Highlights
This release includes 776 merged pull requests from 185 contributors, including 61 first-time contributors. vLLM-Omni v0.34.0 rebases onto vLLM 0.34.0 and moves autoregressive and generation stages to Model Runner V2. It adds MiniMax-H3-Image, FLUX.3 (Klein), PAN2, Hy-Nomas-Diffusers, Haru-4B, and Ming-3.0 speech models, alongside distributed duplex serving and realtime video streaming.
Key Improvements
- Scale duplex serving across nodes. Deploy stage replicas across independent runtimes while keeping each session pinned to its replica group. External orchestrators can also run stages independently and select replicas through the experimental
/v1/runAPI. (#10422, #6883, #9729) - Stream video during request execution. CausVid can deliver decoded chunks while generation is in progress. Noisy Pipeline Parallelism can batch compatible work across requests while preserving each session’s KV state. (#9926, #10049, #10122)
Model Support
- Image generation and editing. Added MiniMax-H3-Image for text-to-image generation and multi-reference editing through the Images API, and FLUX.3 (Klein) with a native BF16 pipeline. (#10120, #9523)
- Video generation. Added PAN2 for text-to-video and image-to-video, and Hy-Nomas-Diffusers for camera-controlled autoregressive video streaming. Kandinsky 6 gains Lite and Pro Distil checkpoints with integrated HD and Full HD super-resolution. (#8602, #9338, #9027)
- Speech and audio. Added Haru-4B for full-duplex spoken dialogue, Ming-3.0-TTS Tiny for streaming speech and voice cloning, Ming-3.0 Audio Flash for speech recognition and streamed spoken responses, and Stable Audio 3 Medium for stereo text-to-audio generation. (#10301, #9119, #9130, #3973)
- Segmentation and depth. Added SAM 3 and SAM 3.1 for promptable image segmentation, SAM 3 for video segmentation and tracking, and Video Depth Anything for video depth estimation. (#9543, #9579, #9714)
- Robotics and driving. Added NVIDIA Alpamayo 2 Super and native Runway Praxis-1 policies for DROID and YAM. Restored π0.5 support and enabled multi-stage policy serving over OpenPI. (#9133, #10197, #10204, #9080, #9120)
Core Architecture & Runtime
- Reduced Omni prefix-cache overhead under concurrent requests and added LMCache CPU offloading for autoregressive KV cache and the hidden states downstream speech stages need. (#9162, #9193, #1330)
- Added CUDA IPC weight sharing so compatible, colocated HunyuanImage3 AR and DiT stages can share a single copy of the backbone weights. (#9333, #9761, #9850)
Audio, Speech & Realtime Serving
- Expanded duplex tool-call controls, added semantic voice activity detection to turn-based Realtime, and improved
/v1/realtimeprotocol compatibility. (#9909, #9170, #9265) - Added experimental Mooncake KV and hidden-state reuse for supported Qwen3-Omni turn-based and duplex deployments. Enabled multi-frame CUDA graph replay for the MiniCPM-o 4.5 Talker. (#10107, #10321, #10288)
Diffusion, Image & Video Generation
- Expanded MiniMax-H3 with more FastH3 checkpoints, ControlNet Union serving, and video continuation from cached latents across requests. (#9167, #7479, #9342)
- Added experimental Physis Same-Step pipelines for Wan, LTX, and MiniMax-H3. Supported diffusion pipelines can also run both classifier-free guidance branches in one denoiser call on a single rank. (#9154, #9797)
Quantization & Memory Efficiency
- Added ConvRot INT8 and INT4 quantization for diffusion transformers on CUDA, and support for the MiniMax-H3 AutoRound checkpoint with an MXFP8 DiT and MXFP4 text encoder. (#9343, #9363)
Serving, Frontend & API Behavior
- Added OpenTelemetry inference tracing and opt-in AudioSeal watermarking for supported audio generation. (#8836, #7366)
Platforms & Hardware Coverage
- Updated the Ascend stack to vLLM-Ascend
0.34.0, added Model Runner V2 deployment profiles for supported ROCm workloads, and enabled Intel XPU sleep mode and diffusion graph capture. (#10773, #10819, #10820, #5485, #9306)
Breaking Changes
- Model Runner V2 is required. Deployments that resolve any stage to
model_runner: v1fail at startup. Review custom deployment files and platform overrides before upgrading. Affected profiles include full Drama 3 and Fish-Omni speech serving, PersonaPlex duplex, VibeVoice, LFM3-Omni, and the Qwen3-TTS and Qwen3-Omni MUSA overrides. (#10365, Deployment compatibility) - Removed first-party support for JoyAI-Image-Edit and dropped the
--shm-threshold-bytesCLI option. (#10393, #9992)
What's Changed
- [Bugfix] Keep coordinator broadcasts scheduled during publish by @amy-why-3459 in #9172
- [Bugfix][Benchmark][Duplex] Fall back to frame-sample judging when the video exceeds the judge context by @kunkunblueberry in #9101
- [Model] Run MiniCPM-o 4.5 speaker embeddings on CUDA by @natureofnature in #9178
- [Bugfix] Import Mooncake connectors lazily so an optional dependency cannot break model inspection by @y-null in #9168
- [New Model] Ming-3-TTS (inclusionAI/Ming-3.0-tts-tiny) Support by @yuanheng-zhao in #9119
- [Doc] Host model recipes in recipes/models/ by @esmeetu in #9184
- [Model] Restore Pi0.5 (lerobot/pi05_base) by @yenuo26 in #9080
- [Doc] Add FastH3 deployment guidance to MiniMax-H3 recipe by @mooncake2026 in #9204
- [Model] Add FastH3 series by @lishunyang12 in #9167
- [Bugfix] Optionally bind fa3-fwd dense FlashAttention on Hopper by @NumberWan in #9201
- [CI/Build] Gate MiniCPM-o Seed-TTS SIM and isolate UTMOS on the eval GPU by @amy-why-3459 in #9160
- [BugFix][Duplex][Benchmark] Add soundtrack mix for Omni-DuplexEval sessions by @amy-why-3459 in #9175
- [CI/Build] Record branch in accuracy JSON and enable GHA on release/v* by @yenuo26 in #9211
- [Bugfix] Retain duplex request IDs until cleanup succeeds by @aayc123 in #7708
- [CI/Build][ROCm] Add HunyuanVideo-1.5 accuracy to MI300 nightly by @haic0 in #9125
- [CI/Build][ROCm] Add HunyuanImage3 LoRA accuracy to MI300 nightly by @haic0 in #9127
- [New Model] Ming3-Audio (inclusionAI/Ming-3.0-audio-flash) Support by @yuanheng-zhao in #9130
- [Bugfix][ROCm] Skip Qwen-Image fused QK-norm assertions on HIP by @haic0 in #9229
- [Bugfix] Support on-demand offload for Cosmos3 UND embeddings by @shijiecao-01 in #8741
- [Model] Optimize MiniCPM-o handoff and optional Code2Wav execution by @amy-why-3459 in #9218
- [Core] Support KV Cache + Hidden State CPU Offloading by @Vivo50E in #1330
- [Bugfix] Preserve queued duplex input reservations across clear by @wuli666 in #7783
- [Bugfix][Diffusion] Make Cache-DiT effective when serving Wan2.2 T2V/I2V/S2V without num_inference_steps by @xuechendi in #7335
- [Bugfix][XPU] Honor full_attn_spans in the XPU FlashAttention path by @iboiko-habana in #9123
- [Experimental] Clarify Noisy KV step_replicas, D, and clean semantics by @ShengDev in #9169
- [2/N][Perf][Core] Cut the per-step cost of the omni prefix cache by @ZhengWG in #9193
- [Kernel][Z-Image] Fuse QK RMSNorm + RoPE with CUDA-matching numerics by @yuweih205 in #7597
- [Model] Add CPU-reference TensorRT speaker extraction for MiniCPM-o by @amy-why-3459 in #9188
- [BugFix][CI][Qwen-Image] Pin accuracy to TORCH_SDPA, eager, and NPU by @NumberWan in #8798
- [Bugfix] Guard Qwen3-Omni Code2Wav dtype on Blackwell with cuDNN 9.20 by @psv666 in #9144
- [Model] Optimize MiniMax-H3 reference KV caching by @zhang-s-a in #8072
- [Frontend] Serve multi-stage policies over OpenPI by @KrzysztofTomala in #9120
- [Kernel][HunyuanVideo-1.5] Fuse video/text QK RMSNorm + RoPE + cat into one Triton launch by @yuweih205 in #7596
- [1/N][Perf][Core] Stop the omni prefix cache from collapsing under concurrent hits by @ZhengWG in #9162
- Enable Cosmos3 MAPS TP, SP, CFG and HSDP by @wkutak in #8633
- [CI/Build][ROCm] Shard READY model-executor GPU coverage across two MI300s by @haic0 in #9227
- [CI/Build][ROCm] Hit the AITER MHA cache from ABI-compatible main by @haic0 in #9224
- [XPU][Feat] enable sleep mode support on intel GPU by @yma11 in #5485
- [CI/Build][ROCm] Add Wan2.2 I2V accuracy to MI300 nightly by @haic0 in #9126
- [Bugfix][Hardware][Intel] Breeze-TTS: replay depth and text graphs eagerly off CUDA by @Joshna-Medisetty in #8652
- [Bugfix][Wan2.2] Cache-DiT: support low-noise-only (single transformer_2) runs and discover cache targets by @xuechendi in #8800
- [Quantization][Hardware][Intel] Wan2.2 S2V: plumb quant_config through the transformer by @xuechendi in #8902
- [Bugfix] Cache-DiT: no separate-CFG step pairing under CFG parallel by @xuechendi in #8998
- [Hardware][Intel] Wan VAE: one distributed-encode piece per rank instead of overlapping tile by @xuechendi in #8966
- [Hardware][Intel] Wan VAE: run wan_vae_fastpath on XPU by @xuechendi in #8968
- [CI/Build][ROCm] Add Qwen-Image diffusers accuracy to MI300 nightly by @haic0 in #9124
- [CI/Build][ROCm] Add MiniMax-H3 accuracy to MI300 nightly by @haic0 in #9128
- [Bugfix][Hardware][Intel] MiMo-Audio: allocate local speech tokens on the embedding device by @Joshna-Medisetty in #8654
- [Model] Add JoyAI duplex tool delegation by @kunkunblueberry in #9149
- [Bugfix][Diffusion] Stamp the CFG branch for TeaCache instead of inferring it from parity by @fusheng-ji in #8483
- [Model] PersonaPlex: stream Code2Wav output as deltas by @linyueqian in #9113
- [Bugfix][Qwen-Image-2.1] Drop idle decode-graph prefix copies by @NumberWan in #9272
- [Bugfix] MiniMax-H3: make audio VAE output independent of request order by @Shan2L in #8802
- [Frontend] Add remote SeedVR2 restoration to ComfyUI by @WEIXILIN0307 in #8505
- [Hardware][Ascend] Add a Drama 3 NPU runner path by @amy-why-3459 in #9106
- [OOT] [vllm-metal] Add external CustomOp registration [1/3] by @WindChimeRan in #8943
- [Bugfix] Make Qwen3-Omni Seed-TTS prompts read questions verbatim by @amy-why-3459 in #9284
- [Bugfix][Diffusion][Qwen-Image-2.1] Fall back to SDPA for dense 4D masks by @wangqiankun13 in #9248
- [Bugfix][Core] Ignore completed duplex append tasks in response gating by @amy-why-3459 in #9282
- [Bugfix] MiniMax-H3: keep native layout for local snapshots with modular_model_index.json by @Shan2L in #8801
- [Bugfix] Recognize mixed-attention warmup requests in MRv2 by @FayeSpica in #9264
- [perf][npu] Remove host sync stalls in the prompt feature setup by @y-null in #9186
- [CI/Build] Enable qwen3-omni accuracy benches on NPU A3 by @psv666 in #9091
- [Bugfix][Rlease 0.32.0] Register native KV specs in fast allocation tests by @yenuo26 in #9289
- [Bugfix][NPU] Support piecewise block-causal attention for full_attn_spans by @YawningPants in #9244
- [BugFix] security: restrict untrusted image decoding formats by @NickCao in #9114
- [Bugfix] Align Wan VAE attention queries for 1x1 tiles by @amy-why-3459 in #9171
- [Frontend] Wire semantic VAD into legacy Realtime by @NickCao in #9170
- [Bugfix] Honor the per-request speaker for Qwen2.5-Omni audio output by @advanced233 in #9266
- [Bugfix] Lock shared memory before opening segments by @WindChimeRan in #9285
- [Bugfix][Duplex] Give the first session open a cold-start timeout budget by @kunkunblueberry in #9107
- [Bugfix] Re-aim a dropped silence continuation at the answer that opened by @armaanamatya in #7880
- [Feat] Add Cosmos3-Nano Pipeline Parallel support by @MrDongsls in #4637
- [3/N][Config] Enable Thinker prefix caching by default for Qwen3-Omni and MiniCPM-o by @ZhengWG in #9203
- fix: guard omni-specific CLI args behind use_omni check by @khairulkabir1661 in #8698
- Add /v1/realtime WebSocket conformance tests (RHAISTRAT-2749) by @khairulkabir1661 in #8258
- [Bugfix][XPU] Fix HunyuanImage3 FLASH_ATTN regression from #9248 by @iboiko-habana in #9302
- [Bugfix][ROCm] Compare Wan SDPA realignment with the math kernel by @haic0 in #9303
- [Bugfix][CI/Build] Place stage-CLI test replicas on separate GPUs by @cipheraxat in #9041
- [Bugfix] Read Cosmos3 Edge PP ownership from the live modules by @haic0 in #9304
- [Feature][XPU] Add SPARGE INT8 sparse attention backend by @yiliu30 in #9254
- [CI/Build][ROCm] Add MiniCPM-o 4.5 accuracy to MI300 nightly by @haic0 in #9267
- [Bugfix] Place the MAGI-2 preview pipeline on the current accelerator by @Joshna-Medisetty in #8655
- [Bugfix] Expect Qwen3-Omni Thinker prefix cache only on CUDA by @hacZ17 in #9312
- [BugFix][OmniVoice] Add ASR for missing reference text and match audio preprocessing by @akshatvishu in #5784
- [CI/Build][ROCm] Add Qwen-Image 2.1 accuracy to MI300 nightly by @haic0 in #9268
- [Kernel] Split the fused GroupNorm+SiLU reduction across CTAs by @yashkgp in #6662
- [Core] Fail recompute preemption on AR stages that cannot replay their decode state by @amy-why-3459 in #9249
- [diffusion][test][ROCm] Make Hunyuan GroupNorm regression detect missing guard by @andyluo7 in #9276
- [Frontend] ComfyUI: add MiniMax-H3 supplied-audio music video workflow (H3-03) by @nikiwang-125 in #9226
- [Bugfix] MiniMax-H3: warn on TaylorSeer order>=2 and document the Ref2VA artifact margin by @Dong1017 in #9270
- [Bugfix] Keep diffusion LoRA buffers resident on the compute device by @he-yufeng in #8874
- [Frontend] MiniMax-H3: Add timeline guides to the ComfyUI extension (GUIDE-02) by @shykoe in #9152
- [Bugfix][Qwen-Image-2.1] Snap the trailing VAE tile so tiled decode leaves no seam by @ZhIMQ in #8868
- [CI/Build][ROCm] Select qualified ROCm tests in AMD CI by @hacZ17 in #9320
- [CI/Build][ROCm] Add Qwen3-Omni Seed-TTS realtime perf to MI300 nightly by @haic0 in #9307
- [Bugfix] Fix MiniMax-H3 ComfyUI workflow schemas by @lm10-arg in #9337
- [BugFix][TTS] Support long-form OmniVoice generation by @akshatvishu in #6409
- [CI/Build] Stabilize host runtime waiter completion wait by @lm10-arg in #9339
- [CI/Build] Fix AMD R2 collection and Qwen graph oracle by @lm10-arg in #9340
- [Bugfix][Hardware][AMD] Initialize Mammoth attention test weights by @lm10-arg in #9341
- [Feature] Quantization: Add ConvRot rotated INT8 and INT4 diffusion methods by @princepride in #9343
- [Doc] Tongyi/Z-Image-Turbo by @MrDongsls in #3283
- [Docs] Clarify MiniCPM-o attention backend comparisons by @amy-why-3459 in #9344
- [CI/Build][ROCm] Add MiniCPM-o Omni-DuplexEval to MI300 nightly by @haic0 in #9269
- [Model] Reduce CosyVoice3 V1 RAS sampler synchronization by @suyanli220 in #8730
- [Core][Metrics] Expose stage post-warmup memory snapshots by @akshatvishu in #5180
- [Benchmark] Accept multiple reference voices in CosyVoice3 concurrency benchmark by @Arifuzzamanjoy in #8456
- [Bugfix] Validate diffusion speech sampling parameters by @Liuchenbing-2026 in #8685
- [Bugfix] Validate served model names in diffusion speech by @Liuchenbing-2026 in #8882
- [CI][ROCm] Add MI325 tiny-diffusion multi-GPU nightly coverage by @lm10-arg in #9369
- [CI/Build] Filter AMD PR nightlies by source changes by @lm10-arg in #9372
- [Bugfix][CI/Build] Add MiniCPM-o Stage-2 timeout diagnostics by @lm10-arg in #9370
- [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #9375
- [Perf] Reduce MiniMax-H3 MP4 metadata and AAC encoding overhead by @xdlsy in #9074
- [Bugfix][AURA] Keep sentence TTS turns alive across vision and segments by @NumberWan in #8829
- [Bugfix][CI/Build] Fix MiniCPM-o watchdog CI regressions by @lm10-arg in #9374
- [Bugfix] Fix Qwen2.5-Omni talker end conditioning by @lm10-arg in #9371
- [Bugfix] Stabilize diffusion master_port precedence tests by @ShengDev in #9219
- [Bugfix] Fix Qwen2.5-Omni speaker test fixture by @lm10-arg in #9389
- [Bugfix] fix flashinfer backend and torch.compile(fullgraph=True) incompatible by @MrDongsls in #4989
- [Doc] MiniMax-H3: document the Ref2VA dialogue prompt contract by @Dong1017 in #9331
- [Bugfix][Diffusion] Reject unsupported pre-tokenized prompt ids at engine admission by @fusheng-ji in #8448
- [Bugfix] Fix grouped chunk interaction latency by @maithilijoshi20 in #8810
- [Bugfix] Reject invalid Omni-DuplexEval judge scores by @psv666 in #9390
- [CI/Build] Avoid busy polling in stage memory lifecycle tests by @amy-why-3459 in #9406
- [CI/Build] Update MiniCPM-o A3 C8 TTFT baseline by @amy-why-3459 in #9388
- [Refactor][3/N Scheduler]Move model metadata interpretation out of scheduler internals into a typed adapter contract. by @wy17003 in #6161
- [Model] Support fishaudio/drama-3-int4 hybrid INT4 checkpoint on CUDA SM80+ by @FayeSpica in #9391
- [Perf][Ascend] Replay Drama 3 decode steps as FULL graphs by @amy-why-3459 in #9308
- [Hardware][Ascend] Enable NPU reference encoding for MOSS-TTS Local by @Wallbreazzz in #9067
- [Bugfix] Reuse one Qwen3-Omni fault payload across concurrent requests by @yenuo26 in #9407
- [Hardware][NPU] Support SeedVR2 video restoration on Ascend NPU by @zengchuang-hw in #8737
- [CI/Build] Install fa3-fwd 0.0.3 with the dev extra by @NumberWan in #9393
- [Bugfix] Keep the offline Omni engine alive after a non-fatal request error by @fusheng-ji in #8450
- [Bugfix] Reject oversized stage budgets on integrated GPUs by @tzhouam in #9404
- [Tests] Add tiny model test for StableDiffusionXLPipeline by @NickCao in #6169
- [Bugfix] Preserve the diffusion parallel knobs when engine args come from parsed CLI args by @Bevisy in #8041
- [Bugfix][Benchmark] Match official Omni-DuplexEval GT reference handling by @RyanYun09 in #9056
- [Bugfix] Align MiniCPM-o reference audio with Transformers while retaining vision SDPA by @amy-why-3459 in #9386
- [Bugfix] Stop vocoder stages from declaring SupportsPP without make_empty_intermediate_tensors by @rk9595 in #6860
- [Bugfix][Benchmark] Enable audio_ttfp goodput SLOs on the current bench CLI by @ZacheryAU in #9322
- [CI/Build] Re-enable turn-based server VAD realtime test by @NickCao in #9300
- [Bugfix][Engine] Preserve non-parallel diffusion engine args across from_cli_args by @he-yufeng in #8585
- [Perf] Exclude wakeup_scope UUIDs from Torch Compile Hash Key by @alex-jw-brooks in #8899
- [CI/Build] Restore SenseNova-U1 smoke and H100 accuracy coverage by @Qing7364 in #8759
- [CI/Build][ROCm] Add Qwen3-Omni scale and performance burn-in by @lm10-arg in #9427
- [Bugfix][XPU] Support auto-round-lib >= 0.15 in SageAttention backend by @tzielinski-habana in #9255
- [Model] Add opt-in HF-numerics mode for the Qwen3-Omni thinker by @cr-gao in #8186
- [Model] Skip Fish Speech DAC reconstruction during reference encoding by @ischencheng in #8140
- [Bugfix][Frontend] fix: park generate() through partial wakes by @knlnguyen1901 in #8891
- [CI/Build] Keep AMD nightly dependency check independent of job count by @amy-why-3459 in #9444
- [Bugfix][ROCm] Make Qwen OOM recovery pressure deterministic by @hacZ17 in #9425
- [Bugfix] Qwen2.5-Omni: assemble talker spans from runner phase metadata by @LJX-xixi in #5019
- [Perf][Core] Materialize duplex Realtime response history at read boundaries by @he-yufeng in #8691
- [Refactor] Retire the generic deferred MTP path in MRV2 by @sphinxkkkbc in #9346
- [Diffusion] Chunked DLO weight transport with H2D/AllGather overlap (stage 1) by @leonail1 in #6374
- [CI/Build][ROCm] Update Qwen3-Omni AMD collection count by @lm10-arg in #9448
- [Bugfix][Kernel] Fix MoT GEMM rejecting a one-sided text/vae bias by @yashkgp in #6490
- [Frontend] Report generated audio duration for /v1/audio/speech by @apapi in #8875
- [Bugfix] Avoid retaining tensors in warning_once exception logging by @princepride in #9445
- [CI/Build][ROCm] Shorten READY R2-01 and AITER consumer jobs by @haic0 in #9439
- [Bugfix] Release diffusion ZMQ resources on loop initialization errors by @princepride in #8419
- [Bugfix][Core] Schedule native MRv2 async chunks atomically by @dongteng in #9060
- [Perf][Qwen2.5-Omni] Run the talker with CUDA graphs by default by @rynwang in #9447
- [CI/Build][ROCm] Add Qwen-Image backend similarity to MI300 nightly by @haic0 in #9434
- [Bugfix][VoxCPM2] Reject outer full CUDA graph modes at startup by @linyueqian in #9458
- [Bugfix][Diffusion] Prepare the MoE LoRA context by @BLANKETusers in #9443
- [Model] Add bounded same-stage summary memory for JoyAI duplex by @natureofnature in #8932
- [Bugfix] CosyVoice3: reject unsupported generation APIs and guard prefill embeddings by @ssfdgsg in #4738
- [Hardware][Ascend] Support MAGI-2 Preview on Ascend NPU by @jiangmengyu18 in #9006
- [Hardware][NPU] Add HunyuanVideo-1.5 accuracy coverage by @zengchuang-hw in #9174
- [Bugfix][NPU] Avoid Helios block-noise Cholesky fault on Ascend by @scestola-h in #7643
- [CI/Build] Trim Omni-DuplexEval artifacts and clean temporary outputs by @natureofnature in #9464
- [Bugfix][Profiler] Warn VLLM_TORCH_PROFILER_DIR was removed and is ignored by @ZacheryAU in #9473
- [CI/Build] Enable validated MiniMax-H3 A3 CI coverage by @Dong1017 in #9030
- [Bugfix] Avoid retaining tensors in Quack FP8 fallback logging by @rahul-steiger-nv in #9262
- [CI/Build][ROCm] Add MiniCPM-o 4.5 duplex function to MI300 nightly by @haic0 in #9431
- [CI/Build][ROCm] Add Qwen-Image expansion to MI300 nightly by @haic0 in #9432
- [Bugfix][NPU] Make Helios block-noise fallback LAPACK-independent by @lm10-arg in #9488
- [Bugfix] Re-bound Wan VAE channels-last bf16 decoder test tolerance by @Dong1017 in #9421
- [CI/Build][ROCm] Add Wan2.2 function tests to MI300 nightly by @haic0 in #9433
- [CI/Build][ROCm] Isolate Qwen3-Omni SHM leak coverage on MI300 by @haic0 in #9497
- [CI/Build][ROCm] Add Qwen3-Omni online core_model to MI300 READY by @haic0 in #9496
- [Core] Add Watermarking Framework & Support Watermarked Audio Generation w/ AudioSeal by @alex-jw-brooks in #7366
- [Hardware][Ascend] Fold HiFT weight norms when loading Token2Wav by @amy-why-3459 in #9463
- [CI/Build][ROCm] Add Qwen3-Omni and BAGEL full_model MI300 nightlies by @haic0 in #9435
- [CI/Build][ROCm] Run Hunyuan reliability without expert parallel by @hacZ17 in #9428
- [CI/Build][ROCm] Add MiMo and Ming full-weight Omni E2E to MI300 weekly by @hacZ17 in #9318
- [CI/Build][ROCm] Add Voxtral TTS perf to MI300 nightly by @hacZ17 in #9309
- [CI/Build][ROCm] Add Qwen3-TTS perf to MI300 nightly by @hacZ17 in #9310
- [CI/Build][ROCm] Add MiniCPM-o 4.5 single-GPU perf to MI300 nightly by @hacZ17 in #9429
- [Bugfix] Keep a deferred duplex commit on cancel; stop the client-VAD replay teardown truncating its last answer by @amy-why-3459 in #9486
- [CI/Build] Drop DreamZero upstream parity and keep the OpenPI wire contract by @yenuo26 in #9469
- [Bugfix] Honor stream=true in pure-diffusion chat completions by @princepride in #9452
- [Bugfix] Restore Lance video generation on current main by @habibrosyad in #4304
- [Bugfix] Preserve Cosmos3 transfer control guidance offset by @rahul-steiger-nv in #9263
- [Bugfix] Reduce Cosmos3 and Wan VAE peak memory by @rahul-steiger-nv in #9424
- [Bugfix] Preserve Omni-DuplexEval video frames and input tails by @psv666 in #9417
- [Experimental] Add first-chunk clean sink to Same-Step inference by @ShengDev in #9482
- [Model] Optimize Ming3 Audio Flash concurrent serving by @Sy0307 in #9196
- [CI/Build][ROCm] Select the marked Qwen2.5-Omni cases on MI300 MERGE by @haic0 in #9546
- [Bugfix] Disable Breeze depth autotune on XPU. by @Joshna-Medisetty in #9547
- [Hardware][Ascend] Support VoxCPM2 on the NPU Model Runner V2 by @DarkMatter9527 in #9507
- [Bugfix][Hardware][Ascend] Fix RainFusion mix video spans by @De-cs in #9241
- [Model] Optimize Ming-3.0-TTS Tiny serving on vLLM 0.32.0 by @Sy0307 in #9198
- [Doc] Update README and installation docs for v0.32.0 by @hsliuustc0106 in #9539
- [Welcome] Add committers and CODEOWNERS: timzsu, haic0, ZhengWG, FayeSpica by @hsliuustc0106 in #9564
- [CI/Build] Trim Qwen3-Omni artifacts and remove duplex Seed-TTS perf jobs by @psv666 in #9531
- [Perf][Qwen3-Omni] Capture the thinker audio encoder with CUDA graphs (part of #7452) by @MrlixiangWE in #9010
- [Bugfix][CI] Raise HunyuanImage-3 online I2I stage init timeout by @BLANKETusers in #9471
- [Model] Add SAM 3 / SAM 3.1 promptable image segmentation by @princepride in #9543
- [Core] Reduce Omni output and idle generation overhead by @amy-why-3459 in #9526
- [Bugfix] Keep Drama 3 on its NPU model state under Model Runner V2 by @amy-why-3459 in #9560
- [CI/Build] Remove MiniCPM-o Seed-TTS BF16 attention-cache test by @amy-why-3459 in #9181
- [Model][MiniMax-H3] Support optional text encoder CPU offload to unlock Cache-DiT by @mayooot in #8093
- [Bugfix] Order NIXL transfers after the stream writes they depend on by @KrzysztofTomala in #9422
- [CI/Build][ROCm] Add MiniCPM-o 4.5 offline, online, and duplex READY coverage by @andyluo7 in #7689
- [Bugfix][Diffusion] Reject empty prompt and num_inference_steps<=0 for FLUX.2-dev by @ShengDev in #9580
- [Bugfix][ROCm] Expect VoxCPM2 v1 audio defaults off CUDA by @hacZ17 in #9561
- [Model] SAM 3: video concept segmentation and tracking by @princepride in #9579
- [Bugfix] Restore MiniCPM-o CUDA duplex long-video recognition by @natureofnature in #9544
- [Bugfix] Warm prefill-sized fused-MoE Triton configs at startup by @IneshReddy249 in #9499
- [Model][Frontend] Add MiniMax-H3 Fun ControlNet Union serving (CTRL-01) by @zhuhu00 in #7479
- [CI/Build][ROCm] Fail-close the Wan2.2 MI300 READY node by @haic0 in #9548
- [CI/Build][ROCm] Select the marked async-engine abort case on MI300 READY by @haic0 in #9549
- [Model] Replace deprecated transformers config validators in Qwen3-TTS and VoxCPM2 by @moezdil in #9596
- [Bugfix] Keep streaming video session history text-only by @armaanamatya in #9568
- [CI/Build][ROCm] Select marked Qwen3-TTS CustomVoice core cases on MI300 READY by @haic0 in #9550
- [CI/Build][ROCm] Add one-card worker GPU cases to MI300 READY by @haic0 in #9552
- [AutoRound][Model][MiniMax-H3] support MXFP8-DiT and MXFP4-TE quantized checkpoint loading by @yiliu30 in #9363
- [Bugfix][NPU] Fix 310P detection after vllm-ascend moved is_310p by @wzk-jazk in #9246
- [OOT] [vllm-metal] Optimize eager code prediction with KV reuse [3/3] by @WindChimeRan in #8949
- [Model] Qwen3-Omni: send one merged prefill tensor to the Talker by @ischencheng in #9183
- [CI/Build][ROCm] Add Drama 3 core text-to-audio to MI300 READY by @haic0 in #9551
- [CI/Build][ROCm] Add diffusion device-reduction GPU cases to MI300 READY by @hacZ17 in #9606
- [Hardware][Ascend] Enable same-shape cross-reference CFM batching for MiniCPM-o 4.5 by @yezhi0312 in #9534
- [OOT] [vllm-metal] Add shared-memory storage hooks [2/3] by @WindChimeRan in #8946
- [CI/Build][ROCm] Add remaining one-card worker copy cases to MI300 READY by @haic0 in #9610
- [CI/Build][ROCm] Add two-rank diffusion pause-keep to MI300 READY by @hacZ17 in #9607
- [Bugfix] Take an odd median for the Qwen-Image diffusers latency gate by @NumberWan in #9520
- [CI/Build][ROCm] Select the marked VeRL-Omni generate case on MI300 READY by @hacZ17 in #9605
- [CI/Build] Remove MiniCPM-o duplex Seed-TTS performance test by @amy-why-3459 in #9635
- [CI/Build] Drop merge Simple Test and move CPU coverage to nightly by @yenuo26 in #9298
- [Core][Duplex] Measure server-side busy RTF per turn (RFC #8325) by @chickeyton in #9182
- [CI/Build][ROCm] Add one-card tiny diffusion base cases to MI300 READY by @hacZ17 in #9604
- [Model] MiniMax-H3: Continue a previous video from cached latents across requests by @princepride in #9342
- [Bugfix] Preserve MRv2 codec output without hidden states for Qwen3-TTS by @sphinxkkkbc in #9626
- [Hardware][Ascend] Bump MindIE-SD to 3.2.0 by @FayeSpica in #9573
- [Bugfix] Complete the duplex turn on a terminal listen by @Zssan12 in #8661
- [Frontend] TTS prefix-cache salts: default and caller-supplied by @RishabhSaini in #8298
- [Bugfix][ROCm] Disable mbtree in the chunked MP4 byte compare by @haic0 in #9658
- [Feature / Core] Support Independent Stage Execution & External Routing by @RishabhSaini in #6883
- [CI/Build][ROCm] Add distributed buffer GPU smokes to MI300 READY by @haic0 in #9659
- [CI/Build][ROCm] Make diffusion device-reduction READY coverage blocking by @haic0 in #9661
- [Model] Add optional LiteVAE decoder for MiniMax-H3 by @xdlsy in #9245
- [Bugfix] Address MAGI-2 NPU review follow-ups by @jiangmengyu18 in #9484
- [CI/Build][ROCm] Make one-card worker READY coverage blocking by @haic0 in #9660
- [CI/Build][ROCm] Promote MiniCPM-o core READY coverage by @haic0 in #9662
- [Doc] Add MiniMax-H3 LiteVAE checkpoint links by @xdlsy in #9674
- [CI/Build] Move Cosmos3 functional tests to weekly by @yenuo26 in #9685
- [perf][VoxCPM2]: async stop-mask D2H transfer with pinned memory by @Lingwei1943 in #8310
- [Hardware][Ascend] Reuse captured ACL graphs on static-shape decode misses by @tanhaoan333 in #9618
- [perf][VoxCPM2] remove weight_norm and cast AudioVAE to bf16 by @Lingwei1943 in #8309
- [CI/Build][ROCm] Require the Qwen3-Omni SHM leak MERGE check by @haic0 in #9678
- [CI/Build][ROCm] Gate the Qwen-Image default smoke on MI300 MERGE by @haic0 in #9681
- [CI/Build][ROCm] Gate MiniCPM-o core cases on MI300 MERGE by @haic0 in #9680
- [Model] Add Kokoro-82M single-stage TTS by @Alex-Wengg in #9637
- [CI/Build][ROCm] Gate Drama 3 text-to-audio on MI300 MERGE by @haic0 in #9679
- [CI/Build][ROCm] Make VeRL-Omni generate a blocking READY check by @hacZ17 in #9668
- [Bugfix] Show Seed-TTS WER/SIM/UTMOS in the TTS benchmark summaries by @MohammadHijjawi97 in #8354
- [CI/Build][ROCm] Qualify MiniCPM-o V1 fallback duplex on MI300 READY by @hacZ17 in #9670
- [Bugfix] Coalesce concurrent CosyVoice3 reference conditioning by @ychu824 in #9632
- [Perf] Share Qwen3-TTS streaming decode backend across single-stage and two-stage serving by @sphinxkkkbc in #9023
- [Model] Qwen3-TTS: cut code_predictor cold-start compile time ~4.5x by @RishabhSaini in #9652
- [Bugfix][Ming-TTS] Fix speaker-only mode handling in Ming TTS by @yuanheng-zhao in #9657
- [Core][Frontend] Add OpenTelemetry tracing for Omni inference by @z2z23n0 in #8836
- [CI/Build] Cut CPU coverage job runtime and keep it under the step timeout by @yenuo26 in #9392
- [Core] Add persistent CUDA IPC weight-sharing backend by @AbelSara in #9333
- [Bugfix][MiniCPM-o] Capture Code2Wav graphs right after load, not on the first request by @y-null in #9207
- [Core] Propose draft tokens on MRv2 AR stages; DSpark on the MiniCPM-o Thinker by @ischencheng in #9696
- [Bugfix] MiniCPM-o turn handoff: pick Thinker rows by position by @ischencheng in #9695
- [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #9716
- [Hardware][Ascend] Capture CosyVoice3 Flow and HiFT with NPUGraph by @tanhaoan333 in #9619
- [Bugfix] Align MOSS Local reference handling and duration conditioning by @Sy0307 in #9557
- [Core] Default Qwen3-Omni CUDA serving to model runner v2 by @amy-why-3459 in #9578
- [Kernel] Add Drama 3 sampler CUDA graph replay coverage by @shunwendongan in #9052
- [Perf][Core] Let AR stages wake after level-2 sleep for RL full-weight sync by @SamitHuang in #9712
- [Model] MiniMax-H3: serve Open-H3-Regenerate-2K by @princepride in #9562
- [Bugfix] Warm MoE prefill and seeded top-k kernels on the MRv2 AR runner by @amy-why-3459 in #9750
- [Doc] Spell out diffusion LoRA scale semantics by @Dong1017 in #9513
- [Bugfix][DLO] Accept shape-changing loader transforms in rank-local mmap by @zengchuang-hw in #9563
- [Misc] Modernize legacy typing: Optional -> X | None, drop collections.abc.Set by @moezdil in #9603
- [Bugfix] Forward the requested layer count on multi-stage image generation by @alex7092 in #9746
- [Diffusion] Share GraphedForward across platforms and add XPUGraphedForward by @iboiko-habana in #9597
- [Model] Avoid redundant CUDA mask discovery in MOSS-TTS Delay sampling by @dongbo910220 in #9694
- [cherry-pick][main] cherry-pick #9515: [Bugfix][Diffusion] Detach the idle MoE LoRA context on Ascend by @BLANKETusers in #9690
- [Bugfix][MammothModa2] Refresh t2i AR-grid constraints under FULL graph replay by @BLANKETusers in #9527
- [CI/Build] Remove redundant unit tests by @amy-why-3459 in #9792
- [Benchmark][Omni-DuplexEval] Survive a stalled close, a frozen reply, and a scoreless judge reply by @y-null in #9260
- [Model] Add Video Depth Anything video depth estimation by @princepride in #9714
- [Bugfix] Fix MiniCPM-o 4.5 duplex KV reanchor on Ascend NPU by @yezhi0312 in #9649
- [Model] Make MiniMax-H3 reference KV host layers contiguous by @zhang-s-a in #9259
- [Docs][Boogu] Mark Ascend NPU support for Boogu-Image-0.1-Turbo by @zengchuang-hw in #8250
- [Doc] Route Ascend diffusion optimization to MindIE-SD skills by @Fishermanykx in #9008
- [CI] Deflake recv drain burst test by measuring polls per chunk by @amy-why-3459 in #9756
- [CI/Build][ROCm] TIME-01: Add CPU-suite timing and execution artifacts by @haic0 in #9779
- [CI] Give the Qwen2.5-Omni CI thinker enough KV on L4 by @amy-why-3459 in #9755
- [CI/Build][ROCm] Guard READY entrypoint execution before blocking by @haic0 in #9774
- [Tools] Serve SheetSage2 transcription over HTTP by @princepride in #9823
- [New Model] Stable Audio 3 (medium) — text-to-audio diffusion (#3787) by @wuli666 in #3973
- [CI/Build][ROCm] Run MiniCPM offline and online MERGE checks at advanced level by @haic0 in #9775
- [Hardware][Ascend] Add NPU codec for persistent IPC weight sharing by @Bounty-hunter in #9757
- [Misc] Replace use of deprecated pkg_resources with importlib.metadata by @NickCao in #5874
- [Hardware][Ascend] Fix AR input embedding buffer width by @panpan0000 in #9810
- [CI/Build][ROCm] Validate Drama 3 text-to-audio at advanced MERGE level by @hacZ17 in #9776
- [CI/Build] Fix CPU coverage collection with multiple workers by @LinzeShi in #9785
- [CI/Build][ROCm] Add Qwen-Image batch coverage to READY and MERGE by @hacZ17 in #9780
- [CI/Build][ROCm] Restore CustomVoice async-chunk MERGE validation by @hacZ17 in #9781
- [CI/Build][ROCm] TIME-02: Add opt-in model-executor CPU shard rebalancing by @haic0 in #9782
- [Experimental][Diffusion] Bind Noisy KV cell path to session history by @ShengDev in #9758
- [CI/Build] Poll for joyai visual-memory observations instead of racing the settle budget by @yenuo26 in #9636
- [CI/Build] Add MiniMax-H3 Ref2VA RainFusion NPU nightly coverage by @fan2956 in #6060
- [Model] Add experimental Physis Same-Step Wan, LTX and MiniMax-H3 by @Dong1017 in #9154
- [Duplex][Benchmark] Add
client_rtf,busy_rtfandstall_ratiointo result files of benchmarks by @ZacheryAU in #9673 - [Hardware][Ascend] Capture MOSS Local reference encoding with ACLGraph by @Wallbreazzz in #9271
- [Model] TaoMate-H3: full-request inference and streamed audio delivery by @Dong1017 in #8676
- [Core] Align vLLM-Omni with vLLM 0.33.0 by @tzhouam in #9648
- [Bugfix][MammothModa2] Route AR MoE experts graph-safely under cudagraph capture by @BLANKETusers in #9764
- [Bugfix] Release MRv2 prefix-cache contexts on non-output TP ranks by @psv666 in #9861
- [Perf][MiniCPM-o] Default token2wav CFM timesteps to 5 by @y-null in #9872
- [CI/Build] Fix prefix-cache Ray tests for vLLM 0.33 by @amy-why-3459 in #9879
- [Doc] Add omni model integration and performance optimization skills by @amy-why-3459 in #8972
- [Doc] Add Qwen-Image-2.1 Ascend 910B1 low-latency serving recipe by @wangqiankun13 in #9796
- [Misc] Remove deprecated vllm_omni.engine output shim files by @moezdil in #9602
- [Bugfix][CI/Build] Enable layerwise offload for the LTX-2.3 distilled I2V accuracy test by @ShengDev in #9868
- [Bugfix] Avoid event ID collisions in scheduled video controls by @Qing7364 in #9857
- [Bugfix][ROCm] Stabilize eager chunked-video byte comparison by @haic0 in #9853
- [Bugfix][MiniMax-H3] Fix missing Regenerate control fields by @RuixiangMa in #9778
- [CI/Build] Fix vLLM 0.33 AMD regressions by @lm10-arg in #9883
- [CI][Qwen3-Omni-Thinker] Update markers to match pipeline definition by @clodaghwalsh17 in #9772
- [Bugfix] Correct TTS throughput improvement direction by @Mengmj in #9099
- [CI/Build][ROCm] Run Qwen2.5-Omni expansion on weekly by @haic0 in #9840
- [Bugfix] Avoid duplicate MRv2 full-payload serialization by @psv666 in #9815
- [CI/Build] Share infra retry and allow retrying passed GPU and NPU jobs by @yenuo26 in #9870
- [Model] Add Alpamayo 2 Super by @KrzysztofTomala in #9133
- [Quantization][SVDQuant]: Add MXFP4 SVD Support by @yiliu30 in #9379
- [CI/Build][ROCm] Add the ROCm-marked Bagel weekly node by @haic0 in #9843
- [CI/Build][ROCm] Add Moss TTS expansion to weekly by @haic0 in #9841
- [Model] Add Hy-Nomas-Diffusers generation and bounded AR streaming by @wtz2333 in #9338
- [Bugfix][Benchmark] Keep warmups from seeding the prefix cache by @ZhengWG in #9485
- [Benchmark][Omni-DuplexEval] Send the drain padding and bound the session writes by @y-null in #9802
- [Bugfix] Save streaming video with standalone client dependencies by @Qing7364 in #9902
- [CI/Build] Use a scoped Buildkite secret for Docker Hub releases by @khluu in #9924
- [NPU] Adapt to vLLM v0.33.0 and new vllm-ascend model_runner path by @DarkMatter9527 in #9890
- [CI/Build][ROCm] Move Stable Audio layerwise offload to nightly by @haic0 in #9842
- [Model] TaoMate-H3: reduce serial request preparation and decode overhead by @Dong1017 in #8919
- [Bugfix] Avoid runtime imports during ROCm model discovery by @lm10-arg in #9932
- [Core] Add engine-side LoRA merging for diffusion and AR rollouts by @knlnguyen1901 in #9766
- [Feature] CUDA K-step multi-frame Talker decode with graph replay by @y-null in #9733
- [XPU][Docker] Drop uv pip check from the XPU image build by @Joshna-Medisetty in #9948
- [Bugfix][CI] Isolate the Code2Wav stats test clock by @haic0 in #9944
- [Bugfix][CI] Use file rendezvous in Sana Video CPU tests by @haic0 in #9914
- [Core][Duplex] Cut API-process CPU and GIL contention under high concurrency by @lrq619 in #9705
- [Bugfix] Handle list-valued scales during distilled LoRA initialization by @LinzeShi in #9836
- [Core] Add HunyuanImage3 AR↔DiT weight-sharing adapter by @wangqiankun13 in #9761
- [Bugfix] Do not import Physis pipelines from package init by @ShengDev in #9949
- [Bugfix][ROCm] Bound Qwen3-Omni cold NFS startup timeouts by @hacZ17 in #9830
- [Perf][NPU] Enable ragged CFM batching for MiniCPM-o 4.5 by @yezhi0312 in #9880
- [CI/Build][ROCm] Add advanced entrypoint sleep/wake MERGE coverage by @hacZ17 in #9897
- [Model] Stream JoyAI duplex speech through async-chunk Code2Wav by @wuli666 in #9798
- [Bugfix] Count streaming video frames incrementally off the event loop by @Qing7364 in #9838
- [CI] Remove obsolete performance test steps from weekly pipeline by @yenuo26 in #9933
- [CI/Build] Share Qwen3-Omni merge runners to cut cold starts by @yenuo26 in #9871
- [Enhancement] Persist diffusion accuracy results with hardware-nested baselines by @yenuo26 in #9732
- [Bugfix] Preserve Qwen3-Omni text cached during chunked prefill by @LJX-xixi in #9492
- [Model] Improve MiniCPM-o concurrency and default CUDA to MRv2 by @amy-why-3459 in #9583
- [Bugfix] Ask for a mandatory final response in the MiniCPM-o duplex recognition driver by @y-null in #9816
- [Bugfix][NPU] Keep MOSS codec fusion attention NPUGraph-capturable by @Dong1017 in #9885
- [CI/Build] Bound AMD image publication stalls by @lm10-arg in #9956
- [Doc] Pass OmniDiffusionSamplingParams in generate() examples by @spped2000 in #9954
- [Frontend] Further improve spec compliance of the turn-based /v1/realtime endpoint by @NickCao in #9265
- [Test][OmniVoice] Opt voice-cloning tests into ASR escalation to fix CI flake by @pujitha24 in #7747
- [CI/Build][ROCm] Add Drama 3 voice-clone MERGE coverage by @hacZ17 in #9900
- [Bugfix][XPU] Stop overriding get_memory_info; scope VLLM_XPU_USE_SAMPLER_KERNEL=0 to Qwen2.5-Omni by @xuechendi in #9794
- [Hardware][Intel] Enable Model Runner V2 for omni AR/generation stages on XPU by @xuechendi in #9613
- [Perf] Diffusion attention: cache the vLLM flash-attn varlen lookup by @xuechendi in #9669
- [CI/Build][ROCm] Add diffusion collective parity nightlies by @hacZ17 in #9896
- [Bugfix] Align MiniCPM-o K-step codec validity across execution paths by @amy-why-3459 in #9976
- [Core] Support explicit replica routing for independent stages by @Yuy1L1 in #9729
- [Bug] Remove the dead --shm-threshold-bytes knob from CLI, examples and docs by @Bevisy in #9992
- [CI/Build][ROCm] Add CustomVoice prefill-tail MERGE regression by @hacZ17 in #9901
- [CI/Build][ROCm] Add MiniCPM-o full-model modality nightlies by @hacZ17 in #9858
- [Bugfix] Stabilize duplex VAD batch failure test by @lm10-arg in #10002
- [CI/Build][ROCm] Add Bagel shared-memory MERGE coverage by @hacZ17 in #9898
- [CI/Build][ROCm] Add Wan2.2 Diffusers-backend similarity nightly by @hacZ17 in #9860
- [CI/Build][ROCm] Add advanced MiniCPM duplex and window MERGE coverage by @hacZ17 in #9899
- [CI/Build][ROCm] Add Qwen3-TTS Base expansion to AMD nightly by @hacZ17 in #9859
- [CI/Build][Diffusion] Run MiniMax-H3 official-reference accuracy with deterministic Inductor by @haic0 in #10004
- [CI/Build][ROCm] Check the AMD R2-01 collection count on CUDA READY by @haic0 in #10005
- [CI/Build][ROCm] Add Qwen-Image nightly performance coverage by @hacZ17 in #9895
- [CI/Build][ROCm] Stabilize CustomVoice prefill-tail waveform guard by @lm10-arg in #10018
- [Model] MiniCPM-o: keep the Thinker prefix cache with a stage-0 draft on turn MRv2 by @ischencheng in #9864
- [Core][Diffusion] Add per-step attention schedules for MiniMax-H3 by @zhongc1211 in #8714
- [CI/Build][ROCm] Stabilize Bagel img2img pixel guard by @lm10-arg in #10027
- [Bugfix][CI/Build][ROCm] Split MiniCPM-o accuracy jobs by @lm10-arg in #10030
- [Bugfix][CI/Build][ROCm] Stabilize Flux2Klein tiny-diffusion startup by @lm10-arg in #10031
- [Bugfix][CI/Build][ROCm] Cap Qwen2.5-Omni BLAS threads by @lm10-arg in #10029
- [Bugfix][CI/Build][ROCm] Fix Qwen-Image profile resolution by @lm10-arg in #10028
- [Bugfix][CI/Build][ROCm] Stabilize remaining AMD nightly cold starts by @lm10-arg in #10032
- [Bugfix] Address MiniCPM-o 4.5 reanchor and NPU CFM batching review follow-ups by @yezhi0312 in #9973
- [Bugfix][CI/Build][ROCm] Avoid Nemotron checkpoint projection QR by @lm10-arg in #10034
- [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in #10035
- [Hardware][Ascend] Enable Qwen3-Omni V2 and Code2Wav graphs by @amy-why-3459 in #9910
- [Bugfix] Honor duplex handshake timeout during WebSocket upgrade by @lm10-arg in #10040
- [Bugfix] Honor the OpenAI
nparameter in omni serving by @KrzysztofTomala in #9592 - [Bugfix] Support mixed FP8/NVFP4 Cosmos3 checkpoint recipes by @rahul-steiger-nv in #9530
- [Bugfix][Hardware][AMD] Reduce ROCm pinned-host backing for module offload by @lm10-arg in #10047
- [Perf][NPU] Use torch_npu.npu_rotary_mul directly for NPU RoPE by @Dong1017 in #9892
- [Bugfix][Diffusion] Publish Qwen-Image-2.1 denoising steps to ForwardContext by @HAAZZZEEEE in #9455
- [Diffusion] Graduate Noisy PP cell path onto the request path by @ShengDev in #9926
- [CI/Build] Share Qwen3-Omni realtime servers to cut entrypoint cold starts by @yenuo26 in #9967
- [Model] MiniCPM-o: verify DSpark draft codes on the MRv2 Talker by @ischencheng in #9701
- [Bug] Stop counting an audio RTF sample for requests without audio by @Bevisy in #10011
- feat(rocm): support explicit online FP8 block scaling by @lm10-arg in #10075
- [CI/Build] Preserve Qwen3-Omni documentation audio evidence by @lm10-arg in #10081
- [CI/Build] Use five duration-weighted AMD Model Executor CPU shards by @hacZ17 in #10078
- [CI/Build] Split AMD Other CPU tests into two duration-weighted shards by @hacZ17 in #10079
- [CI/Build] Rebalance AMD Diffusion CPU shards and bound reference-test threads by @hacZ17 in #10080
- [perf][VoxCPM2]: default vae_decode_every=2 for zero-buffer continuity by @Lingwei1943 in #8929
- [Hardware][Ascend] Reclaim denoise-loop allocator blocks before streaming VAE decode by @dongyang1006 in #9313
- [Bugfix] Fence diffusion controls against request cleanup by @maithilijoshi20 in #9856
- [Bugfix][MiMo-Audio] Disable dropout before CUDA graph capture by @jeffaa729 in #8607
- [Core] Add encode_into buffer path for OmniSerializer by @apex-mochen in #9977
- [Model] Add Flux3 base BF16 image generation pipeline support by @Isotr0py in #9523
- [Ascend] Add fused SHMEM Ulysses all-to-all by @zwhzzz0821 in #9510
- [BugFix] Isolate attention schedule candidate tests from backend env by @ShengDev in #10091
- [Core][Diffusion] Apply attention schedules to Wan2.2 and HunyuanImage-3.0 by @zhongc1211 in #8718
- [Benchmark][Omni-DuplexEval] Match the official judge limits by @amy-why-3459 in #10069
- [Perf] AR-Diffusion: superpage KV cache for TMA attention loads by @chenchaoxu7575 in #9978
- [Bugfix][Step-Audio2] Support long audio input in Thinker by @Hniii98 in #9927
- [Bugfix] Apply sampling defaults across diffusion APIs by @FayeSpica in #9700
- [Doc] Qwen3-TTS: add 0.6B on 1x RTX 4070 Ti SUPER 16GB by @Sheehan20 in #7488
- [Feature] Add FP8 quantization for Qwen2.5-Omni (thinker LM only) by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/3466
- [Model] Add LFM3-Omni-2B on Ascend NPU by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/9827
- [CI/Build] Temporarily pause A5 Ready tests by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/10105
- [Bugfix][MammothModa2] Select mask-keyed CUDA graphs on AR FULL replay by @BLANKETusers in https://github.com/vllm-project/vllm-omni/pull/9911
- [CI] Attach OmniServer stderr to startup early-exit errors by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10098
- [Core] Add versioned duplex history snapshots by @Qing7364 in https://github.com/vllm-project/vllm-omni/pull/9311
- [Doc] Add MiniMax-H3-Image recipes by @princepride in https://github.com/vllm-project/vllm-omni/pull/10121
- [Model] Add PAN2 text-to-video and image-to-video by @wu-qing-157 in #8602
- [Bugfix][Step-Audio2] Use native chat format in offline ASR example by @Hniii98 in https://github.com/vllm-project/vllm-omni/pull/10050
- [CI/Build] Promote three AMD MERGE jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10082
- [CI/Build] Promote three AMD READY jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10083
- [Bugfix] Fix blurred Stable Diffusion 3.5 images with Cache-DiT when CFG is on by @blondeCS in https://github.com/vllm-project/vllm-omni/pull/8747
- [Doc] Add IFM/PAN2 recipe by @wu-qing-157 in https://github.com/vllm-project/vllm-omni/pull/10126
- [Doc] MiniMax-H3-Image: use minimax-h3-image-rocm image on MI350X by @hongxiayang in https://github.com/vllm-project/vllm-omni/pull/10129
- [XPU][Feature][Diffusion] Device graph capture and replay for denoise steps by @iboiko-habana in #9306
- [Core] Add duplex L1 context ledger and stage views by @Qing7364 in https://github.com/vllm-project/vllm-omni/pull/9441
- [CI/Build] Promote six AMD NIGHTLY jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10125
- [Model][Qwen-Image] Shard AdaLN modulation projections under TP by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/9440
- [Feature][Diffusion] Generic Hugging Face text-encoder TP; enable for Qwen-Image and Qwen-Image-2.1 by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/9438
- [CI/Build][ROCm] Guard AMD weekly reliability evidence by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10127
- [Perf][Diffusion] CFG batching: run both CFG branches as one denoiser call on one CFG rank by @xuechendi in #9797
- [CI/Build][ROCm] Target AMD nightly PRs to changed jobs by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10140
- [Bugfix] Initialize HunyuanVideo attention test biases by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10147
- [CI/Build] Keep Qwen3-Omni nightly evaluations for shared runtime changes by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10036
- [Model] LFM3: CUDA support and LFM3-Audio-2B by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/9968
- [CI/Build][ROCm] Expand AMD weekly reliability coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10148
- [Bugfix][ROCm] Fix Audex TTA startup metadata by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10154
- [CI/Build][ROCm] Add Covo-Audio weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10160
- [CI/Build] Share one MiniMax-H3 DP2 DLO server for FL2VA and Ref2VA by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10063
- [NPU] MiniCPM-o 4.5: full-MRv2 three-stage serving (default on NPU, Talker K-step K=8 + stop latch) by @y-null in https://github.com/vllm-project/vllm-omni/pull/10062
- [Model] Add MiniMax-H3-Image text-to-image and image editing by @princepride in #10120
- [Bugfix] Forward images num_frames for LingBot T2I by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10137
- [CI/Build] Calibrate LFM3-Omni streaming HNR floor by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/10166
- [CI/Build] Temporarily remove the Ulysses symmetric memory NPU Ready by @DarkMatter9527 in https://github.com/vllm-project/vllm-omni/pull/10146
- [Refactor][Diffusion] Let samplers control model calls and latent updates by @nodeeeeee in https://github.com/vllm-project/vllm-omni/pull/9088
- [Bugfix][MiniMax-H3] Wait for peers before tearing down the prestart test's process group by @ZhIMQ in https://github.com/vllm-project/vllm-omni/pull/9587
- [NPU][VoxCPM2] Expose cfg_value as VLLM_OMNI_SPEECH_CFG_VALUE for deployment tuning by @Lingwei1943 in https://github.com/vllm-project/vllm-omni/pull/9699
- [CI/Build][ROCm] Stabilize AMD invalid-parameter weekly jobs by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10159
- [CI/Build] Isolate diffusion attention test environments by @Bounty-hunter in https://github.com/vllm-project/vllm-omni/pull/10172
- [Core] Propose draft tokens on MRv2 AR stages; DSpark and padded DiT compilation for CosyVoice3 by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/9692
- [Model] Reduce MiniCPM-o eager codec cache copies by @WindChimeRan in https://github.com/vllm-project/vllm-omni/pull/9723
- [Bugfix][MiniCPM-o] Support portable Torch speaker placement by @WindChimeRan in https://github.com/vllm-project/vllm-omni/pull/9722
- [CI/Build] Repair deterministic pre-merge collection guards by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10182
- [CI/Build] Enable Model Runner V2 on weekly CUDA E2E jobs by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10073
- [CI/Build][ROCm] Extend pause-keep cold AITER JIT budget by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10189
- [Model][NPU] Host-side codec sampling and device K-step chain removal for MiniCPM-o 4.5 by @y-null in https://github.com/vllm-project/vllm-omni/pull/9990
- [Bugfix][NPU] Correct MOSS codec shared streaming attention masks by @heyew in https://github.com/vllm-project/vllm-omni/pull/10104
- [Bugfix][ROCm] Fix Audex Code2Wav startup metadata by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10187
- [Doc][NPU] Add the MiniMax-H3-Image Ascend NPU recipe by @brandneway in https://github.com/vllm-project/vllm-omni/pull/10155
- [Model] Enable MRv2 for JoyAI-VL-Interaction on CUDA by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10149
- [Hardware][Intel][MiniMax-H3] XPU video VAE decoder fast paths and geometry-chosen tile gather by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/9437
- [Bugfix] Handle scalar pooling outputs in stage metrics by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10188
- [Bugfix][Diffusion] Fix Cosmos3 Transfer SeaCache artifacts with target-only indicators by @yzhautouskay in https://github.com/vllm-project/vllm-omni/pull/7939
- [Bugfix] Read SharedMemoryConnector wakeup scope from flat config by @loveysuby in https://github.com/vllm-project/vllm-omni/pull/10171
- [Bugfix][ROCm] Force math SDPA for Wan VAE on gfx11 by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10205
- [Doc] Fix stale import path and moved links in contributor docs by @Williamren97 in https://github.com/vllm-project/vllm-omni/pull/10156
- [Bugfix] Clamp MiniMax-H3 shifted sigmas to [0, 1] by @zhongc1211 in https://github.com/vllm-project/vllm-omni/pull/10158
- [CI/Build][ROCm] Add weekly Audex, CosyVoice3, and Qwen3-TTS coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10180
- [Test] Relax cancel-bounds cleanup abort harness bound to 5s by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10183
- [CI/Build][ROCm] Isolate CustomVoice prefill parity from BF16 decoder drift by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10209
- [Recipe] Add Sophon-Realtime-32B-Tag (full-duplex speech-to-speech) by @BKtensor in https://github.com/vllm-project/vllm-omni/pull/10052
- [CI/Build] Combine Qwen3-Omni duplex C8 accuracy and performance by @psv666 in https://github.com/vllm-project/vllm-omni/pull/9623
- [CI/Build][ROCm] Stabilize Ming weekly stage startup by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10206
- [CI/Build][ROCm] Promote tiny diffusion READY coverage to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10210
- [CI/Build] Move Qwen-Image and Drama 3 to weekly and map Ascend910 to A3 by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10068
- [Bugfix] Reuse OpenCV captures on vLLM 0.33.0 by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10057
- [Bugfix] Register CFG companions before parent dispatch by @wutongyuonce in https://github.com/vllm-project/vllm-omni/pull/7104
- [CI/Build][ROCm] Promote seven NIGHTLY documentation and accuracy jobs by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10211
- [MUSA] MiniMax-H3-Image: enable the VAE graph decode on cuda-alike devices by @yeahdongcn in https://github.com/vllm-project/vllm-omni/pull/10139
- [Bugfix][Diffusion] Dispatch diffusion attention on out-of-tree platforms by @GoDiao in https://github.com/vllm-project/vllm-omni/pull/10071
- [CI/Build][ROCm] Stabilize HunyuanImage weekly startup by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10208
- [CI/Build][ROCm] Stabilize OOM sidecar saturation gate by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10218
- [Bugfix][Core] Route StagePool add_lora through the typed client helper by @JSs-u in https://github.com/vllm-project/vllm-omni/pull/9466
- [Doc] Add omni-perf-ceiling skill for serving performance ceilings by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10001
- [Core] Add runtime coordination for rank-paired weight sharing by @AbelSara in #9850
- [Bugfix] Give LLM stages typed owners for vLLM LoRA engine args by @MinhaoLi0318 in https://github.com/vllm-project/vllm-omni/pull/8574
- [Bugfix] Preserve MiniCPM-o duplex boundaries and sampled turn endings by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10066
- [Doc] Add single RTX 4090 MiniMax-H3-Image recipe by @yiminghub2024 in https://github.com/vllm-project/vllm-omni/pull/10245
- [Bugfix] Feed Voxtral MRV2 postprocess outputs by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10244
- [Quant] Enable native MXFP4 MoE by @yiliu30 in https://github.com/vllm-project/vllm-omni/pull/9749
- [Bugfix][CI] Give the weight-sharing worker test a concrete parallel … by @wangqiankun13 in https://github.com/vllm-project/vllm-omni/pull/10251
- [Model][NPU] MiniCPM-o 4.5: split estimator K/V cache, fused codec-tail sampling and a whole-decode graph on stage 0 by @y-null in https://github.com/vllm-project/vllm-omni/pull/10016
- [CI/Build] Keep MiniCPM-o 4.5 async_chunk cases adjacent to share one OmniServer by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10221
- [Bugfix][MiMo-Audio] Restore seeded sampling in eager and CUDA graphs by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10241
- [Bugfix] Fix GLM-Image positions for requested grids by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/9570
- [Bugfix] Serialize contiguous tensors whose last stride is not 1 by @wuli666 in https://github.com/vllm-project/vllm-omni/pull/10117
- [Bugfix] Remove placeholder tokenizer: gpt2 from skip_tokenizer_init deploy configs by @Shan2L in https://github.com/vllm-project/vllm-omni/pull/10145
- [CI/Build][ROCm] Add VoxCPM2 and Voxtral weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10212
- [CI] Plan CUDA ready and merge shards from timing artifacts by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/9923
- [Bugfix] Preserve HunyuanImage3 AR token identity for KV reuse by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/9323
- [Bugfix][Frontend] reject MP4 voice uploads at validation with the conversion hint by @Lingwei1943 in https://github.com/vllm-project/vllm-omni/pull/10087
- [Kernel] Add copy-engine transport for diffusion Ulysses by @zwhzzz0821 in https://github.com/vllm-project/vllm-omni/pull/9922
- [Model] Alpamayo 2 Super: n trajectories from one request by @KrzysztofTomala in https://github.com/vllm-project/vllm-omni/pull/10076
- [Doc] Add low precision attention docs for Cosmos3 by @bastefaniak in https://github.com/vllm-project/vllm-omni/pull/9538
- [Core] Share Wan2.2 TI2V VAE loading in OmniAutoencoderKLWan by @Lucaskabela in https://github.com/vllm-project/vllm-omni/pull/10196
- [Perf][Diffusion][Qwen-Image] Replace per-bucket varlen loop with flat varlen attention for mixfusion by @renyongxing in https://github.com/vllm-project/vllm-omni/pull/6711
- [CI/Build] Publish XPU wheel index and XPU release image by @Thangnguyenvn98 in https://github.com/vllm-project/vllm-omni/pull/10124
- [Hardware][Ascend] Support MOSS-TTS-Local on the NPU Model Runner V2 by @heyew in https://github.com/vllm-project/vllm-omni/pull/9953
- [Feat] Add HunyuanImage3 full-parameter RL rollout support by @Acerak01-fy in https://github.com/vllm-project/vllm-omni/pull/9535
- [Doc][NPU] Add Atlas 800T A3 performance to the MiniMax-H3-Image recipe by @DarkMatter9527 in https://github.com/vllm-project/vllm-omni/pull/10292
- [CI/Build] Fix invalid-audio MIME error guard by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10302
- [CI/Build] Shrink Ready Simple CPU mp matrices (noisy_kv + wan sharded) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10070
- [Bugfix][NPU] Bound non-streaming Qwen3-Omni Code2Wav batches by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10116
- [Bugfix] Retry MAGI-2 expert tiles after Triton resource rejection by @BLANKETusers in https://github.com/vllm-project/vllm-omni/pull/10291
- [CI][Duplex][MiniCPM-o] Enhance Full Duplex CI test of MiniCPM-o 4.5 by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/9555
- [Bugfix][Qwen-Image-2.1] Fix decode graph pool lifetime and stream recovery by @wangqiankun13 in https://github.com/vllm-project/vllm-omni/pull/10268
- [CI/Build] Fix Talker DSpark drafter targeting wrong stage in MiniCPM-o 4.5 E2E by @y-null in https://github.com/vllm-project/vllm-omni/pull/10316
- [Test] Unify DFX stability runner and add LingBot-World (GPU&NPU) case by @wangwei1254 in https://github.com/vllm-project/vllm-omni/pull/10250
- [Model] Add Hy-Nomas CPU history archive with failure cleanup by @mooncake2026 in https://github.com/vllm-project/vllm-omni/pull/10242
- [CI/Build][ROCm] Bound OOM residual reporting slack by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10309
- [Bugfix] Install pan2-guardrail in the CI image by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10247
- [CI] Move Qwen3-Omni colocate async to full_model nightly; drop HunyuanImage3 shared TP4 jobs by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10295
- [CI/Build][ROCm] Add Moss and Higgs weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10280
- [Bugfix] Bind FlashAttention arguments by name in the packed Ming Whisper attention by @alex7092 in https://github.com/vllm-project/vllm-omni/pull/9751
- [Diffusion] Add Batch for Noisy PP by @ShengDev in #10122
- [CI/Build][ROCm] Add OmniVoice and Step-Audio weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10326
- [CI/Build][ROCm] Promote four NIGHTLY GPU smoke jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10270
- [Bugfix][Quantization] mxfp8: honour ignored_layers match mode by @yiliu30 in https://github.com/vllm-project/vllm-omni/pull/10282
- [CI/Build][ROCm] Promote four NIGHTLY unit and inference jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10271
- [Bugfix] Segment long IndexTTS-2.5 speech requests by @voidreaming in https://github.com/vllm-project/vllm-omni/pull/7567
- [Perf] Wan2.2 S2V: compile the audio injection with the DiT blocks by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/10077
- [Core] FlashInfer autotune for diffusion stages (on by default on SM100+) by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/9998
- [Kernel][Diffusion] Wan VAE fast path: one installer for every Wan-family VAE by @xuechendi in https://github.com/vllm-project/vllm-omni/pull/9553
- [Bugfix][GLM-Image] Route offline prompts through the AR multimodal processor by @tthakkal in https://github.com/vllm-project/vllm-omni/pull/7958
- [Model][Qwen-Image-2.1] ring SP: attend the replicated prefix locally by @yiliu30 in https://github.com/vllm-project/vllm-omni/pull/10283
- [CI/Test] Repair FlashInfer CUDA contracts by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10346
- [Core][Diffusion] Add the weight_stack package foundation (RFC #9536 01-1) by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/9725
- [bugfix] Fix FA test_vllm_dense_adapter_cuda_matches_sdpa_and_lse test for Blackwell by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/9217
- [CI/Build][ROCm] Add Ming GLM IndexTTS and Breeze weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10329
- [CI/Build] Tune CUDA merge timeouts and reorganize Qwen3-Omni tests by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10286
- [Model] Enable CUDA MRv2 K-step for MiniCPM-o 4.5 by @y-null in #10288
- [Core][Diffusion] Select attention profiles from scheduler noise by @zhongc1211 in https://github.com/vllm-project/vllm-omni/pull/8746
- [CI/Build][Bugfix][ROCm] Complete shared TTS weekly coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10350
- [Model][Frontend] Add MiniMax-H3 character replacement ComfyUI workflow (H3-04) by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/9194
- [Model] Add Kandinsky6 Distil and integrated HD/FHD generation by @osipovmekete in #9027
- [Bugfix] Preserve requested layer count for Ming-Image Design-Layer by @Acerak01-fy in https://github.com/vllm-project/vllm-omni/pull/10318
- [Bugfix] Shard fused diffusion LoRA deltas for tensor parallelism by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/9708
- [Bugfix] Synchronize UNI diffusion output handoff by @BeiningWu in https://github.com/vllm-project/vllm-omni/pull/9804
- [1/4][Model] Add native Praxis-1 video/action policy (DROID) by @Lucaskabela in #10197
- [diffusion][model] Enable CPU/layerwise offload for LingBot-Video by @pandalee99 in https://github.com/vllm-project/vllm-omni/pull/5359
- [CI/Build][ROCm][Bugfix] Match remaining TTS weekly paths by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10367
- [CI/Build][ROCm] Add BAGEL weekly coverage and fix routed LoRA by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10382
- [CI/Build] Enable Omni merge sharding and set H100 job timeouts to 25 minutes by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10372
- [2/4][Model] Add YAM two-arm support to Praxis-1 by @Lucaskabela in #10204
- [CI/Build][ROCm] Promote weekly parity coverage to blocking by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10392
- [Hardware][Ascend][perf] Enable MiniCPM-o 4.5 Thinker Ngram on MRv2 by @y-null in https://github.com/vllm-project/vllm-omni/pull/10313
- [Perf][Hy-NoMas]integrate native PixelUMM decoder by @BruceLoveDecimal in https://github.com/vllm-project/vllm-omni/pull/9567
- [Bugfix][Frontend] Keep the stream alive when a queued interaction is rejected by @zyuanbing in https://github.com/vllm-project/vllm-omni/pull/8560
- [Model] Improve LingBot-World streaming performance by @nagisa-kunhah in https://github.com/vllm-project/vllm-omni/pull/10383
- [3/4][Doc] Add the Praxis-1 recipe and worker tests by @Lucaskabela in https://github.com/vllm-project/vllm-omni/pull/10198
- [Hardware][NVIDIA] Enable MiniMax-H3-Image MXFP8 serving on Rubin by @lishunyang12 in https://github.com/vllm-project/vllm-omni/pull/10390
- [Model] Hy-Nomas: quantize chunk frames on device and stage delivery by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/10113
- [4/4][Model] Ship the Praxis-1 DROID and YAM deployments with checkpoint-gated tests by @Lucaskabela in https://github.com/vllm-project/vllm-omni/pull/10202
- [Core][Diffusion] Retain per-step history for Same-Step sessions by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/9817
- [Model] Hy-Nomas: add step-level zero-sync caching and regional compilation by @yiwen-cai in https://github.com/vllm-project/vllm-omni/pull/10246
- [Core] Add Mooncake KV/HS caching for Qwen3-Omni turn serving by @ztang2370 in #10107
- [Core][Diffusion] Place sessions on the replica that holds them by @kunkunblueberry in https://github.com/vllm-project/vllm-omni/pull/9821
- [Bugfix] Fix MOSS-TTS-Realtime request-local sampling by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10411
- [CI/Build][ROCm] Add Drama 3 weekly burn-in coverage by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10413
- [Bugfix] Unwrap multimodal outputs during CUDA graph capture by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10414
- [Bugfix] Enable Cache-DiT for SenseNova-U1 by @Laceprndpm in https://github.com/vllm-project/vllm-omni/pull/8248
- [Bugfix] Load MOSS-TTS checkpoint processors without AutoProcessor by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/10377
- [CI/Build][Ascend] Fix Fast Ulysses NPU Ready by @zwhzzz0821 in https://github.com/vllm-project/vllm-omni/pull/10248
- [CI/Build][ROCm] Add Cosmos3 weekly generation-mode burn-in by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10416
- [Bugfix] Reset MiniCPM-o NPU host sampler at duplex segment boundaries by @heyew in https://github.com/vllm-project/vllm-omni/pull/10319
- [Model] Add Haru-4B full-duplex spoken dialogue model by @linyueqian in #10301
- [Perf][Hy-Nomas] Parallelize PixelUMM decode windows by @wtz2333 in https://github.com/vllm-project/vllm-omni/pull/9961
- [Bugfix] Support up to 10 reference images in Qwen-Image 2.1 by @Nankago in https://github.com/vllm-project/vllm-omni/pull/9770
- [Core] Pin the duplex context boundary through the real scheduler (RFC #7389 W6) by @ztang2370 in https://github.com/vllm-project/vllm-omni/pull/8784
- [Model] MiniCPM-o 4.5 Seed-TTS runtime and deployment optimizations by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10331
- [Core] Add Qwen3-Omni Thinker DFlash support Phase 1 Thinker-only speculative decoding by @YuDeng0102 in https://github.com/vllm-project/vllm-omni/pull/9595
- [Doc] Update vLLM-Omni WeChat QR code by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/10423
- [Perf] diffusion IPC: ship bf16 SHM tensors as raw bits and copy host… by @MaciejBalaNV in https://github.com/vllm-project/vllm-omni/pull/9487
- [Bugfix] Preserve committed camera frames when coalescing duplex turns by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10289
- [Model] Remove Kokoro-82M and JoyAI-Image-Edit support by @yenuo26 in #10393
- [Hardware][Ascend] Reuse LingBot paged attention lengths and block conditions by @Fishermanykx in https://github.com/vllm-project/vllm-omni/pull/10427
- [Hardware][Ascend] Optimize Qwen3-TTS codec and predictor inference by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/9068
- [Model] Add replaceable Qwen3-Omni duplex history policies by @psv666 in https://github.com/vllm-project/vllm-omni/pull/9960
- [Hardware][Ascend] Enable MiniMax-H3-Image VAE graph execution by default by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10184
- [CI/Build] Add JoyAI long-session lifecycle acceptance tests by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10386
- [Model] Move ABot-World live camera control onto the stepwise interaction path by @zyuanbing in https://github.com/vllm-project/vllm-omni/pull/10207
- [CI/Build][ROCm] Promote final AMD weekly parity jobs by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10438
- [Bugfix] Retry OmniServer start on EADDRINUSE by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10176
- [Bugfix][Ascend] Restore MiniCPM-o MRv2 FULL graphs and two-card Nightly by @y-null in https://github.com/vllm-project/vllm-omni/pull/10435
- [Model] Run MOSS-TTS-Nano on Model Runner V2 by @BeatSeat in https://github.com/vllm-project/vllm-omni/pull/9905
- [Bugfix] Remove hardcoded environment inventory count test by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10456
- [CI/Build] Fix Qwen duplex history policy coverage by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10455
- [MiniMax-H3] Select model runner v2 for the encoder stage by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10074
- [Perf][LingBot World] Add WorldKV memory retrieval by @BruceLoveDecimal in https://github.com/vllm-project/vllm-omni/pull/9945
- [CI/Build][ROCm] Promote nightly weekly-parity owners by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10457
- [Doc] Record Ref2VA RainFusion points for 86 GB 950PR by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/9955
- [Perf][Sana-WM] Add opt-in compiled camera-prep leaf optimization by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/8738
- [CI/Build] Parallelize bounded AMD image publication by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10436
- [Core] Support Qwen3-Omni duplex sessions with Mooncake KV/HS caching by @ztang2370 in #10321
- [Bugfix][Wan2.2] Handle the None decode result on non-decoding pipeline-parallel stages by @klin2024 in https://github.com/vllm-project/vllm-omni/pull/10404
- [Bugfix][Breeze-TTS-2] Run Breeze stage 0 on Model Runner V2 by @searlion in https://github.com/vllm-project/vllm-omni/pull/10399
- [Bugfix][Diffusion] Use the diffusers guidance_scale default for SDXL when omitted by @alex7092 in https://github.com/vllm-project/vllm-omni/pull/10462
- [Bugfix][Qwen-Image-Edit] Initialize the text encoder on CPU by @wangqiankun13 in https://github.com/vllm-project/vllm-omni/pull/10451
- [Bugfix] Validate VoxCPM2 reference audio before resolution by @dongbo910220 in https://github.com/vllm-project/vllm-omni/pull/10357
- [Bugfix] Preserve CosyVoice3 DSpark capture contract by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10485
- [CI/Build] Avoid TCP rendezvous races in GLM-Image CPU tests by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10479
- [CI/Build][ROCm] Promote NIGHTLY collective and offloader jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10336
- [CI/Build][ROCm] Promote three NIGHTLY full-model jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10337
- [Model] Improve YuE2 latency and throughput with automatic ABC and paired CFG by @lzqlzzq in https://github.com/vllm-project/vllm-omni/pull/9098
- [Bugfix] Reject blank Qwen3-TTS text before it reaches the Talker by @searlion in https://github.com/vllm-project/vllm-omni/pull/10380
- [CI/Build] Align model reliability tests with the fault-mode library by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10402
- [CI/Build] Enforce 10% performance baseline regression checks by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10395
- [CI/Build][Bugfix] Resolve LTX-2.5 Gemma4 RoPE config by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10467
- [Diffusion] Add request-path CHUNK_READY realtime video output by @ShengDev in #10049
- [Bugfix] Restore Qwen3-Omni Talker FULL CUDA graphs by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10499
- [Bugfix] Restore Qwen3-TTS MRv2 FULL graph replay by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10503
- [Feature][Diffusion] Add the Heun sampler for MiniMax H3 by @nodeeeeee in https://github.com/vllm-project/vllm-omni/pull/10400
- [Bugfix][Step-Audio2] Align example prompts with the upstream message format by @Hniii98 in https://github.com/vllm-project/vllm-omni/pull/10191
- [CI/Build] Re-import DFX diffusion runner without a stale package attribute by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10523
- [Core] Duplex tool-call enhancement by @chickeyton in #9909
- [CI/Build] Add LingBot World accuracy and performance CI by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/9009
- [Bugfix][VoxtralTTS] Tokenizer & LM has hardcoded torch.bfloat16 component, crashing on fp16-only GPUs (T4/sm75) by @y123456y78 in https://github.com/vllm-project/vllm-omni/pull/9324
- [Bugfix][Config] Check strategy device layout after stage overrides are merged by @fusheng-ji in https://github.com/vllm-project/vllm-omni/pull/9586
- [Bugfix][MiniMax-H3] Skip FastH3 checkpoint gate check when a host-weight plan owns the DiT by @david6666666 in https://github.com/vllm-project/vllm-omni/pull/10498
- [Bugfix] Release unused MOSS-TTS codec modules by @RongHuang14 in https://github.com/vllm-project/vllm-omni/pull/10345
- [Bugfix] Expire video job metadata with stored artifacts by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/10535
- [Bugfix] Preserve multimodal payloads across stage-0 replicas by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10539
- [Bugfix] Exclude incomplete Higgs Audio v3 frames before codec decoding by @ss1127201970 in https://github.com/vllm-project/vllm-omni/pull/7942
- [Bugfix] Add Chinese punctuation normalization for IndexTTS-2 by @xiny-spec in https://github.com/vllm-project/vllm-omni/pull/5590
- [CI/Build] Reduce repeated setup in Simple Test CPU coverage by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10469
- [Model] Implement full async_chunk for JoyAI-VL-Interaction by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/10281
- [CI/Build][ROCm] Recognize quoted HIP OOM evidence by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10514
- [CI/Build][ROCm] Promote three NIGHTLY model-validation jobs to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10481
- [Model] Run MOSS-TTS-Realtime on Model Runner V2 by @BeatSeat in https://github.com/vllm-project/vllm-omni/pull/9906
- [Bugfix][Hardware][NVIDIA] Use FlashInfer fa2, not cute-dsl, for auto attention on SM12x by @DuFanYin in https://github.com/vllm-project/vllm-omni/pull/10426
- [Core] Support duplex replicas across stage runtimes and nodes by @chickeyton in #10422
- [Benchmark] Add MOSS-TTS-Realtime RTX 5060 Ti 16GB baseline by @RongHuang14 in https://github.com/vllm-project/vllm-omni/pull/10312
- [Frontend] Report token-budget truncation for Qwen3-TTS, Breeze-TTS-2 and Drama 3 by @YingxuH in https://github.com/vllm-project/vllm-omni/pull/10556
- [CI/Build][ROCm] Promote NIGHTLY tiny diffusion 2-GPU to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10519
- [Frontend] Document and test browser CORS options by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/10389
- [Bugfix] Close Omni connectors during runner shutdown by @lm10-arg in https://github.com/vllm-project/vllm-omni/pull/10559
- [CI/Build] Increase duplex resume test synchronization timeout by @LinzeShi in https://github.com/vllm-project/vllm-omni/pull/10480
- [Frontend] Fish Speech demo: gapless AudioWorklet streaming playback by @nagelanping in https://github.com/vllm-project/vllm-omni/pull/4836
- [BugFix] Respect user specified dtype in Qwen3 TTS by @N-Harish in https://github.com/vllm-project/vllm-omni/pull/8284
- [CI/Build][ROCm] Promote NIGHTLY MiniMax-H3 accuracy to blocking by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10560
- [CI/Build][ROCm] Prepare BAGEL and duplex NIGHTLY qualification by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10561
- [Model][PersonaPlex] Serve duplex calls past the 3000-frame window (RFC #7389 W6) by @ztang2370 in https://github.com/vllm-project/vllm-omni/pull/8783
- [Bugfix][ROCm] Use ieee input_precision in Qwen3-TTS time-major conv on ROCm by @menglcai in https://github.com/vllm-project/vllm-omni/pull/10510
- [CI/Build][ROCm] Require MI300 NIGHTLY performance baselines by @hacZ17 in https://github.com/vllm-project/vllm-omni/pull/10562
- [Bugfix][Diffusion] Support MXFP8 E8M0 scales in DLO AllGather by @yiliu30 in https://github.com/vllm-project/vllm-omni/pull/10403
- [Misc] Read diffusion_attention_schedule directly from the config by @zhongc1211 in https://github.com/vllm-project/vllm-omni/pull/10375
- [Model] Batch cache-aware Nemotron VoiceChat duplex perception by @sinpeyw in https://github.com/vllm-project/vllm-omni/pull/8597
- [CI/Build] Add nightly CosyVoice3 cross-request voice check by @Arifuzzamanjoy in https://github.com/vllm-project/vllm-omni/pull/8689
- [Bugfix][Qwen3-TTS] Reject blank text sent through model_intermediate_buffer by @searlion in https://github.com/vllm-project/vllm-omni/pull/10554
- [Refactor] Wire DiffusionTTSAdapter into diffusion-mode online serving by @sphinxkkkbc in https://github.com/vllm-project/vllm-omni/pull/8011
- [Core] Align with vLLM 0.34.0 and remove ModelRunner V1 by @tzhouam in #10365
- [Bugfix] Guard MiniCPM-o flow encoder graph capture against cyclic GC by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10581
- [Bugfix] Backport diffusion worker startup orphan fix to v0.34.0 by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10619
- [Bugfix][BAGEL] Align two-stage contexts with reference inference by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10623
- Revert "[Bugfix][BAGEL] Align two-stage contexts with reference inference" by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10635
- [CI/Build] Backport Wan I2V reference prompt alignment to v0.34.0 by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10620
- [Bugfix][BAGEL] Backport two-stage context fixes to v0.34.0 by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10637
- [Bugfix] Backport TTS synthesis error handling to v0.34.0 (#9693) by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10709
- [Hardware][Ascend] Backport Qwen3-Omni text-only latency fix to v0.34.0 by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10696
- [CI/Build] Backport BAGEL golden update to v0.34.0 by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10712
- [CI/Build][release/v0.34.0] Backport Qwen3-Omni runtime OOM workload fix by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10720
- [Cherry-pick] [Bugfix] Clean request artifacts on pre-registration rejection by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10433
- [Bugfix] Fix completed duplex playback interruption on v0.34.0 by @y-null in https://github.com/vllm-project/vllm-omni/pull/10706
- [Cherry-pick][Bugfix][test]: accept both staged-write representations in rope shim test by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10702
- [CI/Build] Backport Qwen3-Omni LoRA test startup timeout to v0.34.0 by @psv666 in https://github.com/vllm-project/vllm-omni/pull/10701
- [Bugfix][v0.34.0] Backport MRv2 audio transport fixes (#10721) by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10750
- [Bugfix][BAGEL] Backport single-stage conditioning fixes to v0.34.0 by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10733
- [Bugfix][v0.34.0] Backport duplex silence-continuation jitter fix (#9537) by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10763
- [Bugfix][VoxCPM2] Backport edge digit pronunciation fixes to v0.34.0 by @y-null in https://github.com/vllm-project/vllm-omni/pull/10751
- [Bugfix][FLUX.2] Backport output-size validation to v0.34.0 by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10744
- [Bugfix][release/v0.34.0] Backport HunyuanImage3 MR2 sampling fix by @Bounty-hunter in https://github.com/vllm-project/vllm-omni/pull/10740
- [Bugfix] Backport MiniCPM duplex forced-response latch fix (#10546) to v0.34.0 by @y-null in https://github.com/vllm-project/vllm-omni/pull/10647
- [Bugfix][v0.34.0] Backport NIXL loopback replica-host fix (#9651) by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10777
- [Hardware][Ascend] Backport NPU stack upgrade to v0.34.0 (#10691) by @amy-why-3459 in #10773
- [Cherry-pick][Bugfix][NPU] Fall back to pageable memory when pinned host allocation fails by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10776
- [Cherry-pick][Metrics] Exclude diffusion scheduler queue wait from stage generation time by @ZacheryAU in https://github.com/vllm-project/vllm-omni/pull/10784
- [Bugfix][NPU] Backport profiler stack flag fix to v0.34.0 by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10771
- [CI/Build] [v0.34.0] Limit GPU doc tests to eight model families by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10741
- [Bugfix][release/v0.34.0] Backport HunyuanImage3 request VAE RNG fix (#9382) by @Bounty-hunter in https://github.com/vllm-project/vllm-omni/pull/10793
- [Bugfix] Backport re-saved Z-Image text encoder loading to v0.34.0 by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/10756
- [Bugfix][v0.34.0] Backport buffered speech idle lease fix (#10617) by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10801
- [Cherry-pick][NPU] Log the torch_npu / CANN build identity at platform init by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10789
- [Bugfix][v0.34.0] Backport duplex receiver generation fix (#10755) by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10799
- [Cherry-pick][Bugfix] Fix full-payload accumulation across preemption and aborts (#10692) by @BLANKETusers in https://github.com/vllm-project/vllm-omni/pull/10797
- [CI/Build] [v0.34.0] Backport video server test wait budget (#10687) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10803
- [CI/Build] [v0.34.0] Preserve OmniServer stderr after readiness by @Bounty-hunter in https://github.com/vllm-project/vllm-omni/pull/10796
- [Cherry-pick][Bugfix] Restore native ModelOpt NVFP4 MoE inference (#10264) by @BLANKETusers in https://github.com/vllm-project/vllm-omni/pull/10815
- [NPU] Bump mindiesd from 3.2.0 to 3.3.0 in NPU Dockerfiles by @DarkMatter9527 in https://github.com/vllm-project/vllm-omni/pull/10850
- [Bugfix][release/v0.34.0] Bound Omni-DuplexEval judge frame resolution (#9038) by @natureofnature in https://github.com/vllm-project/vllm-omni/pull/10809
- [Bugfix][ROCm][v0.34.0] Backport Model Runner V2 deploy profiles (#10630) by @IndigoTan in #10819
- [Bugfix][v0.34.0] Avoid the Ascend A5 platform initialization cycle (#10707) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10822
- [Bugfix][v0.34.0] Backport Qwen-Image-2.1 per-request image RoPE (#10376) by @wangqiankun13 in https://github.com/vllm-project/vllm-omni/pull/10804
- [Bugfix][v0.34.0] Handle empty Qwen3-Omni talker input projections (#8155) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10831
- [Bugfix][v0.34.0] Return all Cosmos3 images with guardrails enabled (#10652) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10827
- [Bugfix][v0.34.0] Handle local Q and global KV in packed NPU attention (#8632) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10825
- [Bugfix][v0.34.0] Restore diffusion INT8 kernel dispatch on ROCm (#10708) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10823
- [Bugfix][v0.34.0] Fix FA4 Ring attention compilation (#8988) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10824
- [Bugfix][v0.34.0] Restore the Fish attention fast path for strided KV caches (#5943) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10826
- [Bugfix][v0.34.0] Capture Qwen Code2Wav graphs after model load (#10593) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10829
- [Bugfix][v0.34.0] Handle packed tail padding in SAGE attention (#10729) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10841
- [Bugfix][v0.34.0] Honor the request seed in Voxtral TTS (#6371) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10832
- [Bugfix][Frontend] Backport diffusion-only Anthropic 501 responses to v0.34.0 by @ShengDev in https://github.com/vllm-project/vllm-omni/pull/10807
- [Bugfix][v0.34.0] Pass num_logits to MRv2 LoRA input construction (#10772) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10821
- [Bugfix][v0.34.0] Isolate inline VoxCPM2 voice references (#10361) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10833
- [Bugfix][v0.34.0] Keep full MRv2 graph candidates for tensor-returning stages (#10636) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10828
- [Bugfix][v0.34.0] Preserve the Qwen3-Omni Code2Wav embedding dtype (#8327) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10830
- [Bugfix][v0.34.0] Load the Breeze-TTS-2 codec on Transformers 5.22 (#10651) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10834
- [Bugfix][v0.34.0] Decode voice uploads and normalize audio MIME types (#10673) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10835
- [Bugfix][v0.34.0] Honor REST speech parameters on the speech WebSocket (#10739) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10836
- [Bugfix][v0.34.0] Return HTTP 400 for speech client errors (#8277) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10837
- [Cherry-pick][Bugfix][Profiler] Build the wait/warmup/active schedule in the omni torch-profiler wrappers by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10810
- [CherryPick][Bugfix] Normalize Qwen distill LoRA keys before fusion by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/10806
- [Bugfix][v0.34.0] Stop serving benchmarks on VAD and warmup errors (#10626) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10839
- [CherryPick][Bugfix] Dequantize packed NVFP4 weights for full-precision diffusion layers by @NumberWan in https://github.com/vllm-project/vllm-omni/pull/10785
- [Bugfix][v0.34.0] Warn about unsupported stage overrides (#9650) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10840
- [CI/Build][v0.34.0] Backport HunyuanImage performance and accuracy matrix (#10408) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10816
- [Bugfix][v0.34.0] Honor the duplex client handshake timeout (#9084) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10838
- [CI/Build][v0.34.0] Backport LingBot capture runtime pin removal by @IndigoTan in https://github.com/vllm-project/vllm-omni/pull/10818
- [Bugfix][v0.34.0] Backport the Drama 3 silence fix (#10669) by @zengchuang-hw in https://github.com/vllm-project/vllm-omni/pull/10775
- [Bugfix][ROCm][v0.34.0] Backport MRv2 runtime and deploy fixes (#10683) by @IndigoTan in #10820
- [Bugfix][v0.34.0] Backport MiniCPM-o cleanup fix (#10738) by @y-null in https://github.com/vllm-project/vllm-omni/pull/10866
- [CI/Build][v0.34.0] Backport invalid parameter matrices to CPU tests (#10719) by @yenuo26 in https://github.com/vllm-project/vllm-omni/pull/10862
- [Bugfix][v0.34.0] Backport four TTS and serving fixes by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/10867
- [Model][v0.34.0] Backport Music3/YuE2 MRV2 and Qwen3-TTS/Higgs fixes by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10849
- [Bugfix][v0.34.0] Backport Music3 audio delivery and YuE2 review fixes by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10875
- [Bugfix][v0.34.0] Restore LingBot Video image generation capability (#10886) by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10887
- [Bugfix][v0.34.0] Fix Qwen3-TTS first-packet P50 regression (#10851) by @FayeSpica in https://github.com/vllm-project/vllm-omni/pull/10872
- [Bugfix] Backport PNG encoder fork fix to v0.34.0 by @IndigoTan in https://github.com/vllm-project/vllm-omni/pull/10731
- [CI/Build][v0.34.0] Backport LingBot World Ready and Merge gates (#10863) by @tzhouam in https://github.com/vllm-project/vllm-omni/pull/10865
- [Cherry-pick][Bugfix] MiniMax-H3: validate the LightX2V Turbo 8-step v1.2 LoRA layout by @Dong1017 in https://github.com/vllm-project/vllm-omni/pull/10853
- [Bugfix][v0.34.0] Backport the JoyAI delegate prompt order fix (#10847) by @Gaohan123 in https://github.com/vllm-project/vllm-omni/pull/10905
- [Bugfix][v0.34.0] Fix Qwen-Image-2.1 mixed-length RoPE on Ascend by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10888
- [Bugfix][v0.34.0] Backport JoyAI history and frontend shutdown fixes by @amy-why-3459 in https://github.com/vllm-project/vllm-omni/pull/10902
New Contributors
- @esmeetu made their first contribution in #9184
- @mooncake2026 made their first contribution in #9204
- @aayc123 made their first contribution in #7708
- @shijiecao-01 made their first contribution in #8741
- @Vivo50E made their first contribution in #1330
- @zhang-s-a made their first contribution in #8072
- @KrzysztofTomala made their first contribution in #9120
- @Shan2L made their first contribution in #8802
- @WindChimeRan made their first contribution in #8943
- @YawningPants made their first contribution in #9244
- @advanced233 made their first contribution in #9266
- @cipheraxat made their first contribution in #9041
- @ZhIMQ made their first contribution in #8868
- @xdlsy made their first contribution in #9074
- @tzielinski-habana made their first contribution in #9255
- @LJX-xixi made their first contribution in #5019
- @leonail1 made their first contribution in #6374
- @apapi made their first contribution in #8875
- @ssfdgsg made their first contribution in #4738
- @scestola-h made their first contribution in #7643
- @habibrosyad made their first contribution in #4304
- @mayooot made their first contribution in #8093
- @moezdil made their first contribution in #9596
- @wzk-jazk made their first contribution in #9246
- @Zssan12 made their first contribution in #8661
- @RishabhSaini made their first contribution in #8298
- @Alex-Wengg made their first contribution in #9637
- @MohammadHijjawi97 made their first contribution in #8354
- @panpan0000 made their first contribution in #9810
- @Mengmj made their first contribution in #9099
- @spped2000 made their first contribution in #9954
- @Yuy1L1 made their first contribution in #9729
- @dongyang1006 made their first contribution in #9313
- @apex-mochen made their first contribution in #9977
- @Hniii98 made their first contribution in #9927
- @Sheehan20 made their first contribution in #7488
- @wu-qing-157 made their first contribution in #8602
- @hongxiayang made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10129
- @heyew made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10104
- @Williamren97 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10156
- @BKtensor made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10052
- @JSs-u made their first contribution in https://github.com/vllm-project/vllm-omni/pull/9466
- @Lucaskabela made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10196
- @renyongxing made their first contribution in https://github.com/vllm-project/vllm-omni/pull/6711
- @Thangnguyenvn98 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10124
- @voidreaming made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7567
- @BeiningWu made their first contribution in https://github.com/vllm-project/vllm-omni/pull/9804
- @pandalee99 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5359
- @zyuanbing made their first contribution in https://github.com/vllm-project/vllm-omni/pull/8560
- @yiwen-cai made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10246
- @ztang2370 made their first contribution in #10107
- @Nankago made their first contribution in https://github.com/vllm-project/vllm-omni/pull/9770
- @YuDeng0102 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/9595
- @klin2024 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10404
- @lzqlzzq made their first contribution in https://github.com/vllm-project/vllm-omni/pull/9098
- @RongHuang14 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10345
- @ss1127201970 made their first contribution in https://github.com/vllm-project/vllm-omni/pull/7942
- @xiny-spec made their first contribution in https://github.com/vllm-project/vllm-omni/pull/5590
- @N-Harish made their first contribution in https://github.com/vllm-project/vllm-omni/pull/8284
- @menglcai made their first contribution in https://github.com/vllm-project/vllm-omni/pull/10510
- @sinpeyw made their first contribution in https://github.com/vllm-project/vllm-omni/pull/8597
Full Changelog: v0.32.0...v0.34.0