Attach a unit's whole frame group and size the vision slot budget like Stage0 by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7271
* [Model] Add VoxCPM2 startup LoRA adapter support by @Bezdarnost in https://github.com/vllm-project/vllm-omni/pull/7108
* [Bugfix][Qwen3-TTS] Gate async-chunk Code2Wav on non_streaming_mode (#4371) by @smartDream-chao in https://github.com/vllm-project/vllm-omni/pull/6898
* [Frontend][Benchmark] Add duplex performance metrics for OmniInteract and Omni-DuplexEval benchmark by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/7242
* [Bugfix] Don't shutdown Engine on py_generator=True by @alex-jw-brooks in https://github.com/vllm-project/vllm-omni/pull/6334
* [Bugfix] Use current OpenAI audio event names in duplex client metrics by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7541
* [Refactor][Diffusion] Complete generic backend plan consumption by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7313
* [Bugfix] Restore the async PP sampled-token broadcast in the AR runner by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7393
* [CI][ROCm] Add non-blocking engine and model executor GPU coverage by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7397
* [Frontend] Add MiniMax H3 video upscale workflow (WF-07) by @xuexueligao in https://github.com/vllm-project/vllm-omni/pull/7472
* [Lingbot World] Bound image condition storage and extend temporal RoPE by @wtz2333 in https://github.com/vllm-project/vllm-omni/pull/6838
* [Diffusion][Perf] Add Qwen-Image QK RoPE Triton path by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/5931
* [Tests] Test image edit input limits in tiny model framework by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7434
* [Model] Add Gepard-1.0 OpenAI /v1/audio/speech serving by @mjZhaoElaine in https://github.com/vllm-project/vllm-omni/pull/7499
* [Perf][Cosmos3] Add SeaCache support for Cosmos3 by @yzhautouskay in https://github.com/vllm-project/vllm-omni/pull/6922
* [CI][ROCm] Add non-blocking entrypoint GPU coverage by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7398
* [Bugfix][Core] Keep sample-rate snapshots bounded in delta output by @LOGO127 in https://github.com/vllm-project/vllm-omni/pull/7448
* [NPU][CosyVoice3] Add CPU STFT fallback by @peterDengcx in https://github.com/vllm-project/vllm-omni/pull/7520
* [CI][Bugfix] Update Qwen3-TTS CI Validation Message by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/7276
* [Bugfix] Route Music3 float32 attention to SDPA by @liuyao0322 in https://github.com/vllm-project/vllm-omni/pull/7354
* [Voxtral] Raise BadRequest for ref_audio by @clodaghwalsh17 in https://github.com/vllm-project/vllm-omni/pull/7012
* [Bugfix] Fix Qwen3-Omni streaming crops in mixed-length batches by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7340
* [Bugfix] Tolerate a bounded TTS tail in the audio-text similarity gate by @tlysanhuo in https://github.com/vllm-project/vllm-omni/pull/7358
* [Bugfix][Wan2.2] Apply UniPC flow shift once by @HAAZZZEEEE in https://github.com/vllm-project/vllm-omni/pull/7576
* [CI/Build][Hardware][Ascend] Add torchmetrics to the NPU CI image by @GodHu777777 in https://github.com/vllm-project/vllm-omni/pull/7570
* [Bugfix] Align video frame consumption reporting with prompt sampling by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7329
* [Bugfix] Per-key accumulation strategy for Qwen3-TTS codec frame outputs by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7608
* [Bugfix] Keep async video jobs queued until inference and abort on delete by @ZhengWG in https://github.com/vllm-project/vllm-omni/pull/6759
* [Feature][Qwen3-TTS] Make Code2Wav honor the stage dtype by @gxxx-hum in https://github.com/vllm-project/vllm-omni/pull/6059
* [Diffusion] MAGI-2: regional compile coverage by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/7174
* [Bugfix] Skip generic text-only warmup for DreamZero by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7548
* [LingBot World] Reuse the text encode for a prompt already encoded by @linzhenpl07 in https://github.com/vllm-project/vllm-omni/pull/6845
* [Core][Frontend] Unified Full-duplex Framework by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/7413
* [Attention] [P1] Keep H3 VSA policy in the model and share sparse primitives by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/7535
* [Bugfix][MiniCPM-o] Stabilize code2wav CFM CUDA graph caching with reference audio normalization and mel-frame bucketing by @y-null in https://github.com/vllm-project/vllm-omni/pull/7416
* [Frontend] Add MiniMax-H3 references and Ref2VA workflow by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/7483
* [feat] Extend disaggregated MiniMax-H3 stage with VAE encoders by @asukaqaq-s in https://github.com/vllm-project/vllm-omni/pull/6939
* [Frontend] Add MiniMax H3 ComfyUI text-to-video workflow (WF-01) by @zhuhu00 in https://github.com/vllm-project/vllm-omni/pull/7423
* [Boogu-Image] Fuse QKV/FFN projections + switch to diffusion RMSNorm by @Bounty-hunter in https://github.com/vllm-project/vllm-omni/pull/6649
* [Diffusion] Fuse Qwen-Image-Edit select01 modulation for CUDA by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/5921
* [Bugfix][Omni-DuplexEval] Load local Hugging Face dataset layouts in duplex eval loader by @RyanYun09 in https://github.com/vllm-project/vllm-omni/pull/7331
* [Bugfix] Fail a duplex session that outgrows max_model_len instead of killing the EngineCore by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7277
* [Perf][CosyVoice3] Bounded-window incremental HiFT vocoder streaming by @timzsu in https://github.com/vllm-project/vllm-omni/pull/7521
* [Bugfix] Preserve benchmark metric sample counts in results by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7624
* [Bugfix][Qwen-Image] Restore RotaryEmbedding CUDA RoPE for Diffusers e2e by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7513
* [CI][MiniCPM-o] Lower the Daily-Omni accuracy gate to 0.77 by @y-null in https://github.com/vllm-project/vllm-omni/pull/7657
* [Bugfix] Reject failed diffusion benchmark warmups by @yuweih205 in https://github.com/vllm-project/vllm-omni/pull/7092
* [Misc] Remove dead deploy keys from personaplex.yaml by @THUqliu in https://github.com/vllm-project/vllm-omni/pull/7454
* [Bugfix][Qwen3-Omni] Include deferred residual in captured Thinker states by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7304
* [BugFix][NPU] Fix Fish Speech S2 Pro NPU support by @KrystalRay in https://github.com/vllm-project/vllm-omni/pull/7546
* [Bugfix][Frontend] Route synthetic aborts through final stage by @CarrotSwordsman in https://github.com/vllm-project/vllm-omni/pull/7006
* Preserve audio boundary embeddings in Qwen2.5-Omni interleaved input by @nodeeeeee in https://github.com/vllm-project/vllm-omni/pull/7510
* [BugFix][TTS] Fix MOSS Realtime serving and generation by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/5661
* [Bugfix] Fix Breeze-TTS-2 repeated silence and audio EOS handling by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7667
* [Frontend] Add ComfyUI Generate Music node with initial MiniMax Music 3 support by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/7516
* [Bugfix] Fix Qwen3-TTS word timestamps with async chunking by @Bezdarnost in https://github.com/vllm-project/vllm-omni/pull/7544
* [Frontend] Allow explicit turn-based serving by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/7675
* [Doc] Scope Spark FP8 guidance to the validated revision by @bojiang-li in https://github.com/vllm-project/vllm-omni/pull/7536
* [CI][bugfix] Fix MiniMax H3 FP8 quality test remote-code loading by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7617
* [Bugfix] Update ComfyUI serving-mode mock after helper rename by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/7694
* [Hardware][Ascend] Add Wan2.2 MXFP4 UOS, Smooth and W4A8 fallback by @HAAZZZEEEE in https://github.com/vllm-project/vllm-omni/pull/7210
* [Core] Retain structured configs through runtime startup by @maithilijoshi20 in https://github.com/vllm-project/vllm-omni/pull/6849
* [Bugfix][NPU] Fix talker_mtp ACL graph capture missing force_uniform_decode by @Wallbreazzz in https://github.com/vllm-project/vllm-omni/pull/7008
* [Frontend] Scale Omni serving across API processes by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/6923
* [Model][PersonaPlex] Add paced multi-session validation by @LOGO127 in https://github.com/vllm-project/vllm-omni/pull/7428
* [CI] Fix Structured Diffusion Config Wan Test by @alex-jw-brooks in https://github.com/vllm-project/vllm-omni/pull/7707
* [Diffusion][Model] Add direct FastH3 8-Step V2 checkpoint loading by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7610
* [Model][PersonaPlex] add per-row RingKV offsets and active rows by @MrDongsls in https://github.com/vllm-project/vllm-omni/pull/7670
* [Core] Give the OpenAI Realtime wire codec a reusable home (RFC #6592 P0a) by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7640
* [Bugfix][MiniCPM-o] Size the duplex HD slice reservation from the frame by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7654
* [Voxtral] Follow up to raise BadRequest for ref_audio by @clodaghwalsh17 in https://github.com/vllm-project/vllm-omni/pull/7649
* [Doc][NPU] Refresh MiniMax-H3 Ascend recipes for 0.28-era configurations by @brandneway in https://github.com/vllm-project/vllm-omni/pull/7720
* [Model] Add π0.5 (Pi0.5) VLA model support by @chenchaoxu7575 in https://github.com/vllm-project/vllm-omni/pull/6950
* fix(minicpmo45): deadline-align native duplex silence continuation by @Tiagosf00 in https://github.com/vllm-project/vllm-omni/pull/7059
* [Qwen3-Omni] Remove redundant text embedding in DeepStack path by @howard-shan in https://github.com/vllm-project/vllm-omni/pull/7573
* [Feature][TTS] MOSS-TTS: place the reference-audio encoder on the code2wav stage's GPU by @gcanlin in https://github.com/vllm-project/vllm-omni/pull/7724
* [Frontend] Add a shared realtime web UI for MiniCPM-o and Qwen3-Omni (#7222) by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/7585
* [4/N] Unify worker-side MP4 encoding and wire Wan into it by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7048
* [Kernel] Fuse LongCat paired Q/K RoPE by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/7500
* [Bugfix][MAGI-2] Fix layerwise offload OOM by pointing offload block attrs to 'block' (#7523) by @BeatSeat in https://github.com/vllm-project/vllm-omni/pull/7540
* [Diffusion] Fuse ERNIE-Image Q/K RoPE for eager execution by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/7502
* [Misc] treewide: remove shm_threshold_bytes from deploy config by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7522
* [XPU][HunyuanImage3] Fix HunyuanImage-3.0 text-to-image on XPU by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7674
* [Bugfix] Fix Parallel State Initialization For MoE + Diffusion by @alex-jw-brooks in https://github.com/vllm-project/vllm-omni/pull/7676
* [5/N] Schedule chunk transfers as their own stage by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7406
* [Hardware][Ascend] Integrate MindIE dense and sparse quantized attention for Wan2.2 T2V by @De-cs in https://github.com/vllm-project/vllm-omni/pull/7561
* [CI/Build][ROCm] Stabilize shared AMD test lanes by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7706
* [AR-Diffusion] Bound KV metadata for long rollouts by @wtz2333 in https://github.com/vllm-project/vllm-omni/pull/7498
* [Bugfix][MiniMax-H3] Use Hopper-safe modulation precision by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/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 https://github.com/vllm-project/vllm-omni/pull/7009
* [Test][Ascend] Share HunyuanVideo-1.5 E2E across GPU and NPU by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/7547
* [Perf][Diffusion] Activate NPU/GPU matmul for LoRA delta computation when weights are on CPU by @holykie in https://github.com/vllm-project/vllm-omni/pull/7590
* [CI/Build][MiniCPM-o] Fix the Nightly tests by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/7758
* [BugFix][TTS] Restore MOSS v1 codec streaming by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/6420
* [Bugfix][Core] Fix CLI parallel flags losing to nested deploy parallel_config by @zwhzzz0821 in https://github.com/vllm-project/vllm-omni/pull/7786
* [CI/Build] Retarget skipped HiDream and MammothModa2 E2E tests to current issues by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7755
* [CI][Qwen3-Omni] Compare structured-config sampling defaults against vLLM normalization by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7742
* [CI][Qwen3-Omni] Keep the deploy-YAML literal in the structured-config sampling check by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7803
* [Bugfix][Core] Apply Realtime item truncation once per command by @NolenLiang in https://github.com/vllm-project/vllm-omni/pull/7800
* [XPU][Bugfix] Make AR-Diffusion KV preallocation device-agnostic by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7777
* [Model] Migrate MammothModa2 DiT to the shared diffusion runtime by @Levius-Fubuki in https://github.com/vllm-project/vllm-omni/pull/7134
* [XPU][CI] Drop the USTC PyPI mirror from the XPU image build by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7808
* [CI/Build] Restrict ERNIE fused RoPE tests to NVIDIA CUDA by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7771
* [Refactor][Diffusion] Resolve distributed layerwise topology in the plan by @specture724 in https://github.com/vllm-project/vllm-omni/pull/7326
* [Core] Refactor omni prefix cache into Manager/Controller by @ZhengWG in https://github.com/vllm-project/vllm-omni/pull/6654
* [Bugfix] Use monotonic clocks for test server startup timeouts by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7628
* [Refactor] Reject unknown diffusion config fields by @TaffyOfficial in https://github.com/vllm-project/vllm-omni/pull/5172
* [Model] Add circlestone-labs/Anima by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/4083
* [Bugfix]Include Reference Audio Cache Salt for HiggsAudio v2 by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/7826
* [Bugfix][Qwen-Image] Keep AutoRound W4A16 on CPU when offloading; Fixes #7555 by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7579
* [Bugfix][Test] Assert stable-audio CPU offload savings on max_memory_allocated by @IneshReddy249 in https://github.com/vllm-project/vllm-omni/pull/6826
* [Bugfix] Register statistics for dynamically added replicas by @leegangtoe in https://github.com/vllm-project/vllm-omni/pull/7298
* fix(e2e): re-point HunyuanImage3 offline test to text_to_image.py by @chethanuk in https://github.com/vllm-project/vllm-omni/pull/7525
* [CI/Build][ROCm] Cache AMD test images in registry by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7712
* [Bugfix][RAINFUSION_ATTN] Fix padding-mask rejection on packed [real, pad] layouts and reuse query tail as output padding by @brandneway in https://github.com/vllm-project/vllm-omni/pull/7235
* [Perf][MiniMax-H3] Cut video VAE encode/decode memory peaks for low-memory serving by @brandneway in https://github.com/vllm-project/vllm-omni/pull/7241
* [Hardware][Ascend] Enable VoxCPM2 eager optimizations by @Big2Wheel in https://github.com/vllm-project/vllm-omni/pull/7207
* [Bugfix] Avoid multimodal cache collisions in multistage image editing by @princepride in https://github.com/vllm-project/vllm-omni/pull/7817
* [Model][Frontend] MiniMax-H3: Add latent-mask editing by @xiaoyu-xyz in https://github.com/vllm-project/vllm-omni/pull/7465
* [Bugfix] Stabilize NIXL ownership test claim queries by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/7870
* [Misc] Remove Dynin-Omni and dots.tts support by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7655
* [Bugfix] Honor request-level seed in VoxCPM2 CFM noise by @tlysanhuo in https://github.com/vllm-project/vllm-omni/pull/7866
* [Benchmark][MiniCPM-o] Port Video-MME dataset support to Omni bench s… by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/6987
* [Refactor][Diffusion] Move single-model layers into their model directories by @congw729 in https://github.com/vllm-project/vllm-omni/pull/5908
* [Hardware][Ascend] Use incremental KV-cache decode for MOSS-TTS depth transformer by @jingchengtian in https://github.com/vllm-project/vllm-omni/pull/6967
* [Bugfix] Avoid waveform-list GC scans and release completed MOSS batches by @gcanlin in https://github.com/vllm-project/vllm-omni/pull/7885
* [Bugfix] Replace ffmpeg subprocess with PyAV to avoid HEVC decoder hang (#7364) by @RyanYun09 in https://github.com/vllm-project/vllm-omni/pull/7504
* [Bugfix] Fix Qwen3-Omni realtime playback interruption and conversation history by @psv666 in https://github.com/vllm-project/vllm-omni/pull/7791
* [Feat] Add native Mooncake KV transfer from AR to DiT by @asukaqaq-s in https://github.com/vllm-project/vllm-omni/pull/7166
* [Bugfix] Route text-only chat as per-request comprehension in HunyuanImage3 AR sampler by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/6111
* [Model] Add MiniMax-H3 long-video latent continuation with driving audio by @princepride in https://github.com/vllm-project/vllm-omni/pull/7838
* [Core][Diffusion] Camera interaction for diffusion streaming generation (LingBot World 2 as example) by @fhfuih in https://github.com/vllm-project/vllm-omni/pull/7198
* [Performance] Capture MOSS-TTS codec streaming decode with NPUGraph by @Wallbreazzz in https://github.com/vllm-project/vllm-omni/pull/7280
* feat(comfyui): add MiniMax H3 first and last frame inputs by @avraichur96 in https://github.com/vllm-project/vllm-omni/pull/7449
* [Model] Add online FP8 support for Boogu-Image MLLM by @Dmaner in https://github.com/vllm-project/vllm-omni/pull/7342
* [Core][Model] Enable Qwen3-TTS MRV2 and optimize the TTS pipeline by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7781
* [LingBot World] Eleven bit-identical removals of repeated work in the realtime path by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7648
* [Bugfix] Honor the requested output canvas in BAGEL img2img by @nussejzz in https://github.com/vllm-project/vllm-omni/pull/7287
* [Bugfix][MiniMax-H3] Fix Hopper keyframe and modulation precision by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7913
* [Bugfix][Engine] Preserve text_encoder_tp_size in OmniEngineArgs by @CarrotSwordsman in https://github.com/vllm-project/vllm-omni/pull/7652
* [Feature] Add experimental LingBot World last-step KV reuse by @wtz2333 in https://github.com/vllm-project/vllm-omni/pull/7816
* [LingBot World] Shard the streaming VAE decode across the Ulysses ranks by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7651
* [Test] add stability test case for High-priority model by @zhumingjue138 in https://github.com/vllm-project/vllm-omni/pull/7571
* [Model] Enable online FP8 linears for LingBot World by @tsinghua-code in https://github.com/vllm-project/vllm-omni/pull/7549
* [Frontend] Add MiniMax-H3 latent editing workflows (WF-05) by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/7898
* [Benchmark] Add LingBot-World realtime streaming benchmark by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7645
* [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/7911
* [CI/Build] Promote LingBot-Video dense smoke to ready/merge and wire L4 expansion by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7892
* [BugFix] Fix diffusion TTS voice error responses by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/7798
* [Bugfix] Preserve NextStep rectangular latent dimensions by @0z5a in https://github.com/vllm-project/vllm-omni/pull/7853
* [Bugfix] Isolate weekly CPU tests from inherited attention env and hf_api mocks by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7565
* [Bugfix] Refresh duplex client defaults after session updates by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/7784
* [Model] Integrate AURA into Full-Duplex Runtime by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7633
* [Core][Diffusion] Add pause_generation(mode="keep") for diffusion stages by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/7685
* [Model] Default Qwen3-TTS to experimental Model Runner V2 by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7930
* [Perf][Qwen3-TTS] Enable event-driven orchestration by default by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7088
* [Bugfix] Keep --no-guardrails out of diffusion stage config by @Levius-Fubuki in https://github.com/vllm-project/vllm-omni/pull/7917
* [tts][feature][NPU] Add 2-NPU and 3-NPU disaggregated deploy profiles for MOSS-TTS-Local by @jingchengtian in https://github.com/vllm-project/vllm-omni/pull/7052
* [Model] MammothModa2: enable Cache-DiT acceleration for the DiT stage by @shihongzhi in https://github.com/vllm-project/vllm-omni/pull/7291
* [Bugfix] Add opt-in MiniCPM-o 4.5 Stage-0 sliding window by @0z5a in https://github.com/vllm-project/vllm-omni/pull/7631
* [Misc] Deprecate VLLM_OMNI_VOXCPM_CODE_PATH by @armaanamatya in https://github.com/vllm-project/vllm-omni/pull/7779
* [CI][ROCm] Shorten the blocking CosyVoice test path by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7748
* [Misc] Refactor static stage metadata lookups by @NickCao in https://github.com/vllm-project/vllm-omni/pull/7602
* [Bugfix] Catch asyncio.TimeoutError so duplex serving starts on Python 3.10 by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7412
* [Frontend] Reject built-in name collisions in voice upload and add VLLM_OMNI_SPEAKER_REGISTRATION_POLICY by @hoseung2 in https://github.com/vllm-project/vllm-omni/pull/6848
* [Frontend] Remove obsolete experimental duplex client parameters by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/7647
* [Bugfix] Reject stray <|AUDIO|> placeholder with use_audio_in_video in Qwen2.5-Omni by @LiRunGuo in https://github.com/vllm-project/vllm-omni/pull/7959
* [Bugfix][Model] Fix CoVo-Audio prompt processing and dummy loading by @jeffaa729 in https://github.com/vllm-project/vllm-omni/pull/7909
* [Bugfix][TTS] Fix Audio8 voice clone reference resolution by @liuyihua95 in https://github.com/vllm-project/vllm-omni/pull/7815
* [Core] Optimize Qwen2.5-Omni Token2Wav buffer loading by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/7061
* [Bugfix] Fix GLM-TTS DiT float32 crash with flash-attn backends by @Sworol in https://github.com/vllm-project/vllm-omni/pull/7756
* [CI][Perf] migrate diffusion DFX benches to vllm bench serve --omni by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7737
* [Doc][skip ci] Add design doc for scheduler-managed diffusion paged KV cache by @Acerak01-fy in https://github.com/vllm-project/vllm-omni/pull/6891
* [Misc] Add portable MiniMax H3 skills by @princepride in https://github.com/vllm-project/vllm-omni/pull/7923
* [Model] Reduce MOSS TTS reference audio preparation overhead by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7725
* [CI] Count VLLM_OMNI_SPEAKER_REGISTRATION_POLICY in the inventory snapshot by @armaanamatya in https://github.com/vllm-project/vllm-omni/pull/7977
* [Bugfix] Keep resumable duplex prompts out of async-chunk prewarm by @twu3202 in https://github.com/vllm-project/vllm-omni/pull/7992
* [CI/Build] Give diffusion CPU tests more timeout headroom by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/7983
* [Bugfix] Stabilize NIXL manager receiver tests by @cr-gao in https://github.com/vllm-project/vllm-omni/pull/7984
* [Feat] Add prefetch for Mooncake cross-stage paged KV transfer in HunyuanImage3 by @Acerak01-fy in https://github.com/vllm-project/vllm-omni/pull/7637
* [Frontend][Benchmark] Add duplex performance metrics to refactored full-duplex framework by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/7714
* [CI] Bump the public env var snapshot count after #6848 landed by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7964
* [CI][Voxtral] Resolve failing test_prepare_speech_generation_awaits_voxtral_async  by @clodaghwalsh17 in https://github.com/vllm-project/vllm-omni/pull/7601
* [Bugfix] Mark Omni chat benchmark streams with error events as failed by @Levius-Fubuki in https://github.com/vllm-project/vllm-omni/pull/8000
* [Bugfix] Restore GLM-Image unshifted timesteps after set_timesteps by @cs-fisha in https://github.com/vllm-project/vllm-omni/pull/7953
* [Bugfix] Mark image edit benchmark stream errors as failed by @Levius-Fubuki in https://github.com/vllm-project/vllm-omni/pull/8015
* [Doc] Correct the SessionClosed admission-slot guarantee (#7636 Issue 22) by @yanbao1217 in https://github.com/vllm-project/vllm-omni/pull/7869
* [Rebase] Rebase to vLLM 0.30.0 by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7820
* [Bugfix] Preserve all video-stream audio deltas by @dshah1333 in https://github.com/vllm-project/vllm-omni/pull/7532
* [Bugfix] Merge audio final output into the matching text choice (#7376) by @CarrotSwordsman in https://github.com/vllm-project/vllm-omni/pull/7879
* [Hardware][Ascend] Add opt-in chunked anti-alias FIR upsampling by @Celeste-jq in https://github.com/vllm-project/vllm-omni/pull/7896
* [optimization] Added Wan 2.2 VAE decode optimizations by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/7056
* [bugfix] Fix FP8 Quack kernels when using Context Parallelism and reduce number of recompilations by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/7458
* [Bugfix][Qwen3-TTS] Ignore voice labels for inline Base cloning by @sc-hua in https://github.com/vllm-project/vllm-omni/pull/6974
* [Model] Accelerate Breeze-TTS-2 non-streaming inference by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/7752
* [Bugfix] Restore MRV2 full-payload transport for offline Qwen3-TTS by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/8010
* [CI/Build] Move Breeze-TTS-2 voice design e2e to weekly by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/8032
* [BugFix][Diffusion] Scope Wan RMSNorm patch to NPU by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/8047
* [Bugfix] Fix Qwen3-TTS prefill probe on MRv2 runner by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/8065
* [Feature] Transfer AR-to-diffusion payloads over NIXL by @yuanwu2017 in https://github.com/vllm-project/vllm-omni/pull/6264
* [New Model] Ming-Image (inclusionAI/Ming-Image-0.1-Design) Support by @yuanheng-zhao in https://github.com/vllm-project/vllm-omni/pull/8021
* [Misc] remove legacy full_duplex alias for auto_response by @NickCao in https://github.com/vllm-project/vllm-omni/pull/8024
* [CI][Omni-DuplexEval] Add nightly scoring CI guard by @RyanYun09 in https://github.com/vllm-project/vllm-omni/pull/7773
* [Bugfix] Fix native Mooncake cross-stage KV layout mismatch by @Acerak01-fy in https://github.com/vllm-project/vllm-omni/pull/8003
* [NPU] upgrade to v0.30.0 by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/8027
* [Bugfix][Core] Preserve static KV staging inputs in AR-Diffusion by @MichaelCao0 in https://github.com/vllm-project/vllm-omni/pull/7921
* [Diffusion][HunyuanImage3] Add request paged prefix caching by @zwhzzz0821 in https://github.com/vllm-project/vllm-omni/pull/7719
* [CI/Build][MiniCPM-o] Gate Seed-TTS WER over chat completions instead of the native duplex path by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/7963
* [Bugfix][HunyuanImage3] Fix DiT LoRA binding: namespace alias + GQA QKV de-interleave by @Asthenia0412 in https://github.com/vllm-project/vllm-omni/pull/6802
* [Bugfix] Gate Qwen-Image fused QK-norm+RoPE on short sequences (#7780) by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7965
* [Bugfix][Diffusion] Drop aborted request's orphaned async output (fix #6413) by @zhang-keliang in https://github.com/vllm-project/vllm-omni/pull/6439
* [Bugfix][Core] Fix stage initialization spawn-lock/device-lock inversion by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7926
* [BugFix] Download Anima assets for README CI test by @akshatvishu in https://github.com/vllm-project/vllm-omni/pull/7928
* [Bugfix] Drain the omni connector output on the main thread in the AR runner by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7750
* [Bugfix] Preserve cancellation in duplex session control queues by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/7782
* [Bugfix] Align MiniMax-H3 inference steps with denoiser evaluations by @AbelSara in https://github.com/vllm-project/vllm-omni/pull/7219
* [CI][NPU] Onboard Qwen-Image perf to Ascend A3 nightly by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7970
* [CI/Build] Cut L3 entrypoint BAGEL cold starts via L4 small-model sleep suites (#6851) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/7359
* [CI/Build][NPU] Add  E2E test on Ascend A3 by @DarkMatter9527 in https://github.com/vllm-project/vllm-omni/pull/7775
* [Bugfix] Reject partially bound diffusion LoRA adapters by @provoke210 in https://github.com/vllm-project/vllm-omni/pull/8008
* [Bugfix] Validate Breeze PCM speech content and calibrate HNR guard by @linyueqian in https://github.com/vllm-project/vllm-omni/pull/8044
* [Frontend][Benchmark] Add judger and accuracy metrics for OmniInterac… by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/7317
* [Bugfix][Frontend] Return 500 for unexpected speech failures by @z2z23n0 in https://github.com/vllm-project/vllm-omni/pull/6487
* [Bugfix] Reclaim aborted chunk SHM and consumed connector state by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/8082
* [Bugfix] Fix MiniCPM-o 4.5 native duplex speech alignment by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/7974
* [BugFix][Cosmos3] Register opt-in T2I pipeline so deploy-config applies by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/7971
* [CI/Build] Stabilize MiniCPM-o mixed-input audio consistency test by @0z5a in https://github.com/vllm-project/vllm-omni/pull/7710
* [LingBot World] Speed up VAE decode, FP8 input quant, and default K/V gather by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7749
* [Model] Reduce MiniCPM-o 4.5 reference setup and Talker overhead by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/7968
* [Hotfix][Core] PrefixCache: re-plan same-step prefix hits whose rows are written after the early prefetch by @ZhengWG in https://github.com/vllm-project/vllm-omni/pull/7877
* [CI/Build] Initialize RoPE frequencies in MiniMax H3 TeaCache tests by @gcanlin in https://github.com/vllm-project/vllm-omni/pull/8125
* [Bugfix][Diffusion] Clean up inline worker lifecycle by @andyluo7 in https://github.com/vllm-project/vllm-omni/pull/7986
* [CI/Build][Bugfix] Cover corrupt chat audio data URLs on Qwen3-Omni by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/8108
* [CI/Build][NPU] Run HunyuanVideo-1.5 perf step via run_benchmark.py by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/8107
* [Bugfix] Read hidden layer key from model config instead of hardcoding 24 by @khairulkabir1661 in https://github.com/vllm-project/vllm-omni/pull/5825
* [Perf] Add reusable exact projection caching with MiniMax-H3 integration by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/7987
* [Orchestrator] Implement cache reset RPC forwarding to AR stage engine cores by @Asthenia0412 in https://github.com/vllm-project/vllm-omni/pull/6766
* [CI/Build] Fix Qwen3-Omni Talker benchmark workload by @psv666 in https://github.com/vllm-project/vllm-omni/pull/8057
* [Bugfix] Fix OmniRequestOutput iteration in the MOSS-TTS-Nano offline example by @devangpratap in https://github.com/vllm-project/vllm-omni/pull/6583
* [CI/Build] Make benchmark CLI test hermetic by @HaningZS in https://github.com/vllm-project/vllm-omni/pull/6337
* [Bugfix] Replace hardcoded Vivian speaker default with config-ordered default by @khairulkabir1661 in https://github.com/vllm-project/vllm-omni/pull/5814
* [Bugfix] Sync TeaCache skip decision across SP to prevent deadlock by @yzong-rh in https://github.com/vllm-project/vllm-omni/pull/6228
* [Bugfix][Cosmos3] Resolve per-component transformer quantization config by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/8101
* [Bugfix] Enable pipeline profiling for LingBot Video by @Bezdarnost in https://github.com/vllm-project/vllm-omni/pull/7744
* [Bugfix] Print text alongside audio in multimodal chat example by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/8130
* [Bugfix][Diffusion] Bound diffusion execute RPCs with the DP wave timeout by @ChoHee15 in https://github.com/vllm-project/vllm-omni/pull/7745
* [Doc] Announce PR limitation to avoid agent spam by @Isotr0py in https://github.com/vllm-project/vllm-omni/pull/8131
* [Bugfix][XPU] Enable LTX-2 family on XPU by @Joshna-Medisetty in https://github.com/vllm-project/vllm-omni/pull/7961
* [CI][Diffusion] Add tiny model builder for StableDiffusion3Pipeline by @khairulkabir1661 in https://github.com/vllm-project/vllm-omni/pull/6423
* [CI][Diffusion] Add acceleration test groups for SD3 and Flux2 by @khairulkabir1661 in https://github.com/vllm-project/vllm-omni/pull/6625
* [Bugfix][Frontend] Surface streaming terminal submit failures by @BANANASJIM in https://github.com/vllm-project/vllm-omni/pull/7011
* [BugFix] Add modalities and logprobs validation for /v1/chat/completions by @Shaun-Walsh in https://github.com/vllm-project/vllm-omni/pull/4743
* [Bugfix][Qwen3-Omni] Restore the talker sampling seed by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/7474
* [Bugfix][Qwen3-TTS] Decode full payload from model buffer by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/8132
* [Bugfix] Compensate cancelled duplex open, append and resume; fix serving README drift (#7636 Issues 1, 2, 3, 21) by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/7785
* [Model] Support request-level batching for MammothModa2 DiT on shared runtime by @BeatSeat in https://github.com/vllm-project/vllm-omni/pull/7476
* [Bugfix] Stop cancelled MiniMax-H3 requests at denoising boundaries by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/7844
* [BugFix][Cosmos3] Decode uploaded video paths in v2v preprocess by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/8106
* [Feature][Diffusion] Add BOOGU Image sequence parallel support by @Xenoryn in https://github.com/vllm-project/vllm-omni/pull/5718
* feat(magi2): add fused mHC post-processing by @yeahdongcn in https://github.com/vllm-project/vllm-omni/pull/7261
* [Bugfix][Diffusion] Skip AllGather DLO weight restore on worker shutdown by @specture724 in https://github.com/vllm-project/vllm-omni/pull/8103
* [CI/Build] Fix MiniCPM-o codec readback test under default-device dispatch by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/8147
* [Bugfix][MammothModa2] Keep AR layer 0 in BF16 in the FP8 KV preset by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/8083
* [Bugfix][MammothModa2] Stop allocating AR KV cache for the replaced language model by @MrlixiangWE in https://github.com/vllm-project/vllm-omni/pull/8095
* [Bugfix] Align Ming-Image prompt templates and RNG, fix number of layer to decompose by @yuanheng-zhao in https://github.com/vllm-project/vllm-omni/pull/8118
* [Bugfix] Restore MiniCPM duplex Seed-TTS startup by @NolenLiang in https://github.com/vllm-project/vllm-omni/pull/8152
* [Bugfix][MiniCPM-o] Build the Stage-0 duplex runtime at load time by @chickeyton in https://github.com/vllm-project/vllm-omni/pull/8146
* [Bugfix] Restore torch default device between tests by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/8157
* [Bugfix] Fix MRV2 payload slicing with CUDA Graph padding by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/8159
* [Bugfix] Fix Nemotron VoiceChat duplex nightly handshake by @Sy0307 in https://github.com/vllm-project/vllm-omni/pull/7884
* [Bugfix] Fix MiniCPM-o 4.5 seeded turns and window E2E inputs by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/8160

## New Contributors
* @xRay2016 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6701
* @LyxWxj made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6484
* @AuFlow made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6553
* @andyluo7 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6884
* @Arifuzzamanjoy made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6783
* @JW-L-7 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6571
* @Xenoryn made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5716
* @jingchengtian made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6908
* @BANANASJIM made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6957
* @0z5a made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6268
* @Anai-Guo made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6886
* @MciG-ggg made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6230
* @Sky-Trigger made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6359
* @heyuanliu-intel made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6989
* @MikeyDong1 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6983
* @LOGO127 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7065
* @MaxFreedomPollard made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7111
* @xiaobaixia222 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7136
* @lucasruan1618 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7189
* @zhang-keliang made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7041
* @FredHuangNV made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7027
* @KrystalRay made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6876
* @pujitha24 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7071
* @specture724 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7016
* @Jerry2423 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6961
* @tlysanhuo made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7130
* @liangmenghuang made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7083
* @ShuoleiWang made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5352
* @hmellor made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5381
* @twu3202 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7240
* @eval-dev made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5730
* @NancyFyong made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6157
* @ZenAlexa made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7003
* @XuTianle0101 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6896
* @Sparks-M made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7299
* @chi030303 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7289
* @kunkunblueberry made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7200
* @Shaun-Walsh made their first contribution in https://github.com/vllm-project/vllm-omni/pull/4741
* @CarrotSwordsman made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7007
* @yuweih205 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7155
* @bojiang-li made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7343
* @Hiro208 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7341
* @yancaocn made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6930
* @kuafou made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7430
* @junpengw67-max made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5128
* @gerayking made their first contribution in https://github.com/vllm-project/vllm-omni/pull/4876
* @Holworth made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6982
* @yuanwu2017 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6093
* @Srinivasoo7 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/4770
* @Oliver7th made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7324
* @LiquidGunay made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7445
* @JiahengX made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7436
* @NolenLiang made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7316
* @lzwnoname made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7084
* @nodeeeeee made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7477
* @yadongtan made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6639
* @vuuihc made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7463
* @Big2Wheel made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7192
* @Bezdarnost made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7108
* @xuexueligao made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7472
* @yzhautouskay made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6922
* @peterDengcx made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7520
* @GodHu777777 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7570
* @LinzeShi made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7483
* @zhuhu00 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7423
* @RyanYun09 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7331
* @THUqliu made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7454
* @chenchaoxu7575 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6950
* @Tiagosf00 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7059
* @howard-shan made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7573
* @De-cs made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7561
* @holykie made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7590
* @Levius-Fubuki made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7134
* @leegangtoe made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7298
* @chethanuk made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7525
* @xiaoyu-xyz made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7465
* @avraichur96 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7449
* @tsinghua-code made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7549
* @shihongzhi made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7291
* @armaanamatya made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7779
* @hoseung2 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6848
* @LiRunGuo made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7959
* @jeffaa729 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7909
* @liuyihua95 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7815
* @Sworol made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7756
* @cs-fisha made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7953
* @yanbao1217 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7869
* @sc-hua made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6974
* @Dong1017 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/8047
* @MichaelCao0 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7921
* @DarkMatter9527 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7775
* @provoke210 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/8008
* @z2z23n0 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6487
* @devangpratap made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6583

**Full Changelog**: https://github.com/vllm-project/vllm-omni/compare/v0.28.0...v0.30.0