vllm-project/vllm-omni v0.28.0rc1 on GitHub

Highlights

This release candidate features 212 merged changes, including contributions from 22 new contributors.

vLLM-Omni v0.28.0rc1 focuses on five major areas: 1) expanding multimodal model coverage across speech, music, image, video, and robotics, 2) advancing diffusion scheduling, continuous batching, paged KV cache, and Host Weight Runtime–based distributed layerwise offload, 3) delivering substantial MiniMax H3, TTS, and MiniCPM-o performance improvements, 4) strengthening realtime and full-duplex speech serving, and 5) improving runtime reliability, observability, CI coverage, and hardware portability. The release candidate also rebases the project onto vLLM 0.28.0 and removes additional legacy configuration paths.

Key Improvements

Significantly expanded model coverage with IndexTTS 2.5, dots.tts, Gepard-1.0, MiniMax Music 3, NVIDIA Nemotron VoiceChat, LTX-2.5, SANA-Video, SANA-WM, HiDream-O1-Image, and π0 VLA support. (#5957, #4765, #5666, #6186, #5842, #6089, #6070, #5508, #4061, #5194, #4222)
Advanced scalable diffusion execution with native scheduler-managed paged KV cache, continuous batching for MiniMax H3, request-level batching for Wan2.2, realtime AR-diffusion tick sessions, and a single-GPU UniProc executor. (#6094, #6102, #5810, #5676, #5491, #6308)
Introduced the Host Weight Runtime foundation and integrated it with DLO, including loader-owned host-weight plans, final-layout BF16 artifacts, mmap-backed direct H2D transfer, explicit post-load publication, online FP8 AllGather, and no-AllGather DLO execution. (#6213, #6419, #6427, #6445, #6591, #6279, #6486, #6651)
Improved speech and TTS serving with forced-alignment timestamps, speech token-usage headers, native full-duplex Nemotron VoiceChat, concurrent MiniCPM-o duplex sessions, realtime audio fixes, and multiple hot-path optimizations. (#4795, #4499, #6089, #6021, #6564, #5068, #5174, #5791)
Further optimized MiniMax H3 across accelerators and execution modes, adding global FP8 with DLO, Turbo LoRA, continuous batching, fused kernels, strict Ulysses boundary handling, reduced reference-video scanning, and new RTX PRO 5000 and NPU 950PR recipes. (#5910, #6476, #6550, #5810, #5990, #6281, #6283, #6173, #6064, #5857, #6120)

Core Architecture & Runtime

Rebased the project onto vLLM 0.28.0, updating engine, model-runner, scheduler, configuration, attention, and platform integrations for the vLLM 0.28 release line. (#6606)
Added the Host Weight Runtime, providing a shared lifecycle for preparing, publishing, mapping, and transferring host-resident model weights. (#6419, #6427)
Added loader-owned host-weight plans and final-layout BF16 artifacts for diffusion DLO deployments. (#6213, #6445)
Integrated no-AllGather DLO with Host Weight Runtime and registered mmap-backed host weights for direct H2D transfer. (#6486, #6591)
Added online FP8 support for DLO AllGather and established a FLUX.2-klein BF16 Host Weight Runtime contract. (#6279, #6651)
Added native diffusion KV-cache initialization and scheduler-managed block allocation, together with a native paged-KV backend for diffusion workers. (#6094, #6102)
Added realtime AR-diffusion tick sessions for LingBot World 2.0. (#5491)
Added a UniProc diffusion executor for single-GPU deployments, avoiding unnecessary distributed-process overhead. (#6308)
Added an opt-in event-driven engine orchestration loop. (#5221)
Refactored diffusion parallel state and continued the worker/model-runner correctness cleanup. (#5531, #5452)
Replaced hard-coded full-payload input-stage handling with resolved stage transport capabilities. (#6149)
Added pause, resume, sleep, and wake support for autoregressive stages in AsyncOmni, including admission recovery after wake. (#6084, #6581)
Fixed remote replica membership lifecycle races, replica device splitting when stages omit explicit devices, and per-stage runtime-environment restoration. (#5277, #5445, #6214)

Model Support

Added IndexTTS 2.5 support. (#5957)
Added dots.tts, a continuous-autoregressive 48 kHz TTS model from rednote-hilab. (#4765)
Added Gepard-1.0 native-autoregressive FSQ/NanoCodec TTS for offline inference. (#5666)
Added MiniMax Music 3 text-to-music generation. (#6186)
Added NVIDIA NemotronLabs VoiceChat 11B offline speech-to-speech inference and native full-duplex serving. (#5842, #6089)
Added LTX-2.5 and the LTX standard two-stage pipeline, with improved audio parity and checkpoint-shard validation. (#6070, #5500, #6342, #6234)
Added native SANA-Video 2B text-to-video and image-to-video generation. (#5508)
Added stage-one SANA-WM support. (#4061)
Added HiDream-O1-Image support. (#5194)
Added π0 VLA model support for vision-language-action workloads. (#4222)
Added offline SVDQuant W4A4 diffusion support and AutoRound MXFP4 quantized-model support. (#6162, #5544)
Added a FastVideo VSA attention backend for Wan2.2. (#4820)
Added native full-duplex and configurable concurrent-session support for MiniCPM-o. (#6021, #6619)

MiniMax H3 Productionization

Enabled global FP8 with DLO and added online-FP8 DLO AllGather support. (#5910, #6279)
Added MiniMax H3 diffusion continuous batching. (#5810)
Added Turbo LoRA support through both the legacy LoRA manager and DLO. (#6476, #6550)
Optimized the DLO component lifecycle and integrated no-AllGather execution with Host Weight Runtime. (#6526, #6486)
Added fused Q/K RMSNorm and RoPE, SwiGLU activation, and modulation with FP32 accumulation. (#5990, #6283, #6281)
Added NPU-fused RMSNorm, RoPE, SwiGLU, native GQA, and fused AddNorm paths for the Qwen3-VL encoder. (#5915, #6061, #6167, #6040)
Optimized strict Ulysses partition boundaries and avoided redundant reference-video scans. (#6173, #6064)
Fixed model-level CPU-offload residency and a distributed VAE hang when decoder tiles are fewer than ranks. (#6072, #6345)
Added optional planar video-response encoding. (#6288)
Added deployment recipes for RTX PRO 5000 and Ascend NPU 950PR, plus DLO DP2 smoke coverage. (#5857, #6120, #6555)

Audio, Speech & Omni Production Optimization

Added TTS timestamps through a forced-aligner pooling model. (#4795)
Added speech token-usage response headers and preserved detailed prompt-token accounting. (#4499, #5181)
Optimized GLM-TTS by moving loop-invariant embeddings, RoPE, masks, and classifier-free-guidance batches outside the Euler loop. (#5068)
Optimized the OmniVoice hot path with batched device-to-host transfers, mask caching, and cached text embeddings. (#5174)
Fused QKV and gate_up projections in the Qwen3-TTS code predictor. (#5791)
Enabled asynchronous scheduling for MOSS-TTS and aligned its default sample parameters with the official implementation. (#6241, #6156)
Added CUDA Graph acceleration for the MiniCPM-o 4.5 CFM DiT estimator and NPU Graph replay for Code2Wav. (#6082, #5604)
Restored MiniCPM-o streaming audio caching, fixed first-response stalls and mid-utterance truncation, and capped offline Talker generation to the remaining context window. (#6274, #6346, #6458)
Improved native-duplex MiniCPM-o configuration, concurrent-session handling, hub-cache metadata loading, and async-chunk snapshot management. (#6021, #6276, #6406, #6619)
Suppressed initial silence codec tokens for Qwen3-TTS. (#5048)
Optimized the CosyVoice3 TensorRT stream handoff and fixed sampling, stage handoff, and STFT device mismatches. (#5673, #6424, #6454)
Turned silent asynchronous TTS chunk failures into visible errors and reclaimed resumable requests after completion. (#6033, #6360)
Reduced exposure of user-provided TTS and audio inputs by moving those log messages from INFO to DEBUG. (#6329)

Diffusion, Image & Video Generation

Added scheduler-managed native KV-cache initialization, block allocation, and a paged-KV worker backend. (#6094, #6102)
Added request-level batching for Wan2.2 pipelines and continuous batching for MiniMax H3. (#5676, #5810)
Added native SANA-Video 2B T2V/I2V, LTX-2.5, HiDream-O1-Image, and stage-one SANA-WM support. (#5508, #6070, #5194, #4061)
Added offline SVDQuant W4A4 and online FP8 routing coverage for diffusion models. (#6162, #3027)
Added diffusion request metrics and returned them from the image-edit endpoint. (#4755, #5999)
Added text and image quality metrics. (#6150)
Fixed lost asynchronous diffusion outputs in request-level batching and made the async-output wait timeout configurable. (#6023, #6255)
Released GPU memory after diffusion execution failures and increased worker timeout tolerance for large broadcast payloads. (#6385, #4845)
Fixed Wan spatial-reshard boundaries, HunyuanImage3 accuracy validation, and Hunyuan VAE/patch-embedding performance. (#6062, #5981, #6306)
Fixed LongCat image-edit dimensions and TeaCache negative-branch CFG arguments. (#6222, #6181)
Fixed FLASH_ATTN cross-attention key-padding unpadding. (#5866)
Added a dedicated WORLD process group for distributed VAE communication. (#6401)
Prevented LTX-2.5 from silently loading incomplete, unindexed Diffusers checkpoints. (#6234)

Quantization, Attention & Memory Efficiency

Added online FP8 execution with DLO AllGather and MiniMax H3 global FP8 support. (#6279, #5910)
Added offline SVDQuant W4A4 support for diffusion pipelines. (#6162)
Added offline AutoRound MXFP4 quantized-model support. (#5544)
Fixed Qwen3-Omni AWQ quantization name mapping and stabilized compiled thinker MRoPE execution. (#5687, #6449)
Scoped Bagel FP8 configuration to the diffusion stage. (#6085)
Fixed compatibility between HSDP and the new online FP8 linear implementation. (#5677)
Added fused Q/K RMSNorm and RoPE, fused SwiGLU, FP32-accumulation modulation, and a mask-free TensorRT-LLM packed-padding path. (#5990, #6283, #6281, #6542)
Enabled fused SwiGLU on MUSA and avoided unsupported complex-RoPE alias guards. (#6364, #6110)
Added native GQA and fused AddNorm for the MiniMax H3 encoder on NPU. (#6040)
Upgraded Cache-DiT to 1.5.0 and Diffusers to 0.40.0. (#6065, #6459)

Serving, Frontend & API Behavior

Added native full-duplex serving for NVIDIA Nemotron VoiceChat. (#6089)
Added configurable concurrent MiniCPM-o duplex sessions and strengthened native-duplex session lifecycle handling. (#6021, #6318, #6619)
Added pause/resume and sleep/wake controls for autoregressive AsyncOmni stages. (#6084)
Added forced-alignment TTS timestamps and speech token-usage response headers. (#4795, #4499)
Added a local OmniInteract realtime benchmark. (#6522)
Relaxed TTS request validation for models without uploaded speakers. (#5878)
Fixed model-tag synchronization and command-line flag normalization in the API server. (#3805)
Preserved object-storage URIs during model resolution and respected media redirect policies for image references. (#5036, #6122)
Restored Hugging Face snapshot-path model detection, with the initial implementation subsequently reverted for further validation. (#6624, #6642)
Improved SenseNova configuration and model detection. (#5877)
Hardened request validation against denial-of-service overflow cases. (#6598)

Output, Configuration & Compatibility

Refactored OmniRequestOutput to inherit directly from vLLM RequestOutput, removing the nested request-output wrapper. (#5146)
Updated downstream users after removal of the legacy OmniRequestOutput.request_output accessor. (#6172)
Preserved ec_transfer_params, cache-creation-token counts, and prompt-token usage details on Omni outputs. (#6152, #5181)
Reused upstream vLLM configuration objects in VllmOmniConfig. (#6050)
Removed internal stage_configs_path plumbing and the legacy stage_args YAML loader. (#5741, #6200)
Removed stale stage-configuration references from documentation and examples. (#6270, #6347)
Moved sampling-parameter overrides and capability metadata from legacy dispatch logic into TTS adapters. (#5272, #6138)

Platforms, Distributed Execution & Hardware Coverage

Upgraded Ascend NPU support to v0.27.0. (#6096)
Added MiniMax H3 recipes for RTX PRO 5000 and NPU 950PR. (#5857, #6120)
Added NPU-native GQA, fused AddNorm, RoPE, SwiGLU, and RMSNorm optimizations for MiniMax H3. (#6040, #6061, #6167, #5915)
Added MiniCPM-o Code2Wav NPU Graph replay and strengthened NPU CI coverage. (#5604, #6275)
Added ROCm MI300X coverage for verified recipes and migrated the v0.27.x ROCm CI fleet to MI300X. (#6207, #5886)
Fixed ROCm CPU-test and circular-import failures. (#6267, #6287)
Fixed XPU CPU-offload device placement and made asynchronous AR and image-output paths device-agnostic. (#6125, #5569, #5571)
Added a dedicated WORLD group for distributed VAE communication and fixed replica-layout validation. (#6401, #5742)
Fixed Wan2.2-S2V complex indexing on NPU and NPU MoE registration. (#6320, #6350)

CI, Benchmarks & Documentation

Added per-model, per-entry-mode code-coverage collection and expanded dependency-based E2E test selection for shared model code. (#5593, #5994)
Added MiniCPM-o 4.5 performance coverage to the ready gate and strengthened its NPU accuracy jobs. (#6079, #6275)
Added tiny-model testing for Qwen-Image Edit and EditPlus and parallelized diffusion tiny-model tests. (#5656, #6339)
Added API-server surface guardrails and re-enabled previously skipped E2E and example tests. (#6202, #5641)
Added nightly Docker Hub publishing and cleanup to the release pipeline. (#6048)
Strengthened pre-commit checks with Markdown linting, SPDX validation, and repository policy hooks. (#6273)
Split scheduled L4/L5 pipelines, moved slower L1/E2E coverage to weekly runs, and demoted expensive nightly cases. (#6311, #5944)
Improved performance-test stability by matching warmups to benchmark concurrency and reusing judge and server processes. (#6356, #6208, #6613)
Added an environment-variable configuration reference and reconciled it with subsequent changes on the main branch. (#6217, #6631)
Reorganized user-guide feature and quantization taxonomies, refreshed community documentation, and split diffusion attention and CPU-offload guides. (#6045, #6074, #6141, #6075)
Unified recipe serving commands on vllm serve --omni. (#6221)
Added a dedicated vLLM-Omni simplification-review skill. (#6363)

Breaking Changes

The project now targets the vLLM 0.28 release line. Out-of-tree integrations, platform plugins, attention backends, and code relying on internal vLLM APIs should be validated against vLLM 0.28 before upgrading. (#6606)
OmniRequestOutput now directly inherits from RequestOutput. Code accessing the previous nested request_output property must migrate to the corresponding fields on OmniRequestOutput. (#5146, #6172)
The legacy stage_args YAML loader and remaining internal stage_configs_path plumbing have been removed. Deployments and integrations must use registry-backed deployment configuration and the unified configuration model. (#5741, #6200)
Sampling overrides and TTS capability metadata now live in model adapters. Out-of-tree TTS integrations relying on legacy dispatch tables may need to implement or update adapter metadata. (#5272, #6138)
Diffusers is now pinned to 0.40.0, Cache-DiT to 1.5.0, and the project is aligned with the updated NPU software stack. Custom environments should review extension and dependency compatibility before upgrading. (#6459, #6065, #6096)

Note

This is a pre-release version. It is intended for evaluation and integration testing before the final release; production deployments should validate model quality, performance, and platform compatibility in their target environments.
Host Weight Runtime, no-AllGather DLO, native diffusion paged KV cache, continuous batching, and the event-driven orchestration loop are newly introduced or rapidly evolving execution paths. Applications should validate memory residency, checkpoint compatibility, request cancellation, fault recovery, and long-running concurrency behavior before production adoption.
Realtime and full-duplex serving behavior varies by model. Nemotron VoiceChat and MiniCPM-o users should validate session lifecycle, interruption handling, context limits, audio buffering, and concurrency requirements.
Hardware capabilities remain backend-specific. Quantization formats, fused kernels, graph execution, attention implementations, CPU offload, and DLO configurations may require platform-specific dependencies and recipes.
#6557, which changed Wan DMD pipeline-test alignment, was reverted in #6574 and the revert was itself reverted in #6578. The final release state therefore retains the original #6557 change.
The Hugging Face snapshot-path model-detection change in #6624 was reverted by #6642 and is not included as an active behavior change in this release candidate.

What's Changed

[Docs] Add MiniMax-H3 recipe for RTX PRO 5000 by @lishunyang12 in #5857
docs: align User Guide feature taxonomy by @hsliuustc0106 in #6045
[New Model] Support IndexTTS 2.5 by @Sy0307 in #5957
[Misc] Build canonical image task prompts by @hsliuustc0106 in #6049
[Bugfix] Fix HunyuanImage3 accuracy test by @BLANKETusers in #5981
[CI][MiniCPM-o] Add MiniCPM-o 4.5 perf coverage to the ready gate by @amy-why-3459 in #6079
[Bugfix] Fix Wan spatial reshard boundary by @rahul-steiger-nv in #6062
docs: align quantization overview with navigation by @hsliuustc0106 in #6074
[CI/Build] Add examples policy to PR skills by @hsliuustc0106 in #6046
[Feature] timestamps for TTS via a forced-aligner pooling by @wjinxu in #4795
[Model] Support NVIDIA-NemotronLabs-VoiceChat-11B offline speech-to-speech by @yuekaizhang in #5842
[CI/Build] Add nightly DockerHub publish and cleanup to release pipeline by @yenuo26 in #6048
[Bugfix][MiniCPM-o] Align Daily-Omni offline loading and duplex soft-… by @amy-why-3459 in #6095
[Perf] MiniMax H3 Qwen3-VL support fused RMSNorm on NPU by @wjialish in #5915
[Refactor][OutputProcessor 2/3]: OmniRequestOutput should inherit RequestOutput and no nested wrap-up by @bowieshi in #5146
[Perf][TTS] glm_tts: hoist loop-invariant text embed, RoPE, mask and CFG batch out of Euler loop by @JuanPZuluaga in #5068
Add cosmos-guardrail dependency and update error message by @MaciejBalaNV in #6107
[Perf][TTS] omnivoice hot path: D2H batching, mask caching, cached text embeddings by @JuanPZuluaga in #5174
[Bugfix] Fix diffusion TTS adapter lookup in speech serving by @HaningZS in #6121
[Tests] Enable tiny model testing for Qwen-Image Edit and EditPlus by @NickCao in #5656
[CI/Build] Fix ReadTheDocs build under RTD's seeded pip 23.1 by @mjZhaoElaine in #6129
[Dependency] Upgrade Cache-DiT to 1.5.0 by @hsliuustc0106 in #6065
[Misc] Consolidate LingBot text-to-video runner by @hsliuustc0106 in #6076
[Perf][MiniCPM-O-4.5]Add CUDA Graph For CFM DiT estimator by @stringl1l1l1l in #6082
[Bugfix] Scope Bagel FP8 config to diffusion stage by @natureofnature in #6085
[Metrics] Add Diffusion Metrics by @vraiti in #4755
[BugFix] Use MOSS-TTS-Local official sample params by @gcanlin in #6156
[Doc] Refresh community documentation by @hsliuustc0106 in #6141
[NPU] Upgrade to v0.27.0 by @FayeSpica in #6096
[Refactor] Remove internal stage_configs_path plumbing by @zwhzzz0821 in #5741
[Feat] LTX Standard Two-Stage Pipeline by @mglyn in #5500
[Bugfix][Higgs-Audio-V3] Disable XQA decode by @Sy0307 in #6068
[Model][TTS] Add dots.tts (rednote-hilab): continuous-AR 48kHz TTS by @Moore-Z in #4765
[Kernel] Fuse Q/K RMSNorm and RoPE by @bobboli in #5990
[doc]: Update latest news section for verl-omni release by @SamitHuang in #6187
[Perf] Cap diffusion worker thread count before spawn by @ZJLi2013 in #6165
[CI] Per-model, per-entry-mode code coverage collection by @ShengleiFu in #5593
[ROCm] [CI] Migrate CI to mi300x for v0.27.x by @tjtanaa in #5886
[Bugfix] Fix HSDP compatibility with the new online FP8 linear method by @baonudesifeizhai in #5677
[Perf][Diffusion] Avoid redundant MiniMax-H3 reference video scans by @yeahdongcn in #6064
[Bugfix][MUSA][Qwen Image] Avoid complex RoPE alias guards by @yeahdongcn in #6110
[Perf][NPU] Fuse MiniMax H3 Qwen3-VL RoPE by @wjialish in #6061
[Perf][NPU] Fuse MiniMax H3 Qwen3-VL SwiGLU by @wjialish in #6167
[Feat/Bugfix] Fix API Server model tag <-> model sync & flag normalization by @alex-jw-brooks in #3805
[Bugfix][XPU][Tests]Scope XPU pytest to explicit paths to fix collection crash by @Joshna-Medisetty in #6175
[Frontend] Loosen Validation for TTS Models with No Uploaded Speakers by @alex-jw-brooks in #5878
[Bugfix] Stop reading the removed OmniRequestOutput.request_output accessor by @MrlixiangWE in #6172
[CI] Expand E2E source_file_dependencies for shared model code by @yenuo26 in #5994
[Model] Add Gepard-1.0 native-AR FSQ/NanoCodec TTS (offline inference) by @mjZhaoElaine in #5666
[Bugfix] Preserve object storage URIs during model resolution by @Ma1oneZhang in #5036
[Docs] Split diffusion attention and CPU offload guides by @hsliuustc0106 in #6075
[codex] Fix diffusion worker timeout for large broadcast payloads by @hotTeaFun in #4845
[Model][Performance] Optimize MiniMax-H3 strict Ulysses boundaries by @mo-ke-ke in #6173
[Bugfix] Carry ec_transfer_params and num_cache_creation_tokens on OmniRequestOutput by @MrlixiangWE in #6152
[BugFix][Nightly CI] Adjust Qwen-Image accuracy thresholds for stable FA-deterministic mode by @NumberWan in #5963
[Model] Add LTX-2.5 support by @lishunyang12 in #6070
[Bugfix] Fix lost async diffusion outputs in request-level batching by @SamitHuang in #6023
[CI] Refresh Voxtral-4B-TTS perf baselines from a clean window by @linyueqian in #6032
[Bugfix] Restore per-stage runtime env during launch by @m0g3r in #6214
[Diffusion][Quantization] Enable MiniMax-H3 global FP8 with DLO by @lishunyang12 in #5910
[Perf] Support request-level batching for Wan2.2 pipelines by @nagisa-kunhah in #5676
Add MiniMax-H3 recipe for NPU 950PR by @yiminghub2024 in #6120
[Bugfix] Prevent DiffusionResultPump crash on cancelled futures (#5793) by @anurag12-webster in #5983
[Docs] Unify recipe serve commands on vllm serve --omni by @linyueqian in #6221
[Core] Add realtime AR-Diffusion tick sessions for LingBot World 2.0 by @Jack47 in #5491
[Diffusion] Add loader-owned host-weight plans for DLO (TP=1) by @hsliuustc0106 in #6213
[Refactor] Remove legacy stage_args YAML loader by @zwhzzz0821 in #6200
[BugFix][TTS] R1.1-R1.4: turn silent async-chunk failures into visible ones by @linyueqian in #6033
[CI]demote slow nightly cases and trim Qwen-Image-Edit coverage by @yenuo26 in #5944
[Migrate] Move sampling parameter overrides from legacy dispatches to TTS adapters by @sphinxkkkbc in #5272
[Bugfix] Respect media redirect policy for image references by @HaningZS in #6122
[ROCm][Recipe] Add MI300X coverage for verified recipes by @akshatvishu in #6207
[CI/Build] Gate one-word pronunciation on a success rate, not per request by @ShengleiFu in #5681
Add π0 VLA model support by @yicwang in #4222
[ROCm] [CI] Fix cpu test on rocm by @tjtanaa in #6267
[MiniCPM-o][NPU] Enable MiniCPM-o Code2Wav NPUGraph replay by @Zhou248 in #5604
[Docs] Classify PD disaggregation as an experimental feature by @hsliuustc0106 in #6115
[CI/Build][MiniCPM-o] Pin Seed-TTS scorer to npu:1 in the NPU accuracy job by @psv666 in #6275
Add speech token usage headers by @JLiu4Coding in #4499
[Bugfix][NPU] Break the pytest DiffusionOutput circular import by @amy-why-3459 in #6293
[Bugfix][MiniCPM-o] Load Daily-Omni QA metadata from the hub cache by @psv666 in #6276
[Model] Add MiniMax Music 3 text-to-music by @linyueqian in #6186
[Diffusion] Add native KV cache initialization and Scheduler-managed block allocation by @zwhzzz0821 in #6094
[Bugfix][Test] fix Qwen3-omni OOM reliability and drop voxcpm2 stability/reliability test cases by @zhumingjue138 in #6299
[ROCm] [CI] Fix circular import for rocm due to patch by @tjtanaa in #6287
[Config] Reuse vLLM configs in VllmOmniConfig by @Acerak01-fy in #6050
[XPU][SDXL] Fix text encoder input device under CPU offload by @Joshna-Medisetty in #6125
[Bugfix] Avoid eager pi0 runtime import in pipeline registry by @hsliuustc0106 in #6322
[Core] Support online FP8 with DLO AllGather by @hsliuustc0106 in #6279
[BugFix][NPU] Wan2.2-S2V RoPE: replace complex64 advanced indexing with index_select by @FayeSpica in #6320
[Bugfix | Model] Fix SenseNova & Use Well-defined Model Configs by @alex-jw-brooks in #5877
[BugFix][TTS] Move user input from INFO to DEBUG in TTS/audio log lines by @dougbtv in #6329
[Perf] Fuse MiniMax H3 SwiGLU activation by @gcanlin in #6283
[Perf] Fuse MiniMax H3 modulation with FP32 accumulation by @gcanlin in #6281
[REFACT]Refactor diffusion parallel state by @bjf-frz in #5531
[Model] Add SANA-WM support(Stage-1 only) by @BruceLoveDecimal in #4061
[Ming] Cleanup Ming-family shared modules by @yuanheng-zhao in #6119
[Perf] Enable async schedule for MOSS-TTS by @Sy0307 in #6241
[Tests][Refactor] P0.1: Add API server surface guardrails by @herotai214 in #6202
[Bugfix][MiniCPM-o] Restore streaming audio cache and unstall the first duplex response by @BruceLoveDecimal in #6274
[skip ci][Doc] Remove obsolete DiffusionParallelConfig in many example scripts by @fhfuih in #6347
[Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #6369
[Bugfix][FLUX.2-klein] Fix duplicated image position ids by @Joshna-Medisetty in #6130
[Bugfix] Record device-agnostic torch.Event on XPU to fix async D2H image corruption [XPU] by @tthakkal in #5571
[Docs] Remove stale stage config references by @zwhzzz0821 in #6270
[XPU][Bugfix] Make Omni AR async output path device-agnostic by @Joshna-Medisetty in #5569
[Perf][CI] Reuse the Whisper judge worker across a test module by @ShengleiFu in #6208
[Bugfix][MiniCPM-o] Stop Talker mid-utterance truncation and load cod… by @amy-why-3459 in #6346
[BugFix][Diffusion] Fix MiniMax-H3 VAE hang when decoder tiles are fewer than ranks by @linzhenpl07 in #6345
[MUSA] Fix SwiGLU to use fused op by @yeahdongcn in #6364
[CI/Build] Match perf warmups to benchmark concurrency by @Sy0307 in #6356
[Bugfix] Reclaim resumable async-chunk requests on finish by @natureofnature in #6360
[Bugfix] Use dedicated WORLD group for distributed VAE communication by @david6666666 in #6401
[Bugfix][MiniCPM-o][NPU] Skip Code2Wav dynamo unwrap when flow.encode… by @amy-why-3459 in #6397
[Refactor] Worker/ModelRunner runner correctness fixes (G2/N) by @tzhouam in #5452
[CI/Build] Strengthen pre-commit with markdownlint, SPDX, and policy hooks by @yenuo26 in #6273
[Bugfix] Return diffusion metrics for image-edit endpoint by @kTorp in #5999
[Model] Suppress silence codec tokens for the first N Qwen3-TTS decod… by @IneshReddy249 in #5048
[Perf][Hunyuan] Optimize for vae and patch_embed by @Bounty-hunter in #6306
Fix npu moe registration by @BLANKETusers in #6350
[Bugfix][Core] Fix replica device split when a stage omits devices by @yashkgp in #5445
[Tests] Run diffusion tiny model tests in parallel by @NickCao in #6339
Cosmos3 transfer fix by @MaciejBalaNV in #5614
[CI/Build] Update perf baselines base on 8/1-8/7 7 days avg by @congw729 in #6201
Cosmos3 logging and prompt improvements by @MaciejBalaNV in #6325
[Perf][Qwen3-TTS] Fuse QKV and gate_up projections in code predictor by @l-wave in #5791
[Bugfix] Preserve prompt token usage details by @ieaves in #5181
[Core] Add Host Weight Runtime foundation by @hsliuustc0106 in #6419
[Model] Add HiDream-O1-Image support by @yixiaoer in #5194
[Core] Add explicit post-load Host Weight Runtime publication by @hsliuustc0106 in #6427
[Tests] Re-enable previously skipped e2e/example tests by @yenuo26 in #5641
[Skills] Add vLLM-Omni simplification review skill by @princepride in #6363
[Docs] Restore the Star History chart by @congw729 in #6446
[Bugfix][MiniCPM-o] Fix async-chunk snapshot replacement and prompt cleanup by @natureofnature in #6406
[Bugfix][Diffusion] Fix MiniMax-H3 model-level CPU offload residency by @yeahdongcn in #6072
[BugFix]: CosyVoice3 STFT window device mismatch by @princepride in #6454
[Bugfix] Drop librosa reintroduced by Step-Audio2 by @NickCao in #6467
[Bugfix] Fix LongCat TeaCache CFG negative-branch guidance kwargs by @yzong-rh in #6181
[Model][Frontend] MiniMax-H3: Add opt-in planar video response encoding by @MosCloud in #6288
[Misc] Bump diffusers pin to 0.40.0 by @NickCao in #6459
[Bugfix][MiniCPM-o] Cap offline Talker generation at remaining context by @amy-why-3459 in #6458
[Refactor] Move model-specific capability metadata into TTS adapters by @sphinxkkkbc in #6138
[Diffusion] Add final-layout BF16 Host Weight Runtime artifacts by @hsliuustc0106 in #6445
fix: keep DLO hooks outside regional compilation by @hsliuustc0106 in #6073
[Diffusion] Add native paged KV backend for diffusion workers by @Acerak01-fy in #6102
[Core] Add UniProc diffusion executor for single-GPU by @rahul-steiger-nv in #6308
[Bugfix] Make the async-output wait bound configurable and default it higher by @ivanusto in #6255
[Bugfix] Preserve pipeline sampling constraints by @maithilijoshi20 in #6182
[Bugfix] Release GPU memory after diffusion execution failures by @nagisa-kunhah in #6385
[Bugfix][Qwen3-Omni] Stabilize thinker MRoPE compilation by @tzhouam in #6449
[BugFix] Eliminate distributed test port race by switching to file:// rendezvous by @NickCao in #6468
[CI]Remove SoulX-Singer support by @yenuo26 in #6362
[CI/Build] Move L1/E2E coverage to weekly and split scheduled L4/L5 pipelines by @yenuo26 in #6311
[tts][bugfix] Fix CosyVoice3 sampling and stage handoff by @akshatvishu in #6424
[CI] Remove AudioX and MammothModa2 support by @yenuo26 in #6353
[diffusion][test] Cover FLUX.2-dev online FP8 routing by @Songrui625 in #3027
[Bugfix] Fix Qwen3-Omni AWQ quantization name mapping by @lishunyang12 in #5687
[Feature] Support MiniMax-H3 Turbo LoRA with the legacy manager by @mglyn in #6476
[FEAT]Add FastVideo VSA backend for Wan2.2 by @bjf-frz in #4820
[Feature]: Support pause / resume and sleep / wake for AR stages in AsyncOmni by @knlnguyen1802 in #6084
[Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #6535
[CI]Fix Wan DMD pipeline test alignment by @bjf-frz in #6557
[Bugfix] Prevent LTX-2.5 from silently loading unindexed Diffusers shards by @mglyn in #6234
[Feature][MiniMax-H3] Support diffusion continuous batching by @princepride in #5810
Revert "[CI]Fix Wan DMD pipeline test alignment" by @Gaohan123 in #6574
Revert "Revert "[CI]Fix Wan DMD pipeline test alignment"" by @hsliuustc0106 in #6578
[CI] Resume AR admission after sleep-mode wake by @hsliuustc0106 in #6581
[Diffusion] Integrate no-AllGather DLO with Host Weight Runtime by @hsliuustc0106 in #6486
[Test] Cleanup for helpers by @yenuo26 in #6523
[Model] Add native SANA-Video 2B T2V and I2V support by @liuyao0322 in #5508
[Bugfix] Invalidate stale ref_audio cache on local file modification (#4873) by @pranavthakur0-0 in #5670
[Bugfix] Honor height/width overrides in LongCatImageEditPipeline by @yzong-rh in #6222
[Feature][MiniCPM-o] Support configurable concurrent duplex sessions by @natureofnature in #6021
[Bugfix] Fix remote replica membership lifecycle races by @hbhflw2000 in #5277
[minor, fix] Remove cuda sync on wake_up in AsyncOmni by @knlnguyen1802 in #4092
[Refactor][2/N Scheduler]Replace _FULL_PAYLOAD_INPUT_STAGES with a resolved stage transport capability. by @wy17003 in #6149
[perf] Add mask-free TRTLLM packed-padding path by @bobboli in #6542
[Test] add marker for stability and add Wan2.2 npu test case by @zhumingjue138 in #6343
[BugFix] Qwen-image performance regressed - Avoid mapping diffusion_batch_size onto scheduler max_num_seqs by @NumberWan in #6525
[CI] Fix response_format json_schema error expectation by @clumsylad21 in #6290
[Misc] move text-to-audio online examples to a unified folder by @zzehli in #4807
[CI/Build] Clean up Qwen3-Omni nightly tests by @psv666 in #6570
[AutoRound] Add offline quantized MXFP4 model support by @jl9876 in #5544
[CI] Reduce Qwen-Image Function and share step-execution perf server by @NumberWan in #6613
[Model] Optimize MiniMax-H3 DLO component lifecycle by @hsliuustc0106 in #6526
[Diffusion] Register HWR mmap for direct DLO H2D by @hsliuustc0106 in #6591
[Benchmark] Add local OmniInteract realtime benchmark by @natureofnature in #6522
[Bugfix] Handle DoS Overflow Cases by @alex-jw-brooks in #6598
[CI/Build] Pin GGUF plugin for diffusion nightly tests by @wxwxwwxxx in #6303
[Doc] Add environment variables configuration reference by @hsliuustc0106 in #6217
[Bugfix] Reconcile env-var inventory with post-#6217 main drift by @hsliuustc0106 in #6631
[Perf][NPU] Enable native GQA & fused AddNorm for MiniMax-H3 encoder by @MarkPoloChina in #6040
[CI]Remove DreamID-Omni and MagiHuman support by @yenuo26 in #6357
text and image metric by @AbelSara in #6150
[Bugfix][MiniCPM-o] Serve native duplex from shipping YAMLs and fence… by @amy-why-3459 in #6619
[Model] Optimize CosyVoice3 TensorRT stream handoff by @EchoHayate in #5673
[Bugfix] Restore name-based model detection for HF cache snapshot paths by @linyueqian in #6624
Revert "[Bugfix] Restore name-based model detection for HF cache snapshot paths" by @Gaohan123 in #6642
[Feature] Support MiniMax-H3 Turbo LoRA with DLO by @mglyn in #6550
[Doc] Add MiMo-Audio recipe for RTX 5090/5090D 32GB by @smartDream-chao in #6559
[Test] Add MiniMax-H3 DLO DP2 T2VA smoke by @david6666666 in #6555
[Bugfix] Fix missing Realtime audio completion by @psv666 in #6564
[Model][Bugfix] Improve LTX audio parity and similarity guards by @mglyn in #6342
[Perf][Engine] Event-driven orchestration loop (opt-in) — S1 of #4855 by @linyueqian in #5221
[Recipe] zai-org/GLM-TTS (2x non-standard RTX 4090 48GB) by @harley619 in #5769
[Bugfix] Reject instructions changes on locked native duplex sessions by @anurag12-webster in #6318
[Bugfix][Diffusion] Fix FLASH_ATTN cross-attention key-padding unpad by @cr-gao in #5866
[Bugfix] Validate per-stage device layout before spawning workers (#5003) by @ChoHee15 in #5742
[Model] Add native full-duplex Nemotron VoiceChat serving by @Sy0307 in #6089
[diffusion][feature] Add offline SVDQuant W4A4 support by @lishunyang12 in #6162
[Diffusion] Add FLUX.2-klein BF16 HWR contract by @hsliuustc0106 in #6651
[Rebase] Rebase to vllm 0.28.0 by @tzhouam in #6606

New Contributors

@wjialish made their first contribution in #5915
@HaningZS made their first contribution in #6121
@stringl1l1l1l made their first contribution in #6082
@Moore-Z made their first contribution in #4765
@MrlixiangWE made their first contribution in #6172
@hotTeaFun made their first contribution in #4845
@mo-ke-ke made their first contribution in #6173
@m0g3r made their first contribution in #6214
@anurag12-webster made their first contribution in #5983
@yicwang made their first contribution in #4222
@yzong-rh made their first contribution in #6181
@MosCloud made their first contribution in #6288
@ivanusto made their first contribution in #6255
@pranavthakur0-0 made their first contribution in #5670
@wy17003 made their first contribution in #6149
@clumsylad21 made their first contribution in #6290
@jl9876 made their first contribution in #5544
@wxwxwwxxx made their first contribution in #6303
@MarkPoloChina made their first contribution in #6040
@EchoHayate made their first contribution in #5673
@smartDream-chao made their first contribution in #6559
@harley619 made their first contribution in #5769

Full Changelog: v0.27.0rc1...v0.28.0rc1