[Bugfix][MiniMax-H3] Apply pre-commit formatting

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>
[Bugfix][MiniMax-H3] Reconcile temporal decode plans across VAE ranks
2026-09-03 08:44:26 +08:00 · 2026-09-03 08:20:00 +08:00 · 2026-09-02 07:08:40 +08:00 · 2026-09-01 23:11:29 +08:00 · 2026-09-01 23:11:29 +08:00 · 2026-09-01 23:11:28 +08:00
106 changed files with 7080 additions and 507 deletions
@@ -238,13 +238,10 @@ steps:
      grade: Blocking
      timeout_in_minutes: 50
      commands:
-        - |
-          timeout 40m bash -c '
-            export VLLM_LOGGING_LEVEL=DEBUG
-            export VLLM_WORKER_MULTIPROC_METHOD=spawn
-            export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
-            pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m 'advanced_model and cuda' --run-level 'advanced_model'
-          '
+        - export VLLM_LOGGING_LEVEL=DEBUG
+        - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+        - export VLLM_ALLOW_LONG_MAX_MODEL_LEN="1"
+        - timeout 40m pytest -s -v tests/e2e/online_serving/test_qwen3_tts_base.py tests/e2e/offline_inference/test_qwen3_tts_base.py -m "advanced_model and cuda" --run-level "advanced_model"

 # TODO: Bagel test on ROCm is very unstable. @tjtanaa
 # Need to debug before reneable numerical changes across large PRs
@@ -237,3 +237,38 @@ steps:
              pytest -s -v tests/e2e/online_serving/test_hunyuan_video_15_expansion.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
            "
        mirror_hardwares: h100_1
+
+      - label: "Diffusion · MiniMax-H3 L3 DLO DP2 and FastH3"
+        source_file_dependencies:
+          - tests/e2e/online_serving/minimax_h3/
+          - tests/helpers/assertions.py
+          - tests/helpers/mark.py
+          - tests/helpers/media.py
+          - tests/helpers/runtime.py
+          - vllm_omni/diffusion/models/minimax_h3/
+          - vllm_omni/diffusion/lora/
+          - vllm_omni/diffusion/offloader/
+          - vllm_omni/diffusion/distributed/
+          - vllm_omni/entrypoints/openai/
+        # Keep four server lifecycles in one Buildkite job, but never in one
+        # pytest process. DLO uses multiprocessing shared-memory queues; the
+        # parent's resource_tracker can retain a stale /psm_* entry during
+        # teardown and strand the next Ref2VA collective.
+        # A per-file process plus a hard timeout also makes a regression fail
+        # instead of leaving the merge queue stuck indefinitely.
+        timeout_in_minutes: 150
+        commands:
+          - |
+            set -eu
+            export VLLM_IMAGE_FETCH_TIMEOUT=60
+            if [ -f vllm_omni/diffusion/models/minimax_h3/lora.py ]; then
+              export VLLM_TEST_MINIMAX_H3_TURBO_LORA=$(python3 -c 'from huggingface_hub import hf_hub_download; print(hf_hub_download(repo_id="lightx2v/Minimax-h3-Turbo", filename="minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors"))')
+            fi
+            if [ -f vllm_omni/diffusion/models/minimax_h3/fasth3.py ]; then
+              export VLLM_TEST_MINIMAX_H3_FASTH3_LORA=$(python3 -c 'from huggingface_hub import hf_hub_download; print(hf_hub_download(repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA", filename="dense-datafree/adapter_model.safetensors"))')
+            fi
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_dlo_dp2_fl2va.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_dlo_dp2_ref2va.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_turbo_lora.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+            timeout 35m pytest -s -v tests/e2e/online_serving/minimax_h3/test_minimax_h3_fasth3_lora.py -m 'advanced_model and cuda and diffusion' --run-level 'advanced_model'
+        mirror_hardwares: h100_2
@@ -16,6 +16,7 @@ Easy, fast, and cheap omni-modality model serving for everyone

 *Latest News* 🔥

+- [2026/08] We released [0.28.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.28.0), featuring production-ready [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) serving on GPU and NPU, a unified AR/DiT paged KV cache runtime, and enhanced realtime full-duplex serving for the [MiniCPM-o series](recipes/OpenBMB/MiniCPM-o-4_5.md).
 - [2026/08] [VeRL-Omni](https://github.com/verl-project/verl-omni) `v0.2.0` is released: faster diffusion RL powered by vLLM-Omni (request-level/step-wise batching with FA3), rebuilt Qwen3-Omni multimodal training (DPO & GSPO), plus LTX-2.3, Qwen-Image-Edit support and more. See the [release notes](https://github.com/verl-project/verl-omni/releases/tag/v0.2.0).
 - [2026/08] We released [0.26.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.26.0) - aligned with the vLLM 0.26 release line, featuring [MiniMax H3](recipes/MiniMaxAI/MiniMax-H3.md) joint video/audio generation, an experimental full-duplex realtime runtime for [MiniCPM-o 4.5](recipes/OpenBMB/MiniCPM-o-4_5.md), distributed layerwise diffusion offload, and broader model, hardware, streaming, TTS, and quantization support.
 - [2026/07] We released [0.24.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.24.0) - aligned with the vLLM 0.24 release line, expanding production-ready coverage across TTS, speech, diffusion, image/video generation, and robot-policy serving, with major Omni stage runtime refactoring, diffusion request-level batching, async output materialization, quantization/cache/memory improvements, and broad CUDA/ROCm/XPU/NPU support.
@@ -57,9 +58,9 @@ vLLM-Omni is flexible and easy to use with:
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:

 - **Omni-modality models** (e.g. Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL)
- **TTS models** (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
- **Diffusion models** — image, video, and audio generation (e.g. MiniMax H3, Qwen-Image, Wan2.2, FLUX)
- **Robot-policy and action models** (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)
+- **TTS models** (e.g. Qwen3-TTS, IndexTTS 2.5, dots.tts, CosyVoice3)
+- **Diffusion models** — image, video, and audio generation (e.g. MiniMax H3, LTX-2.5, SANA-Video, Wan2.2)
+- **Robot-policy and action models** (e.g. π0, GR00T-N1.7, DreamZero-DROID, InternVLA-A1)

 ## Getting Started

@@ -67,6 +68,7 @@ Visit our [documentation](https://vllm-omni.readthedocs.io/en/latest/) to learn

 - [Installation](https://vllm-omni.readthedocs.io/en/latest/getting_started/installation/)
 - [Quickstart](https://vllm-omni.readthedocs.io/en/latest/getting_started/quickstart/)
+- [Nightly CUDA Docker images](https://hub.docker.com/r/vllm/vllm-omni/tags?name=nightly) built from the latest `main`
 - [List of Supported Models](https://vllm-omni.readthedocs.io/en/latest/models/supported_models/)
 - [Deployment Recipes](https://recipes.vllm.ai) for vLLM-Omni model serving

@@ -5,14 +5,22 @@ nav:
    - getting_started/quickstart.md
    - getting_started/installation/*
  - Serving:
-    - OpenAI-Compatible API:
-      - Chat Completions API: serving/chat_completions_api.md
-      - Image Generation: serving/image_generation_api.md
-      - Image Edit: serving/image_edit_api.md
-      - Text to Speech: serving/speech_api.md
-      - Videos: serving/videos_api.md
-      - Audio Generation: serving/audio_generate_api.md
-      - Streaming Video Input: serving/video_stream_api.md
+    - API Server:
+      - Overview: serving/README.md
+      - Core APIs:
+        - Chat Completions: serving/chat_completions_api.md
+        - Text to Speech: serving/speech_api.md
+        - Audio Generation: serving/audio_generate_api.md
+        - Image Generation: serving/image_generation_api.md
+        - Image Edit: serving/image_edit_api.md
+        - Videos: serving/videos_api.md
+      - Streaming and Realtime:
+        - Realtime Audio: serving/realtime_api.md
+        - Full Duplex: serving/full_duplex_api.md
+        - Streaming Video Input: serving/video_stream_api.md
+        - Streaming Video Output: serving/streaming_video_output_api.md
+        - OpenPI Robot Policy: serving/openpi_api.md
+      - Standalone Experimental Servers: serving/standalone_servers.md
  - Examples:
    - examples/README.md
    - Offline Inference:
@@ -29,7 +37,7 @@ nav:
      - Image-To-Video: user_guide/examples/online_serving/image_to_video.md
      - Speech-To-Video: user_guide/examples/online_serving/speech_to_video.md
      - Text-To-Image: user_guide/examples/online_serving/text_to_image.md
-      - Text-To-Speech: user_guide/examples/online_serving/text_to_speech.md
+      - Text-To-Speech (Online Serving): user_guide/examples/online_serving/text_to_speech.md
      - Text-To-Video: user_guide/examples/online_serving/text_to_video.md
  - General:
    - usage/*
@@ -119,6 +127,8 @@ nav:
        - Async Omni Output Materialization: design/feature/omni_async_output_materialization.md
        - Automatic Prefix Caching: design/feature/prefix_caching.md
        - Realtime AR-Diffusion: design/feature/realtime_ar_diffusion.md
+      - Model-Specific Capabilities:
+        - MiniMax H3 Temporal VAE Chunks: design/feature/minimax_h3_temporal_vae_chunks.md
      - Communication:
        - OmniConnector Implementations:
          - Mooncake Store Connector: design/feature/omni_connectors/mooncake_store_connector.md
@@ -57,9 +57,9 @@ vLLM-Omni is flexible and easy to use with:
 vLLM-Omni seamlessly supports most popular open-source models on HuggingFace, including:

 - Omni-modality models (e.g. Qwen3-Omni, MiniCPM-o 4.5, Cosmos3, HunyuanImage, BAGEL)
- TTS models (e.g. Qwen3-TTS, VoxCPM2, Ming-Omni-TTS, CosyVoice3)
- Diffusion models — image, video, and audio generation (e.g. MiniMax H3, Qwen-Image, Wan2.2, FLUX)
- Robot-policy and action models (e.g. GR00T-N1.7, DreamZero-DROID, InternVLA-A1, Cosmos3 action policy)
+- TTS models (e.g. Qwen3-TTS, IndexTTS 2.5, dots.tts, CosyVoice3)
+- Diffusion models — image, video, and audio generation (e.g. MiniMax H3, LTX-2.5, SANA-Video, Wan2.2)
+- Robot-policy and action models (e.g. π0, GR00T-N1.7, DreamZero-DROID, InternVLA-A1)

 For more information, checkout the following:

@@ -2,7 +2,7 @@

 This section lists the most common options for running vLLM-Omni.

-For options within a vLLM Engine, please refer to the [vLLM 0.26 configuration guide](https://docs.vllm.ai/en/v0.26.0/configuration/index.html).
+For options within a vLLM Engine, please refer to the [vLLM 0.28 configuration guide](https://docs.vllm.ai/en/v0.28.0/configuration/index.html).

 Each model defines fixed topology in a registered `PipelineConfig` and runtime
 overrides in a deploy YAML.
@@ -151,7 +151,7 @@ their keys only.
 ## Inherited vLLM variables

 vLLM-Omni also reads variables through its aligned vLLM dependency. Refer to
-the [vLLM 0.26 environment-variable reference](https://docs.vllm.ai/en/v0.26.0/configuration/env_vars.html)
+the [vLLM 0.28 environment-variable reference](https://docs.vllm.ai/en/v0.28.0/configuration/env_vars.html)
 for their definitions. This includes vLLM launch, cache, logging, plugin, ROCm,
 XPU, ModelScope, and FlashInfer workspace settings.

@@ -170,6 +170,10 @@ because the user-visible output is diffusion media rather than a token stream.
 The
 [MiniMax-H3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/MiniMaxAI/MiniMax-H3.md)
 contains the hardware-specific profiles and warmup requirements.
+The internal, opt-in
+[temporal VAE chunk callback](feature/minimax_h3_temporal_vae_chunks.md)
+documents the narrower released-source contract for publishing committed video
+frames during decode. It does not change the complete-output serving topology.

 ### Cosmos3: unified MoT reasoner and generator

@@ -0,0 +1,260 @@
+# MiniMax H3 temporal VAE chunk callback
+
+!!! warning "Experimental internal capability"
+
+    This feature is an opt-in model integration seam. The current MiniMax H3
+    pipeline still uses complete-output VAE decode, and no CLI option, serving
+    path, transport, or MP4 sink enables temporal chunks by default.
+
+## Status and scope
+
+MiniMax H3 decodes video latents in temporal windows, but the released VAE
+normally exposes only the complete reconstructed video. The temporal callback
+publishes frames after the released decoder has completed overlap blending and
+padding removal. This gives a future runtime consumer a stable point at which
+it can schedule representation conversion, D2H transfer, and encoding while
+the VAE continues decoding later windows.
+
+The implementation remains model-local:
+
+- `vae.py` owns VAE topology and the thin
+  `decode_latent_with_chunks()` facade;
+- `chunked_decode.py` owns the normalized callback contract, metadata
+  validation, and fixed VAE-group rendezvous points; and
+- `temporal_chunks.py` owns the source-gated compatibility planner and the
+  released temporal assembly control flow.
+
+Crop policy, typed media, transport, backpressure, encoding, and public output
+remain runtime responsibilities. The model code does not schedule requests or
+select a presentation format.
+
+## Execution flow
+
+The default and opt-in paths are intentionally separate:
+
+```text
+Default:
+latent -> MiniMaxH3VideoVAE.decode_latent() -> complete RGB video
+
+Opt-in producer seam:
+latent
+  -> MiniMaxH3VideoVAE.decode_latent_with_chunks()
+  -> group-local preflight and callback-owner validation
+  -> released temporal decode / overlap blending / padding removal
+  -> borrowed decoder-domain chunk
+  -> normalize + clone into owned RGB float32 BCTHW [0, 1]
+  -> synchronous callback
+  -> complete RGB video is still returned
+```
+
+Calling `decode_latent_with_chunks()` is the opt-in. Supplying no callback on a
+VAE peer does not select the default decoder; every rank in a VAE patch group
+must enter the chunked method, and only the output rank receives the callback.
+
+## Producer contract
+
+The normalized callback has this structural signature:
+
+```python
+def callback(
+    frames: torch.Tensor,
+    /,
+    *,
+    chunk_index: int,
+    total_chunks: int,
+    frame_start: int,
+    is_final: bool,
+) -> None:
+    ...
+```
+
+For a successful decode:
+
+- `frames` is owned, contiguous RGB float32 in `BCTHW` layout and `[0, 1]`;
+- callback storage does not alias the returned complete video;
+- the consumer may retain or mutate a chunk;
+- `chunk_index` is zero-based, contiguous, and strictly increasing;
+- `total_chunks` is positive and stable for the decode;
+- `frame_start` equals the end of the previous published range;
+- chunks are nonempty and contain only temporally committed frames;
+- exactly one chunk has `is_final=True`; and
+- concatenating chunks along time exactly reconstructs the returned video.
+
+The callback runs synchronously on the producer's current device stream. A
+consumer that reads the tensor on another stream must establish ordering, for
+example with an event, and retain the allocation for that stream's lifetime.
+A slow callback also blocks VAE production. An overlap-oriented runtime
+consumer should therefore acquire bounded capacity, enqueue side-stream work,
+and return promptly instead of encoding frames inside the callback.
+
+Chunks retain the decoded VAE canvas dimensions. A future runtime integration
+must apply the same requested-height and requested-width crop as the complete
+MiniMax H3 pipeline before declaring the media transport-ready.
+
+The raw compatibility callback is deliberately narrower. It receives borrowed
+decoder-domain tensors and must not retain or mutate them. The coordinator is
+the ownership boundary that normalizes and clones raw chunks before invoking
+the public callback.
+
+## Temporal planning and terminal publication
+
+The released configuration advances five latent tokens per temporal window and
+includes a two-token overlap, so each `_adaptive_decode` input normally contains
+seven tokens. After temporal scaling, pre-padding removal, and five-frame
+overlap blending, the normal committed unit is 17 output frames. The final
+remainder depends on the requested frame plan.
+
+Representative released frame plans include:
+
+| Output frames | Published chunks |
+| ---: | --- |
+| 124 | `7 x 17 + 5` |
+| 209 | `12 x 17 + 5` |
+| 243 | `14 x 17 + 5` |
+| 362 | `21 x 17 + 5` |
+
+The terminal callback is held until the complete frame plan passes its logical
+frame, padding, written-frame, chunk-count, and finality checks. Earlier chunks
+are semantically stable, but they do not imply that the overall request has
+succeeded; a consumer must discard or abort partial output if a later decoder
+failure terminates the request.
+
+## Source and configuration gate
+
+The released VAE has no public temporal-callback API. vLLM-Omni therefore owns
+a compatibility implementation for one exact released code contract. The
+reference revision identifies its provenance, but runtime acceptance is based
+on source content, configuration, and structure rather than repository or
+weight metadata:
+
+- reference Hugging Face revision, not independently checked at runtime:
+  `42ed227ee7df40d41602854ae760620d6eb651fe`;
+- sorted 15-file top-level `video_vae/*.py` manifest (excluding
+  `__init__.py`) SHA-256:
+  `30e64afbaa940696bf8bfe2c86003a4b1666c84b22fa3aa12403a3e6a03f705d`;
+- clip length `17`;
+- token chunk, overlap, and drop `5`, `2`, and `3`;
+- temporal ratio `4`;
+- frame pre-padding and overlap `3` and `5`; and
+- no isolated first or last frame.
+
+The adapter does not verify Hugging Face revision metadata, repository identity,
+or model weights. The manifest pin detects Python compatibility drift; it is
+not a security boundary for remote-code execution. The source manifest, listed
+temporal attributes, structural contract, plan, and dtype are validated before
+the first `_adaptive_decode` call. A mismatch disables only the opt-in chunk path.
+`decode_latent()` remains available and performs no source fingerprint or chunk
+rendezvous.
+
+## Supported models
+
+Support in this table means the internal producer contract above, not public
+serving or end-to-end transport support.
+
+| Model or component | Source contract | Status | Validation |
+| --- | --- | --- | --- |
+| MiniMax H3 FL2VA video VAE | Released manifest/listed temporal attributes above (reference `42ed227e`) | Supported, opt-in | Same-environment full-output parity and L20X/B300 feasibility measured on the implementation through PR commit `60e9bbad` |
+| MiniMax H3 Ref2VA video VAE | Same released manifest/listed temporal attributes | Source-compatible; not separately benchmarked | Ref2VA and FL2VA use byte-identical video-VAE source and configuration in the reference snapshot |
+| MiniMax H3 with a different Python manifest or listed temporal attribute | Different enforced contract | Unsupported for chunk mode | Fails closed before decoder collectives; complete decode remains available |
+| Same manifest/temporal attributes with different weights or other component metadata | Not checked by this gate | Not independently qualified | The adapter does not inspect weights, repository identity, revision metadata, or arbitrary component configuration |
+| Other video models | Any | Unsupported | No generic temporal VAE producer ABI is declared by this feature |
+
+## Feature compatibility
+
+| Dimension or combination | Status | Constraint |
+| --- | --- | --- |
+| Existing complete-output decode | Supported and default | No source check, callback object, or callback rendezvous |
+| Explicit chunk decode | Supported internally | Caller invokes `decode_latent_with_chunks()` directly |
+| VAE patch parallel size 1 | Supported by the model seam | Caller supplies the callback only on the request output owner; no WORLD rendezvous is added |
+| VAE patch parallel size equal to the full DiT group | Supported when data parallel size is 1 | Every VAE-group rank calls; exactly group rank 0 owns the callback |
+| Partial VAE patch subgroup | Unsupported | MiniMax H3 currently accepts size 1 or the full DiT group only |
+| Data parallel size greater than 1 with VAE patch parallel size 1 | Model seam only | Per-replica callback ownership is possible, but no serving/runtime consumer is wired |
+| Data parallel size greater than 1 with VAE patch parallel size greater than 1 | Unsupported | Fails before decoder collectives because no replica-local VAE patch group is defined |
+| Too few spatial tiles for the VAE group | Correctness fallback | Decodes rank-locally and restores native tiling state; slower than patch parallel decode |
+| Device memory and full-output assembly | Not reduced or bounded | The compatibility path still assembles and returns the full decode while allocating owned callback chunks; no peak-HBM reduction is claimed |
+| DiT cache, attention, TP, SP, HSDP, LoRA, or quantization features | No direct callback dependency | This seam runs after denoising and does not independently qualify combinations beyond the MiniMax H3 model's own matrix |
+| Layerwise or distributed component offload | Not separately validated | Component residency must place the video VAE before decode |
+| Step execution or request batching | Not integrated | No pipeline caller binds a runtime chunk consumer |
+| Requested-size crop and device uint8 preparation | Not implemented here | Runtime-owned follow-up |
+| Pinned D2H, SHM/ZMQ, or cross-process transport | Not implemented here | Runtime-owned follow-up |
+| Complete MP4 or fragmented MP4 sink | Not implemented here | Existing serving behavior still waits for complete output |
+| Audio VAE chunks and incremental A/V composition | Not supported | Video chunks do not imply audio or request finality |
+| Cancellation and multi-output lease ownership | Not implemented here | Requires a bounded runtime lifecycle |
+| NVIDIA CUDA | Feasibility evidence | L20X and B300 measurements apply to the implementation through `60e9bbad`; the current coordination refactor was not reprofiled |
+| AMD ROCm | Not validated for this feature | Base MiniMax H3 support is documented separately and does not imply temporal-chunk validation |
+| Ascend, Intel, or other accelerators | No feature support claim | Consult the model support table independently; this design adds no platform qualification |
+
+## Distributed and failure semantics
+
+When VAE patch parallelism is active, all VAE-group ranks run the same
+preflight before entering checkpoint collectives. Source/planner errors and the
+resolved temporal concat dtype are reconciled at fixed group rendezvous points.
+
+Callback ownership is also a group decision. Rank `r` contributes `r + 1` when
+it owns a callback and zero otherwise. A group SUM equals one only when rank 0
+is the sole owner; every invalid owner configuration therefore fails
+identically on all ranks before decode.
+
+Callback exceptions are retained on the output rank. Further publication stops,
+but every rank completes the same remaining native decoder collective schedule.
+The original exception is then raised on the owner and
+`MiniMaxH3ChunkCallbackPeerError` on peers.
+
+Exceptions raised inside the checkpoint's native distributed decode cannot in
+general be reconciled by the adapter: a peer may already be blocked inside a
+checkpoint-owned collective. Resolving such failures requires native collective
+handling or supervisor-level timeout and worker-group teardown. The process
+group must not be assumed reusable after such a failure.
+
+## Validation evidence
+
+The contract suite covers ordinary and padded frame plans, blend boundaries,
+owned storage and mutation isolation, source-manifest/listed-temporal-attribute
+drift, callback failure, and healthy subsequent requests. Real Gloo processes cover
+rank-asymmetric preflight failures, dtype disagreement, two-rank callback
+failure propagation, three-rank wrong-owner rejection, insufficient-tile
+fallback, and group recovery for failures that occur at adapter-owned
+rendezvous points.
+
+Four-device L20X and independent four- and eight-device B300 experiments on the
+implementation through PR commit `60e9bbad` showed exact serial-versus-chunked
+frame/MP4 parity within each environment and lower VAE-to-complete-MP4 latency
+when an experimental consumer was attached. Hashes are not claimed to match
+across different hardware/software environments. The later coordination and
+module-boundary refactors were revalidated with CPU and real Gloo tests but were
+not reprofiled on GPUs. These are feasibility measurements, not a promise for
+default serving, because this change does not contain the transport or encoder
+consumer. See [RFC #6872](https://github.com/vllm-project/vllm-omni/issues/6872)
+and [PR #6885](https://github.com/vllm-project/vllm-omni/pull/6885) for the
+measurement provenance and limitations.
+
+## Generalization boundary and follow-ups
+
+This feature does not define a repository-wide `SupportsVideoChunkDecode`
+interface. H3 temporal blending, Wan causal caches, LTX temporal/spatial halos,
+and pipeline-level generation chunks have different commit and failure
+boundaries. A generic producer contract should be extracted only after a second
+model proves the same semantic event without `Any` arguments or model-name
+dispatch.
+
+The reusable future boundary is the emitted, committed video-media chunk, not a
+universal latent-to-VAE method. It should extend the typed media work in
+[RFC #6541](https://github.com/vllm-project/vllm-omni/issues/6541) and
+[PR #6615](https://github.com/vllm-project/vllm-omni/pull/6615), while keeping
+model-specific halo, cache, padding, blending, distributed, and source-gating
+logic inside each model adapter.
+
+Runtime follow-ups own:
+
+- typed video-plus-audio media and request/output identity;
+- requested-size crop and device preparation;
+- bounded leases, backpressure, cancellation, and cleanup;
+- D2H/IPC transport and observability; and
+- complete-MP4 and fragmented-MP4 consumers.
+
+## Related documentation
+
+- [Diffusion model integration](../module/diffusion/diffusion_model_integration.md)
+- [VAE patch parallelism](vae_parallel.md)
+- [Async diffusion output](async_diffusion_output.md)
+- [MiniMax H3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/MiniMaxAI/MiniMax-H3.md)
@@ -25,6 +25,10 @@ implementation contract; it is not, by itself, a general support claim.
 - [Automatic Prefix Caching in Omni Models](feature/prefix_caching.md)
 - [Realtime AR-Diffusion Sessions](feature/realtime_ar_diffusion.md)

+### Model-specific capabilities
+
+- [MiniMax H3 Temporal VAE Chunks](feature/minimax_h3_temporal_vae_chunks.md)
+
 ### Communication

 #### OmniConnector implementations
@@ -57,3 +57,10 @@ cross-stage routing.

 Test registry selection, checkpoint loading, minimal inference, input and output
 contracts, and every declared optional capability.
+
+## Model-specific capability documents
+
+- [MiniMax H3 temporal VAE chunk callback](../../feature/minimax_h3_temporal_vae_chunks.md)
+  defines the source-gated producer seam and its boundary with future runtime
+  media transport. It is an internal opt-in capability, not a public serving
+  feature.
@@ -3,4 +3,6 @@
 vLLM-Omni's examples are split into two categories:

 - If you are using vLLM-Omni from within Python code, see the *Offline Inference* section.
- If you are using vLLM-Omni from an HTTP application or client, see the *Online Serving* section.
+- If you are using vLLM-Omni from an HTTP application or client, start with
+  the [API Server guide](../serving/README.md) to choose an endpoint, then use
+  the *Online Serving* section for model-specific examples.
@@ -34,7 +34,7 @@ vLLM-Omni is a Python library that supports the following GPU variants. The libr

 ### Pre-built wheels

-Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, 0.23.0, 0.24.0, 0.25.0, and 0.26.0; the 0.28.0 wheels are published together with the v0.28.0 release this branch produces. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).
+Note: Pre-built wheels are currently available for vLLM-Omni 0.11.0rc1, 0.12.0rc1, 0.14.0rc1, 0.14.0, 0.16.0, 0.18.0, 0.20.0, 0.21.0, 0.22.0, 0.23.0, 0.24.0, 0.25.0, 0.26.0, and 0.28.0. If you need a newer unreleased revision, please [build from source](https://docs.vllm.ai/projects/vllm-omni/en/latest/getting_started/installation/gpu/#build-wheel-from-source).

 === "NVIDIA CUDA"

@@ -5,7 +5,7 @@
 # --8<-- [end:requirements]
 # --8<-- [start:set-up-using-python]

-vLLM-Omni depends on the matching major/minor release of vLLM. The vLLM-Omni 0.26.x release line uses vLLM 0.26.x.
+vLLM-Omni depends on the matching major/minor release of vLLM. The vLLM-Omni 0.28.x release line uses vLLM 0.28.x.

 !!! note
    PyTorch installed via `conda` will statically link `NCCL` library, which can cause issues when vLLM tries to use `NCCL`. See <gh-issue:8420> for more details.
@@ -95,6 +95,13 @@ uv pip install --no-build-isolation --editable .

 vLLM-Omni offers an official docker image for deployment. These images are built on top of vLLM docker images and available on Docker Hub as [vllm/vllm-omni](https://hub.docker.com/r/vllm/vllm-omni/tags). The version of vLLM-Omni indicates which release of vLLM it is based on.

+Nightly CUDA images built from the latest `main` are also available:
+
+- [`vllm/vllm-omni:nightly`](https://hub.docker.com/r/vllm/vllm-omni/tags?name=nightly) is the rolling multi-architecture tag.
+- [`vllm/vllm-omni:nightly-<commit>`](https://hub.docker.com/r/vllm/vllm-omni/tags?name=nightly-) provides commit-pinned multi-architecture tags. The newest 14 are retained.
+
+Use a commit-pinned tag when you need a reproducible nightly environment.
+
 Here's an example deployment command that has been verified on 2 x H100's:
 ```bash
 docker run --runtime nvidia --gpus 2 \
@@ -9,8 +9,8 @@ The recommended way to use vLLM-Omni on NPU is through the vLLM-Ascend pre-built

 ```bash
 # vLLM-Ascend image
-# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.26.0
-# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.26.0-a3
+# Atlas A2: quay.io/atlas-ci/vllm-ascend:v0.28.0
+# Atlas A3: quay.io/atlas-ci/vllm-ascend:v0.28.0-a3
 docker run --rm \
    --name vllm-omni-npu \
    --shm-size=1g \
@@ -28,11 +28,11 @@ docker run --rm \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v /root/.cache:/root/.cache \
    -p 8000:8000 \
-    -it quay.io/atlas-ci/vllm-ascend:v0.26.0-a3 bash
+    -it quay.io/atlas-ci/vllm-ascend:v0.28.0-a3 bash

 # Inside the container, install vLLM-Omni from source
 cd /vllm-workspace
-git clone -b v0.26.0 https://github.com/vllm-project/vllm-omni.git
+git clone -b v0.28.0 https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 pip install -v -e . --no-build-isolation
 # or VLLM_OMNI_TARGET_DEVICE=npu pip install -v -e .
@@ -50,21 +50,11 @@ We are keeping [issue #886](https://github.com/vllm-project/vllm-omni/issues/886

 # --8<-- [start:installation-main]

-You can also build vLLM-Omni from the latest main branch if you want to use the latest features or bug fixes. (But sometimes it will break for a while. You can check [issue #886](https://github.com/vllm-project/vllm-omni/issues/886) for the status of the latest commit of vLLM-Omni main branch on NPU.)
+You can also install vLLM-Omni from the latest main branch if you want to use the latest features or bug fixes. Use the aligned vLLM-Ascend v0.28.0 image described above, and check [issue #886](https://github.com/vllm-project/vllm-omni/issues/886) for the status of the latest vLLM-Omni commit on NPU.

 ```bash
-# Pin vLLM and vLLM-Ascend to the v0.26 release line
-git clone -b v0.26.0 https://github.com/vllm-project/vllm.git
-cd vllm
-VLLM_TARGET_DEVICE=empty pip install -v -e .
-cd ..
-
-git clone -b releases/v0.26.0rc https://github.com/vllm-project/vllm-ascend.git
-cd vllm-ascend
-pip install -v -e .
-cd ..
-
-# Install vLLM-Omni from the latest main branch
+# Inside the vLLM-Ascend v0.28.0 container, install vLLM-Omni from main
+cd /vllm-workspace
 git clone https://github.com/vllm-project/vllm-omni.git
 cd vllm-omni
 pip install -v -e . --no-build-isolation
@@ -88,8 +78,8 @@ Supported images as following.
 Here's an example deployment command that has been verified on 4 x NPUs:

 ```bash
-# Atlas A2: quay.io/ascend/vllm-omni:v0.26.0
-# Atlas A3: quay.io/ascend/vllm-omni:v0.26.0-a3
+# Atlas A2: quay.io/ascend/vllm-omni:v0.28.0
+# Atlas A3: quay.io/ascend/vllm-omni:v0.28.0-a3
 docker run --rm \
    --name vllm-omni-a3 \
    --shm-size=4g \
@@ -114,7 +104,7 @@ docker run --rm \
    -v /etc/ascend_install.info:/etc/ascend_install.info \
    -v ~/.cache:/root/.cache \
    -p 8091:8091 \
-    -it quay.io/ascend/vllm-omni:v0.26.0-a3 bash
+    -it quay.io/ascend/vllm-omni:v0.28.0-a3 bash
 ```

 !!! tip
@@ -91,7 +91,7 @@ if __name__ == "__main__":

 For more usages, please refer to [offline inference](../user_guide/examples/offline_inference/qwen2_5_omni.md)

-## Online Serving with OpenAI-Completions API
+## Online Serving with the API Server

 Text-to-image generation quickstart with vLLM-Omni:

@@ -100,20 +100,16 @@ vllm serve Tongyi-MAI/Z-Image-Turbo --omni --port 8091
 ```

 ```bash
-curl -s http://localhost:8091/v1/chat/completions \
+curl -s http://localhost:8091/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
-    "messages": [
-      {"role": "user", "content": "a cup of coffee on the table"}
-    ],
-    "extra_body": {
-      "height": 1024,
-      "width": 1024,
-      "num_inference_steps": 50,
-      "guidance_scale": 4.0,
-      "seed": 42
-    }
-  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2 | base64 -d > coffee.png
+    "prompt": "a cup of coffee on the table",
+    "size": "1024x1024",
+    "response_format": "b64_json",
+    "seed": 42
+  }' | jq -r '.data[0].b64_json' | base64 -d > coffee.png
 ```

-For more details, please refer to [online serving](../user_guide/examples/online_serving/text_to_image.md).
+See the [API Server guide](../serving/README.md) to choose an endpoint. For
+model-specific details, refer to the [text-to-image online serving
+example](../user_guide/examples/online_serving/text_to_image.md).
@@ -21,7 +21,7 @@ th {
 |--------------|--------|-------------------|------------|---------|-----|-----------|--------|
 | `Qwen3OmniMoeForConditionalGeneration` | Qwen3-Omni | `Qwen/Qwen3-Omni-30B-A3B-Instruct` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Qwen/Qwen3-Omni.md) |
 | `Qwen2_5OmniForConditionalGeneration` | Qwen2.5-Omni | `Qwen/Qwen2.5-Omni-7B`, `Qwen/Qwen2.5-Omni-3B` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
-| `MingFlashOmniForConditionalGeneration` + `MingImagePipeline` | Ming-flash-omni-2.0 (omni-speech + imagegen<sup>1</sup>) | `Jonathan1909/Ming-flash-omni-2.0` | ✅︎ |   |   |   | — |
+| `MingFlashOmniForConditionalGeneration` + `MingImagePipeline` | Ming-flash-omni-2.0 (omni-speech + imagegen<sup>1</sup>) | `Jonathan1909/Ming-flash-omni-2.0` | ✅︎ | | | | — |
 | `BagelForConditionalGeneration` | BAGEL (DiT-only) | `ByteDance-Seed/BAGEL-7B-MoT` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `InternVLAA1Pipeline` | InternVLA-A1 | `InternRobotics/InternVLA-A1-3B` | ✅︎ | ✅︎ | | | — |
 | `Gr00tN1d7Pipeline` | GR00T N1.7 | `nvidia/GR00T-N1.7-3B` | ✅︎ | | | | — |
@@ -62,7 +62,7 @@ th {
 | `LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatVideoAvatarPipeline` | LongCat-Video-Avatar-1.5 A2V/AI2V (native single-speaker and multi-speaker AI2V/AVC) | `meituan-longcat/LongCat-Video-Avatar-1.5` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/meituan-longcat/LongCat-Video-Avatar-1.5.md) |
-| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit` | ✅︎ | | | ✅︎ | — |
+| `BooguImagePipeline` | Boogu-Image | `Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit`, `Boogu/Boogu-Image-0.1-Edit-Turbo` | ✅︎ | | | ✅︎ | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Boogu/Boogu-Image.md) |
 | `StableDiffusionXLPipeline` | Stable-Diffusion-XL | `stabilityai/stable-diffusion-xl-base-1.0` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `StableDiffusion3Pipeline` | Stable-Diffusion-3 | `stabilityai/stable-diffusion-3.5-medium` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `CosyVoice3Model` | CosyVoice3 | `FunAudioLLM/Fun-CosyVoice3-0.5B-2512` | ✅︎ | ✅︎ | | ✅︎ | — |
@@ -95,7 +95,7 @@ th {
 | `MiMoV2ASRForCausalLM` | MiMo-V2.5-ASR | `XiaomiMiMo/MiMo-V2.5-ASR` | ✅︎ | ✅︎ | | | — |
 | `Flux2Pipeline` | FLUX.2-dev | `black-forest-labs/FLUX.2-dev` | ✅︎ | ✅︎ | | | — |
 | `FishSpeechSlowARForConditionalGeneration` | Fish Speech S2 Pro | `fishaudio/s2-pro` | ✅︎ | ✅︎ | | ✅︎ | — |
-| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | | — |
+| `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only), SenseNova-U1.5 | `SenseNova/SenseNova-U1-8B-MoT`, `sensenova/SenseNova-U1.5-8B-MoT` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/SenseNova/SenseNova-U1.5.md) |
 | `LancePipeline` | Lance | `bytedance-research/Lance` | ✅︎ | | | | — |
 | `HunyuanVideo15Pipeline` | HunyuanVideo-1.5-T2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | ✅︎ | ✅︎ | | | — |
 | `HunyuanVideo15ImageToVideoPipeline` | HunyuanVideo-1.5-I2V | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v`, `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_i2v` | ✅︎ | ✅︎ | | | — |
@@ -0,0 +1,187 @@
+# API Server
+
+vLLM-Omni exposes OpenAI-compatible HTTP APIs plus vLLM-Omni extensions for
+image, audio, video, realtime, and robot-policy workloads. Use this page to
+choose an endpoint; follow the linked reference page for model-specific fields,
+wire protocols, and response details.
+
+Each server instance hosts one model. An endpoint is usable only when the
+loaded model supports that task.
+
+This guide focuses on vLLM-Omni's primary task endpoints. Compatible models
+may also expose standard routes inherited from vLLM, such as
+`POST /v1/completions`; those routes remain model- and task-dependent.
+
+## Start and Verify the Server
+
+Start a model with the shared serving command:
+
+```bash
+vllm serve <model> --omni --port 8091
+```
+
+Some models require additional flags or a deployment configuration. Use the
+model-specific guide when one is provided.
+
+After the server starts, check its health and served model name:
+
+```bash
+export VLLM_OMNI_BASE_URL=http://localhost:8091
+
+curl "$VLLM_OMNI_BASE_URL/health"
+curl "$VLLM_OMNI_BASE_URL/v1/models" | jq .
+```
+
+Use `http://localhost:8091/v1` as the `base_url` for an OpenAI SDK client. If
+the server was started with `--api-key`, include an `Authorization` header with
+`Bearer <api-key>` in requests.
+
+## Choose a Core HTTP Endpoint
+
+Prefer a task-specific endpoint when one matches your workload. Use Chat
+Completions for conversational or heterogeneous omni pipelines, rather than as
+the default wrapper for every generation task.
+
+| Task | Endpoint | Request | Response | Details |
+|------|----------|---------|----------|---------|
+| Conversation or multimodal understanding/generation | `POST /v1/chat/completions` | JSON | OpenAI-style JSON or SSE | [Chat Completions](chat_completions_api.md) |
+| Text-to-speech | `POST /v1/audio/speech` | JSON | Audio bytes or SSE | [Speech](speech_api.md) |
+| Sound, music, or ambient audio generation | `POST /v1/audio/generate` | JSON | Audio bytes | [Audio Generation](audio_generate_api.md) |
+| Text-to-image | `POST /v1/images/generations` | JSON | Base64 JSON or an image file | [Image Generation](image_generation_api.md) |
+| Image editing | `POST /v1/images/edits` | Multipart form | Base64 JSON or SSE | [Image Edit](image_edit_api.md) |
+| Video generation (recommended) | `POST /v1/videos` | Multipart form | Asynchronous job JSON | [Videos](videos_api.md) |
+| Video generation (blocking) | `POST /v1/videos/sync` | Multipart form | Video bytes | [Videos](videos_api.md#synchronous-response) |
+
+`POST /v1/videos/sync` is intended for tests and simple scripts. Use the
+asynchronous job API for production requests so clients can poll, download,
+and delete generated videos independently.
+
+## Minimal Requests
+
+The examples below assume a compatible model is already running. TTS voices,
+media inputs, generation controls, and output capabilities vary by model.
+
+=== "Chat"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/chat/completions" \
+      -H "Content-Type: application/json" \
+      -d '{
+        "messages": [{"role": "user", "content": "Describe vLLM-Omni briefly."}],
+        "modalities": ["text"]
+      }'
+    ```
+
+=== "Speech"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/audio/speech" \
+      -H "Content-Type: application/json" \
+      -d '{
+        "input": "Hello from vLLM-Omni.",
+        "voice": "vivian"
+      }' \
+      --output speech.wav
+    ```
+
+=== "Audio"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/audio/generate" \
+      -H "Content-Type: application/json" \
+      -d '{
+        "input": "Ocean waves on a quiet beach",
+        "audio_length": 5.0
+      }' \
+      --output audio.wav
+    ```
+
+=== "Image"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/images/generations" \
+      -H "Content-Type: application/json" \
+      -d '{
+        "prompt": "A small robot reading beside a window",
+        "size": "1024x1024",
+        "response_format": "file"
+      }' \
+      --output image.png
+    ```
+
+=== "Image edit"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/images/edits" \
+      -F "image=@input.png" \
+      -F "prompt=Turn the background into a snowy mountain" \
+      | jq -r '.data[0].b64_json' \
+      | base64 --decode > edited.png
+    ```
+
+=== "Video"
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/videos" \
+      -F "prompt=A cinematic tracking shot of a mountain lake at sunrise" \
+      -F "size=1280x720" \
+      -F "seconds=5"
+    ```
+
+    The response contains a video job `id`. Poll and download it with:
+
+    ```bash
+    curl "$VLLM_OMNI_BASE_URL/v1/videos/<video_id>"
+    curl -L "$VLLM_OMNI_BASE_URL/v1/videos/<video_id>/content" --output video.mp4
+    ```
+
+## Choose a Streaming or Realtime Endpoint
+
+These WebSocket APIs have different event schemas and cannot be used
+interchangeably.
+
+| Workload | Endpoint | Interaction model | Details |
+|----------|----------|-------------------|---------|
+| Incremental text input for speech synthesis | `WS /v1/audio/speech/stream` | Send text events and receive audio | [Streaming Text to Speech](speech_api.md#streaming-text-input-websocket) |
+| Live video understanding | `WS /v1/video/chat/stream` | Send video frames and receive text/audio | [Streaming Video Input](video_stream_api.md) |
+| Turn-based realtime audio | `WS /v1/realtime` | Stream one audio input and receive transcript/audio events | [Realtime Audio](realtime_api.md) |
+| Continuous speech-to-speech interaction | `WS /v1/realtime?duplex=1` or `WS /v1/duplex` | Listen and speak concurrently with session control | [Full Duplex](full_duplex_api.md) |
+| Generated video chunks | `WS /v1/realtime/video` | Start a diffusion request and receive fragmented MP4 | [Streaming Video Output](streaming_video_output_api.md) |
+| Robot policy inference | `WS /v1/realtime/robot/openpi` | Send MessagePack observations and receive action arrays | [OpenPI Robot Policy](openpi_api.md) |
+
+All six routes are model- or configuration-dependent. In particular,
+`/v1/realtime` is not full duplex unless the client sets `duplex=1` and the
+deployment explicitly enables duplex sessions. Clients should also verify the
+duplex capability payload because the query-parameter form falls back to the
+ordinary realtime handler when duplex is unavailable.
+
+## Related Endpoints
+
+| Purpose | Endpoints | Reference |
+|---------|-----------|-----------|
+| Discovery and readiness | `GET /health`, `GET /v1/models` | This page |
+| Batched conversations | `POST /v1/chat/completions/batch` | [Batch requests](chat_completions_api.md#batch-requests) |
+| Batched speech | `POST /v1/audio/speech/batch` | [Batch speech generation](speech_api.md#batch-speech-generation) |
+| TTS voice management | `GET/POST /v1/audio/voices`, `DELETE /v1/audio/voices/{name}` | [Voices](speech_api.md#voices-endpoint) |
+| Video job lifecycle | `GET /v1/videos`, `GET/DELETE /v1/videos/{video_id}`, `GET /v1/videos/{video_id}/content` | [Video endpoints](videos_api.md#endpoints) |
+| Release and restore stage memory | `POST /v1/omni/sleep`, `POST /v1/omni/wakeup` | [Sleep Mode](../features/sleep_mode.md) |
+
+## Standalone Experimental Servers
+
+PersonaPlex also provides a Moshi-compatible WebSocket server, while JoyVL
+provides a stateful interaction orchestrator in front of another model server.
+These are separate processes with their own routes; they are not additional
+paths on every `vllm serve --omni` instance. See [Standalone Experimental
+Servers](standalone_servers.md) before deploying either one.
+
+## Compatibility Notes
+
+- A registered path does not mean every model supports it. Unsupported tasks
+  return an OpenAI-style error or a WebSocket `error` event.
+- Use JSON for Chat, Speech, Audio Generation, and Image Generation. Use
+  `multipart/form-data` when uploading images, videos, or other files.
+- Generation can take minutes for large diffusion models. Configure client
+  timeouts accordingly; prefer asynchronous video jobs over long blocking
+  requests.
+- Model-specific parameters are vLLM-Omni extensions. Check the endpoint
+  reference before assuming an OpenAI SDK exposes them directly.
@@ -4,14 +4,14 @@ vLLM-Omni provides an API for text-to-audio generation using diffusion-based mod

 Unlike the [Speech API](speech_api.md) which targets text-to-speech synthesis, the Audio Generate API is designed for general-purpose audio generation from text descriptions (sound effects, music, ambient soundscapes, etc.).

-Each server instance runs a single model (specified at startup via `vllm-omni serve <model> --omni`).
+Each server instance runs a single model (specified at startup via `vllm serve <model> --omni`).

 ## Quick Start

 ### Start the Server

 ```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
+vllm serve stabilityai/stable-audio-open-1.0 \
    --host 0.0.0.0 \
    --port 8091 \
    --gpu-memory-utilization 0.9 \
@@ -326,7 +326,7 @@ curl http://localhost:8091/health
 Enable debug logging:

 ```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
+vllm serve stabilityai/stable-audio-open-1.0 \
    --host 0.0.0.0 \
    --port 8091 \
    --gpu-memory-utilization 0.9 \
@@ -1,78 +1,110 @@
 # Chat Completions API

-vLLM-Omni supports generating and editing images via the `/v1/chat/completions`
-endpoint using diffusion models. This page explains how to pass generation
-parameters (such as `num_inference_steps`, `height`, `width`) to diffusion
-models through this endpoint.
+Use `/v1/chat/completions` for conversational and multimodal pipelines that
+follow the OpenAI Chat API schema. Depending on the loaded model, a request can
+contain text, image, audio, or video input and can return text, audio, images,
+or other model-specific output.

-!!! tip
-    For dedicated endpoints that accept generation parameters as top-level
-    fields, see [Image Generation API](image_generation_api.md) and
-    [Image Edit API](image_edit_api.md).
+Some diffusion pipelines also support image generation and editing through
+Chat Completions. Prefer the dedicated [Image Generation
+API](image_generation_api.md) or [Image Edit API](image_edit_api.md) for
+task-specific clients; their request fields and response contracts are more
+direct.

-## Passing Generation Parameters
-
-The `/v1/chat/completions` endpoint follows the OpenAI Chat API schema, which
-does not natively include diffusion-specific fields like `num_inference_steps`
-or `height`. How you pass these extra fields depends on your client.
-
-### curl / Python `requests`
-
-Wrap generation parameters inside an `"extra_body"` key in the JSON body:
+## Basic Request

 ```bash
-curl -s http://localhost:8091/v1/chat/completions \
+curl http://localhost:8091/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
-      {"role": "user", "content": "A beautiful landscape painting"}
+      {"role": "user", "content": "Describe vLLM-Omni briefly."}
    ],
-    "extra_body": {
-      "num_inference_steps": 50,
-      "seed": 42
-    }
+    "modalities": ["text"]
  }'
 ```

-### OpenAI Python SDK
+Use `stream: true` for Server-Sent Events. Input media syntax, supported output
+`modalities`, and response choices depend on the model; see the
+[model-specific online serving examples](../user_guide/examples/online_serving/qwen3_omni.md).

-Use the `extra_body` **keyword argument**. The SDK automatically merges these
-fields into the top-level request body:
+!!! tip
+    Each server hosts one model. Query `GET /v1/models` for its served model
+    name when your client requires the `model` field.

-```python
-response = client.chat.completions.create(
-    model="Qwen/Qwen-Image",
-    messages=[{"role": "user", "content": "A beautiful landscape painting"}],
-    extra_body={
-        "num_inference_steps": 50,
-        "seed": 42,
-    },
-)
+## vLLM Extension Parameters
+
+Send standard and vLLM-specific Chat parameters as top-level fields in a
+direct HTTP request. When using the OpenAI Python SDK, pass vLLM-specific
+fields through the SDK's `extra_body` keyword; the SDK merges them into the
+top-level JSON sent to the server.
+
+=== "curl"
+
+    ```bash
+    curl http://localhost:8091/v1/chat/completions \
+      -H "Content-Type: application/json" \
+      -d '{
+        "messages": [{"role": "user", "content": "Write a short haiku."}],
+        "top_k": 40
+      }'
+    ```
+
+=== "OpenAI Python SDK"
+
+    ```python
+    from openai import OpenAI
+
+    client = OpenAI(base_url="http://localhost:8091/v1", api_key="none")
+
+    response = client.chat.completions.create(
+        model="your-served-model",
+        messages=[{"role": "user", "content": "Write a short haiku."}],
+        extra_body={"top_k": 40},
+    )
+    ```
+
+??? note "Diffusion compatibility"
+
+    Diffusion models exposed through Chat Completions can accept supported
+    fields such as `num_inference_steps`, `seed`, `height`, and `width`. Send
+    them at the top level with direct HTTP, or in the SDK's `extra_body`
+    keyword argument.
+
+    A literal nested `"extra_body"` JSON object is accepted for compatibility,
+    but it is not recommended for new direct HTTP clients. Do not provide the
+    same parameter in multiple locations; the server returns a `400` error for
+    duplicate diffusion parameters. Prefer the dedicated [Image Generation
+    API](image_generation_api.md) or [Image Edit API](image_edit_api.md) when
+    either endpoint matches the task.
+
+## Batch Requests
+
+`POST /v1/chat/completions/batch` accepts the same shared generation fields,
+but `messages` is a list of conversations. The response contains one choice
+per conversation in input order.
+
+```bash
+curl http://localhost:8091/v1/chat/completions/batch \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      [{"role": "user", "content": "Summarize vLLM in one sentence."}],
+      [{"role": "user", "content": "Summarize vLLM-Omni in one sentence."}]
+    ],
+    "max_tokens": 64
+  }'
 ```

-!!! note "SDK `extra_body` vs. JSON `extra_body`"
-    These two `extra_body` usages look similar but work differently under the
-    hood. The SDK flattens the dict into the top-level request JSON, while the
-    curl/requests approach sends it as a nested `"extra_body"` key. Both are
-    handled correctly by the server.
-
-!!! note "About the `ignored fields` warning"
-    You may see a log message like:
-
-    ```
-    WARNING: The following fields were present in the request but ignored: {'height', 'width', ...}
-    ```
-
-    This is **harmless**. It is emitted by vLLM's request validation layer
-    because these fields are not part of the standard OpenAI
-    `ChatCompletionRequest` schema. The fields are still stored internally
-    and correctly forwarded to the diffusion pipeline.
+Batch Chat Completions does not support streaming, tools, beam search, or
+`n > 1`.

 ## Model-Specific Examples

-For complete examples with full request/response details, see the model-specific
-guides:
+For complete examples with model-specific inputs and outputs, see:

+- [Qwen3-Omni](../user_guide/examples/online_serving/qwen3_omni.md)
+- [Qwen2.5-Omni](../user_guide/examples/online_serving/qwen2_5_omni.md)
 - [Text-to-Image (Qwen-Image)](../user_guide/examples/online_serving/text_to_image.md)
 - [Image-to-Image (Qwen-Image-Edit, Qwen-Image-Layered)](../user_guide/examples/online_serving/image_to_image.md)
 - [GLM-Image](../user_guide/examples/online_serving/glm_image.md)
@@ -0,0 +1,116 @@
+# Full-Duplex WebSocket API
+
+vLLM-Omni provides an experimental full-duplex runtime for models that can
+continue receiving speech while producing speech. It adds persistent session
+state, model-specific turn policy, overlap handling, playback acknowledgement,
+and optional session resume.
+
+Full duplex is distinct from the turn-based [Realtime Audio API](realtime_api.md).
+
+## Choose an Endpoint
+
+| Endpoint | Protocol | Recommended use |
+|----------|----------|-----------------|
+| `WS /v1/realtime?duplex=1` | OpenAI Realtime-style event projection | Applications and browser clients |
+| `WS /v1/duplex` | Native vLLM-Omni duplex events | Runtime integration and low-level testing |
+
+Both endpoints use the same duplex engine and require the same model-side
+adapter. Prefer the Realtime projection unless the native lifecycle events are
+specifically required.
+
+## Enable Full Duplex
+
+The route becomes usable only when the deployment configuration explicitly
+sets:
+
+```yaml
+session_mode: duplex
+```
+
+The selected model pipeline must also provide a duplex serving adapter.
+Model-native deployments configure the corresponding engine runtime extension
+and control plane as part of their registered pipeline.
+
+!!! warning
+
+    `WS /v1/duplex` fails with `Duplex API is not available` when duplex is not
+    enabled. By contrast, `/v1/realtime?duplex=1` falls back to the ordinary
+    turn-based realtime handler when no duplex handler exists. Confirm that
+    `session.created.session.capabilities` is present before treating the
+    connection as full duplex.
+
+The current unified-runtime integrations are:
+
+- MiniCPM-o 4.5, using `vllm_omni/deploy/minicpmo_4_5_duplex.yaml`;
+- PersonaPlex, whose default `vllm_omni/deploy/personaplex.yaml` enables duplex.
+
+JoyVL is a separate HTTP interaction orchestrator and does not use these
+WebSocket endpoints. See [Standalone Experimental Servers](standalone_servers.md).
+
+## MiniCPM-o Quick Start
+
+Start the duplex deployment:
+
+```bash
+vllm serve openbmb/MiniCPM-o-4_5 --omni \
+  --deploy-config vllm_omni/deploy/minicpmo_4_5_duplex.yaml \
+  --trust-remote-code \
+  --port 8091
+```
+
+Stream a mono, PCM16, 16 kHz WAV file with the provided client:
+
+```bash
+python examples/online_serving/minicpmo/realtime_duplex_demo.py \
+  --url 'ws://localhost:8091/v1/realtime?duplex=1' \
+  --model openbmb/MiniCPM-o-4_5 \
+  --input-wav input_16k_mono.wav \
+  --ref-audio reference_voice.wav \
+  --output-dir /tmp/minicpmo-duplex
+```
+
+## Realtime Event Lifecycle
+
+A typical `/v1/realtime?duplex=1` session follows this lifecycle:
+
+1. Send `session.update` with the model, modalities, audio formats, and session
+   options.
+2. Wait for `session.created`; inspect `session.capabilities` instead of
+   assuming every model supports the same controls.
+3. Send `input_audio_buffer.append` events while microphone audio arrives.
+4. Send `input_audio_buffer.commit` at a user-turn boundary when required by
+   the model policy.
+5. Consume `response.created`, transcript deltas, `response.audio.delta`, and
+   `response.done` or `response.listen` events.
+6. Send `playback.ack` after audio has been played when the session advertises
+   playback acknowledgement support.
+7. Send `session.close` and wait for `session.closed`.
+
+Unlike the turn-based realtime endpoint, input may continue while a response
+is active. The server can emit `overlap.decision` to describe whether input was
+deferred, treated as a short acknowledgement, or used to interrupt output.
+
+## Capabilities and Model Differences
+
+The `session.created` payload includes capability fields such as
+`supports_barge_in`, `supports_playback_ack`, `supports_multi_session`,
+`supports_session_resume`, and `chunk_period_ms`. Treat this payload as the
+runtime contract.
+
+For example, PersonaPlex supports native overlapping speech but currently
+advertises `supports_barge_in=false`; destructive output interruption and
+model-state rewind have not been validated for that integration. Capacity and
+session-resume behavior also depend on the selected deployment configuration.
+
+## Native Protocol
+
+`WS /v1/duplex` exposes lower-level lifecycle names including
+`session.create`, `input_audio_buffer.append`, `turn.signal`, `playback.ack`,
+and `session.close`. It returns native session, input, response, overlap, and
+error events. This protocol is experimental and may evolve with the runtime;
+applications should use the provided Realtime client where possible.
+
+See the [MiniCPM-o example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/minicpmo),
+[PersonaPlex example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/personaplex),
+and [full-duplex runtime design](https://github.com/vllm-project/vllm-omni/blob/main/vllm_omni/experimental/fullduplex/DESIGN.md)
+for model-specific validation and architecture details.
@@ -0,0 +1,109 @@
+# OpenPI Robot Policy WebSocket API
+
+Use `WS /v1/realtime/robot/openpi` for low-latency robot-policy inference with
+OpenPI-compatible clients. A persistent WebSocket carries observations to the
+loaded policy model and returns action arrays.
+
+Despite the `/realtime` prefix, this endpoint does not use the OpenAI Realtime
+event schema. It uses binary MessagePack frames with NumPy extensions.
+
+## Availability
+
+The endpoint is available only when the loaded diffusion policy configuration
+contains `policy_server_config`. The server sends that model-specific
+configuration to the client as the first binary frame.
+
+For example, DreamZero declares:
+
+```yaml
+policy_server_config:
+  image_resolution: [180, 320]
+  n_external_cameras: 2
+  needs_wrist_camera: true
+  needs_stereo_camera: false
+  needs_session_id: true
+  action_space: joint_position
+```
+
+Without this configuration, the WebSocket returns `Robot policy not available`
+and closes.
+
+## DreamZero Quick Start
+
+Start the bundled DreamZero deployment:
+
+```bash
+vllm serve GEAR-Dreams/DreamZero-DROID --omni --port 8091 \
+  --served-model-name dreamzero-droid \
+  --deploy-config vllm_omni/deploy/dreamzero_tp1_cfg2.yaml \
+  --enforce-eager --disable-log-stats
+```
+
+Install the optional client dependencies, download the sample camera inputs,
+and run the client:
+
+```bash
+pip install openpi-client websockets opencv-python
+
+hf download YangshenDeng/vllm-omni-dreamzero-assets \
+  --repo-type dataset \
+  --local-dir outputs/dreamzero/assets
+
+python examples/online_serving/dreamzero/openpi_client.py \
+  --host 127.0.0.1 \
+  --port 8091 \
+  --video-dir outputs/dreamzero/assets
+```
+
+## Protocol
+
+The connection is request-response after an initial server handshake:
+
+```text
+connect
+  <- msgpack(policy_server_config)
+
+infer
+  -> msgpack({"endpoint": "infer", "session_id": "...", ...observation})
+  <- msgpack(ndarray | dict[str, ndarray])
+
+reset
+  -> msgpack({"endpoint": "reset"})
+  <- msgpack({"status": "reset successful"})
+```
+
+If `endpoint` is omitted, the server treats the message as `infer`. Each
+inference message produces one action response after the engine request
+completes; action tokens or intermediate tensors are not streamed.
+
+NumPy arrays use the marker format implemented by `openpi-client`. The server
+also accepts the legacy vLLM NumPy marker representation on input. JSON text
+frames are not observation messages.
+
+## Sessions and Reset
+
+- The API layer's current-session and first-call counters are scoped to each
+  WebSocket connection.
+- `session_id` identifies model-side state across observations on that
+  connection. If omitted, it defaults to `default`.
+- The first inference for a session is sent to the model with `reset=true`.
+- Changing `session_id`, or sending the `reset` command, causes the next
+  inference to start with `reset=true`.
+- The policy pipeline owns observation transforms and persistent model state,
+  normally keyed by `session_id`; the API layer forwards the raw observation
+  dictionary.
+
+## Limits and Errors
+
+- Maximum inbound payload size is 64 MiB.
+- The server closes an idle connection after 30 seconds.
+- Invalid binary input returns a MessagePack
+  `{"type":"error","message":"Invalid request payload"}` response.
+- Inference failures return a generic `Internal inference error` without
+  exposing an internal traceback.
+- Unsupported NumPy object, structured, and complex dtypes are rejected.
+
+Observation keys, camera layout, state tensors, and action shapes are defined
+by the loaded policy rather than by this transport. Read the handshake before
+constructing observations. See the [DreamZero example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/dreamzero)
+for a complete OpenPI client and DROID simulation loop.
@@ -0,0 +1,82 @@
+# Realtime Audio WebSocket API
+
+Use `WS /v1/realtime` to stream PCM audio into a compatible omni model and
+receive incremental transcription and synthesized response audio. This is a
+turn-based streaming API: it starts one generation over an audio stream and
+finishes after the client closes that input stream.
+
+For a model that listens while it is already speaking, use the experimental
+[Full-Duplex API](full_duplex_api.md) instead.
+
+## Quick Start
+
+Start a compatible model. The currently documented example uses Qwen3-Omni:
+
+```bash
+vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni --port 8091
+```
+
+Run the provided client with a mono, 16-bit PCM, 16 kHz WAV file:
+
+```bash
+python examples/online_serving/qwen3_omni/openai_realtime_client.py \
+  --url ws://localhost:8091/v1/realtime \
+  --model Qwen/Qwen3-Omni-30B-A3B-Instruct \
+  --input-wav input_16k_mono.wav \
+  --output-wav response.wav \
+  --output-text response.txt
+```
+
+The client dependency is `websockets`.
+
+## Protocol
+
+Messages in both directions are JSON text frames. Audio carried inside an
+event is base64-encoded raw PCM16.
+
+| Direction | Event | Purpose |
+|-----------|-------|---------|
+| Server to client | `session.created` | Confirms that the WebSocket connection is ready |
+| Client to server | `session.update` | Selects and validates the served model |
+| Client to server | `input_audio_buffer.commit` with `final: false` | Starts generation over the incoming stream |
+| Client to server | `input_audio_buffer.append` | Appends base64 PCM16 audio |
+| Client to server | `input_audio_buffer.commit` with `final: true` | Marks the end of input |
+| Server to client | `transcription.delta` | Carries incremental response text |
+| Server to client | `transcription.done` | Carries final text and token usage |
+| Server to client | `response.audio.delta` | Carries incremental PCM16 response audio |
+| Server to client | `response.audio.done` | Marks the end of response audio |
+| Server to client | `error` | Reports an invalid event, model, or audio payload |
+
+A minimal client sends events in this order:
+
+```json
+{"type":"session.update","model":"Qwen/Qwen3-Omni-30B-A3B-Instruct"}
+{"type":"input_audio_buffer.commit","final":false}
+{"type":"input_audio_buffer.append","audio":"<base64-pcm16>"}
+{"type":"input_audio_buffer.commit","final":true}
+```
+
+The initial non-final commit intentionally starts generation before all audio
+has arrived. Continue sending `input_audio_buffer.append` events while the
+engine is consuming the stream.
+
+## Audio Handling
+
+- Input is mono PCM16 at 16 kHz for the Qwen3-Omni example.
+- `response.audio.delta.audio` contains base64-encoded PCM16 bytes.
+- Read `sample_rate_hz` from each audio event instead of assuming an output
+  rate. Qwen3-Omni output is typically 24 kHz.
+- Concatenate audio deltas in receive order to construct the output waveform.
+- Set `input_audio_buffer.commit.final` to `true` after the last input chunk so
+  the server can terminate the streaming request.
+
+## Availability and Limitations
+
+The path is registered on the unified API server, but it is usable only when
+the loaded pipeline implements realtime audio input and produces compatible
+audio output. Unsupported deployments return an `error` event. The endpoint
+does not provide duplex session resume, playback acknowledgement, overlap
+policy, or barge-in controls.
+
+See the [Qwen3-Omni online serving example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/qwen3_omni)
+for concurrency options, chunk pacing, and per-delta audio debugging.
@@ -2,7 +2,7 @@

 vLLM-Omni provides an OpenAI-compatible API for text-to-speech (TTS) generation. Supported TTS models include:

- **Qwen3-TTS** (`Qwen/Qwen3-TTS-12Hz-*`) -- Qwen3-based TTS with CustomVoice, VoiceDesign, and Base (voice cloning) task types. Output: 24 kHz.
+- **Qwen3-TTS** (`Qwen/Qwen3-TTS-12Hz-*`) -- Qwen3-based TTS with CustomVoice, VoiceDesign, and Base (voice cloning) task types. Native output: 24 kHz; API output can be resampled to 8 kHz.
 - **Fish Speech S2 Pro** (`fishaudio/s2-pro`) -- Dual-AR TTS with DAC codec. Supports text-to-speech and voice cloning via reference audio. Output: 44.1 kHz.
 - **Voxtral TTS** (`mistralai/Voxtral-4B-TTS-2603`) -- AR + FlowMatching TTS with preset voices. Output: 24 kHz.
 - **CosyVoice3** (`FunAudioLLM/Fun-CosyVoice3-0.5B-2512`) -- 2-stage talker + flow-matching code2wav. Voice cloning via `ref_audio` + `ref_text` (no presets). Output: 24 kHz.
@@ -116,6 +116,7 @@ Content-Type: application/json

 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
+| `sample_rate` | integer | model native | Target output sample rate. Qwen3-TTS supports 8000 or 24000 Hz; the model remains native 24 kHz internally and is resampled before encoding. |
 | `task_type` | string | "CustomVoice" | TTS task type: CustomVoice, VoiceDesign, or Base |
 | `language` | string | "Auto" | Language (see supported languages below) |
 | `instructions` | string | "" | Voice style/emotion instructions |
@@ -0,0 +1,109 @@
+# Standalone Experimental Servers
+
+Most vLLM-Omni endpoints run on the unified server started with
+`vllm serve <model> --omni`. The experimental full-duplex package also contains
+two separate user-facing server processes. Their routes are not added to the
+unified API server.
+
+| Server | Transport | Role |
+|--------|-----------|------|
+| PersonaPlex compatibility server | WebSocket | Hosts the official PersonaPlex browser protocol and a raw-PCM alternative |
+| JoyVL interaction server | HTTP | Adds state, memory, proactive decisions, and delegation in front of a separate OpenAI-compatible model backend |
+
+## PersonaPlex Compatibility Server
+
+Use this server when a client requires the official Moshi/PersonaPlex wire
+protocol or the bundled browser UI. For the engine-owned unified duplex path,
+use [`WS /v1/realtime?duplex=1`](full_duplex_api.md) instead.
+
+Start it with:
+
+```bash
+HF_TOKEN=... CUDA_VISIBLE_DEVICES=0 python -m \
+  vllm_omni.experimental.fullduplex.personaplex.serving.server \
+  --port 8091
+```
+
+| Route | Purpose |
+|-------|---------|
+| `GET /` | Official PersonaPlex browser client, when its assets are available |
+| `GET /health` | Readiness check |
+| `WS /api/chat` | Moshi-compatible tagged binary protocol with Opus audio |
+| `WS /v1/audio/duplex` | Simpler float32 raw-PCM protocol |
+
+### `/api/chat` Protocol
+
+The first byte of every binary frame is a message tag:
+
+| Direction | Tag | Payload |
+|-----------|-----|---------|
+| Server to client | `0x00` | Ready handshake |
+| Client to server | `0x01` | Opus microphone audio at 24 kHz |
+| Server to client | `0x01` | Opus response audio |
+| Server to client | `0x02` | UTF-8 inner-monologue text |
+
+The optional `text_prompt` and `voice_prompt` query parameters select a persona
+and voice. The default server admits one conversation; use `--batch-size N` to
+configure multiple lockstep slots. Connections beyond capacity close with
+WebSocket code 1013.
+
+### `/v1/audio/duplex` Protocol
+
+Send `{"type":"open","persona":"...","voice":"..."}` as a JSON text frame,
+then exchange binary float32 PCM frames. The server returns generated float32
+audio as binary frames and text as `{"type":"text","text":"..."}`. Send
+`{"type":"close"}` and wait for `{"type":"done"}` to flush and finish.
+
+See the [PersonaPlex example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/personaplex)
+for the browser and headless clients.
+
+## JoyVL Interaction Server
+
+JoyVL is an orchestration layer, not a model-serving engine. It calls a
+separate OpenAI-compatible backend and adds per-session frame history, memory,
+persona policy, and optional delegation.
+
+Start the model backend and orchestrator separately:
+
+```bash
+vllm serve jdopensource/JoyAI-VL-Interaction-Preview \
+  --served-model-name JoyAI-VL-Interaction-Preview \
+  --port 8092 \
+  --max-model-len 131072 \
+  --enable-prefix-caching \
+  --limit-mm-per-prompt '{"image":256,"video":1}'
+
+python -m vllm_omni.experimental.fullduplex.joyvl.serving.server \
+  --port 8091 \
+  --main-backend-url http://127.0.0.1:8092/v1 \
+  --main-model JoyAI-VL-Interaction-Preview
+```
+
+| Method and route | Purpose |
+|------------------|---------|
+| `GET /health` | Readiness check |
+| `GET /v1/models` | Reports the configured interaction model |
+| `POST /v1/chat/completions` | Processes one frame or interaction tick |
+| `POST /reset` | Resets a session |
+| `POST /v1/streaming/reset` | Alias for `/reset` |
+| `POST /v1/streaming/persona` | Changes the session persona |
+
+Send one multimodal Chat Completions request per video frame, normally around
+one frame per second, and identify the session with `x-session-id`. Responses
+include an `interaction` block whose action is `silence`, `response`, or
+`delegate`.
+
+```bash
+curl http://127.0.0.1:8091/v1/chat/completions \
+  -H 'content-type: application/json' \
+  -H 'x-session-id: demo' \
+  -d '{
+    "messages": [{"role": "user", "content": [
+      {"type": "text", "text": "Alert me if a fire appears"},
+      {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}
+    ]}]
+  }'
+```
+
+See the [JoyVL recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/JD/JoyAI-VL-Interaction.md)
+for memory, personas, delegation backends, and the browser UI.
@@ -0,0 +1,92 @@
+# Streaming Video Output WebSocket API
+
+Use `WS /v1/realtime/video` to receive fragmented MP4 chunks while a compatible
+diffusion video model is generating. This endpoint streams generated video
+output; it is unrelated to [`WS /v1/video/chat/stream`](video_stream_api.md),
+which streams live video input into a conversational model.
+
+## Quick Start
+
+Start a model that supports step execution and enable diffusion streaming
+output:
+
+```bash
+vllm serve BestWishYsh/Helios-Distilled --omni \
+  --diffusion-streaming-output \
+  --port 8091
+```
+
+Install the client dependency and run the provided client:
+
+```bash
+pip install av websockets
+
+python examples/online_serving/streaming_video_generation/streaming_video_client.py \
+  --host 127.0.0.1 \
+  --port 8091 \
+  --model BestWishYsh/Helios-Distilled \
+  --prompt "A serene lakeside sunrise with mist over the water" \
+  --output output.mp4
+```
+
+## Protocol
+
+Control events are JSON text frames. Each `video.chunk_metadata` event is
+immediately followed by one binary WebSocket frame containing fragmented MP4
+bytes.
+
+| Direction | Event or frame | Purpose |
+|-----------|----------------|---------|
+| Client to server | `session.start` | Starts generation with a prompt and video parameters |
+| Server to client | `video.start` | Confirms the request ID, format, and accepted configuration |
+| Server to client | `video.chunk_metadata` | Describes the next binary media frame |
+| Server to client | Binary frame | Carries one fragmented MP4 chunk |
+| Client to server | `session.interaction` | Queues an optional prompt update during generation |
+| Server to client | `session.interaction.queued` | Confirms that an interaction was queued |
+| Client to server | `session.ping` | Refreshes the stall clock |
+| Server to client | `session.pong` | Acknowledges a ping |
+| Client to server | `session.stop` | Cancels the active generation |
+| Server to client | `session.done` | Reports completion or cancellation and the chunk count |
+| Server to client | `error` | Reports validation, generation, or timeout failures |
+
+Start a minimal session with:
+
+```json
+{
+  "type": "session.start",
+  "model": "BestWishYsh/Helios-Distilled",
+  "prompt": "A serene lakeside sunrise",
+  "format": "m4s",
+  "width": 640,
+  "height": 384,
+  "fps": 16,
+  "num_frames": 99
+}
+```
+
+Only `m4s` output is currently accepted. Concatenate the binary frames in
+receive order, then remux the fragmented stream if the target player requires
+a progressive MP4 file.
+
+## Mid-Generation Interaction
+
+Models that implement interaction updates can change the active prompt:
+
+```json
+{
+  "type": "session.interaction",
+  "interaction": {
+    "event_id": "sunlight-update",
+    "event": {"prompt": "Sunlight breaks through the mist"},
+    "transition_chunks": 2
+  }
+}
+```
+
+This feature is pipeline-dependent. An enabled endpoint does not imply that
+the loaded model supports midway prompt updates.
+
+The default start timeout is 10 seconds and the server reports a stall after
+about 60 seconds without generation progress or a `session.ping`. See the
+[streaming video generation example](https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/streaming_video_generation)
+for prompt-update scheduling and browser playback.
@@ -138,7 +138,7 @@ The following tables show which models support each feature:
 | **Qwen-Image-Edit**      |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Qwen-Image-Edit-2509** |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |        ✅ (decode)         |          ✅           |       ❌        |        ❌         |
 | **Qwen-Image-Layered**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
-| **SenseNova-U1**         |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
+| **SenseNova-U1 / U1.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Stable-Diffusion-XL**  |     ❌     |     ❌      |           ✅           |       ✅        |         ✅         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Stable-Diffusion3.5**  |     ❌     |     ✅      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **Z-Image**              |     ✅     |     ✅      |           ✅           |       ❓        |   ✅ (TP=2 only)   |          ❌          |   ✅    |             ❌             |      ✅ (decode)      |       ✅        |        ❌         |
@@ -49,7 +49,7 @@ scenario below for live barge-in validation on the target GPU.
 Default:

 ```bash
-vllm-omni serve openbmb/MiniCPM-o-4_5 \
+vllm serve openbmb/MiniCPM-o-4_5 \
    --omni \
    --deploy-config vllm_omni/deploy/minicpmo_4_5.yaml \
    --trust-remote-code \
@@ -37,7 +37,7 @@ During generation the client normally sends only `session.start` and then receiv
 ## Install Client Dependency

 ```bash
-pip install websockets
+pip install av websockets
 # For the Gradio demo:
 pip install 'vllm-omni[demo]' websockets
 ```
@@ -8,7 +8,7 @@ the received bytes to disk when the session finishes or is interrupted.
 Image/reference input is intentionally not supported in this example yet.

 Requirements:
-    pip install websockets
+    pip install av websockets
 """

 import argparse
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """OpenAI-compatible client for Qwen3-TTS via /v1/audio/speech endpoint.

 This script demonstrates how to use the OpenAI-compatible speech API
@@ -74,6 +77,8 @@ def run_tts_generation(args) -> None:
        "voice": args.speaker,
        "response_format": args.response_format,
    }
+    if args.sample_rate is not None:
+        payload["sample_rate"] = args.sample_rate

    # Add optional parameters
    if args.instructions:
@@ -246,6 +251,13 @@ def parse_args():
        choices=["wav", "mp3", "flac", "pcm", "aac", "opus"],
        help="Audio output format (default: wav)",
    )
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=None,
+        choices=[8000, 24000],
+        help="Output sample rate in Hz (default: model native)",
+    )
    parser.add_argument(
        "--output",
        "-o",
@@ -1,11 +1,14 @@
 # Boogu-Image

-> Text-to-image and image-editing online serving (Boogu-Image-0.1-Base / -Edit)
+> Text-to-image and image-editing online serving
+(Boogu-Image-0.1-Base / -Edit / -Edit-Turbo)

 ## Summary

 - Vendor: Boogu
- Model: `Boogu/Boogu-Image-0.1-Base` (text-to-image), `Boogu/Boogu-Image-0.1-Edit` (image editing)
+- Model: `Boogu/Boogu-Image-0.1-Base` (text-to-image),
+  `Boogu/Boogu-Image-0.1-Edit` (image editing), and
+  `Boogu/Boogu-Image-0.1-Edit-Turbo` (four-step image editing)
 - Task: Text-to-image generation and text-guided image editing (TI2I)
 - Mode: Online serving with the OpenAI-compatible API
 - Maintainer: Community
@@ -13,19 +16,23 @@
 ## When to use this recipe

 Use this recipe when you want a known-good starting point for serving
-`Boogu/Boogu-Image-0.1-Base` with vLLM-Omni's native pipeline (no
+`Boogu/Boogu-Image-0.1-Base`, `Boogu/Boogu-Image-0.1-Edit`, or
+`Boogu/Boogu-Image-0.1-Edit-Turbo` with vLLM-Omni's native pipeline (no
 `--diffusion-load-format diffusers`, and the upstream `boogu` package is not
 required).

 Boogu-Image-0.1 is an Apache-2.0 unified image generation and editing model
-family. This recipe covers the Base text-to-image checkpoint, which pairs a
-Qwen3-VL multimodal encoder with a Diffusion Transformer (DiT) and a flow-match
-Euler scheduler with time-shift. It handles photorealistic generation and
-Chinese/English text rendering.
+family. The Base text-to-image checkpoint pairs a Qwen3-VL multimodal encoder
+with a Diffusion Transformer (DiT) and a flow-match Euler scheduler with
+time-shift. It handles photorealistic generation and Chinese/English text
+rendering. The Edit and Edit-Turbo checkpoints use the same native
+`BooguImagePipeline` for image editing.

 ## References

 - Upstream model card: <https://huggingface.co/Boogu/Boogu-Image-0.1-Base>
+- Edit model card: <https://huggingface.co/Boogu/Boogu-Image-0.1-Edit>
+- Edit-Turbo 1K hotfix: <https://huggingface.co/Boogu/Boogu-Image-0.1-Edit-Turbo/tree/hotfix-1k-20260708>
 - Project page: <https://boogu.org>
 - GitHub: <https://github.com/boogu-project/Boogu-Image>
 - Related example: [`examples/online_serving/text_to_image/`](../../examples/online_serving/text_to_image/README.md)
@@ -93,7 +100,7 @@ curl -s http://localhost:8091/v1/chat/completions \
 - **Memory usage:** ~34.6 GiB on GPU. Use `--vae-use-slicing --vae-use-tiling`
  to trim peak VRAM if needed.
 - **Key flags:**
-  - `--omni` — enables vLLM-Omni diffusion serving.
+    - `--omni` — enables vLLM-Omni diffusion serving.
 - **Guidance:** Boogu-Image uses `guidance_scale` (mapped to the upstream
  `text_guidance_scale`); the default is `4.0`. Classifier-free guidance is
  active whenever `guidance_scale > 1.0`.
@@ -111,13 +118,13 @@ the image-editing (TI2I) path activates automatically when a request carries a
 reference image. The Base text-to-image path is unaffected (no reference image
 is sent).

-#### Command
+### Command

 ```bash
 vllm serve Boogu/Boogu-Image-0.1-Edit --omni --port 8091
 ```

-#### Verification
+### Verification

 Edit an image with `/v1/images/edits` (the model-card example — change a photo
 to a colored-pencil drawing). Diffusion parameters are plain multipart form
@@ -157,14 +164,14 @@ curl -s http://localhost:8091/v1/chat/completions \
  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2- | base64 -d > edited.png
 ```

-#### Notes
+### Notes

 - **Single reference image:** only one input image is supported for now (the
  upstream "Only support 1 reference image for now" limit).
 - **Guidance semantics:**
-  - `guidance_scale` = text guidance (upstream `text_guidance_scale`, default
+    - `guidance_scale` = text guidance (upstream `text_guidance_scale`, default
    `4.0`); `> 1.0` enables text CFG. Editing typically uses `5.0`.
-  - `guidance_scale_2` = image guidance (upstream `image_guidance_scale`,
+    - `guidance_scale_2` = image guidance (upstream `image_guidance_scale`,
    default `1.0` = off). Setting it `> 1.0` enables the double-guidance path
    (3 model predictions per step), steering more strongly toward the reference
    image.
@@ -178,3 +185,69 @@ curl -s http://localhost:8091/v1/chat/completions \
  supports Boogu-Image-Edit directly
  (`--model Boogu/Boogu-Image-0.1-Edit --guidance-scale 5.0`, optional
  `--guidance-scale-2`).
+
+## Fast image editing (Boogu-Image-0.1-Edit-Turbo)
+
+Edit-Turbo is the distilled image-editing checkpoint. Its `model_index.json`
+declares `BooguImagePipeline`, so it uses the same native TI2I path as Edit.
+The upstream 1K hotfix is recommended for stable results and is pinned below
+to avoid silently loading the older checkpoint from the repository's default
+revision.
+
+### Command
+
+```bash
+vllm serve Boogu/Boogu-Image-0.1-Edit-Turbo \
+  --omni \
+  --revision hotfix-1k-20260708 \
+  --port 8091
+```
+
+### Verification
+
+Edit an image with the distilled four-step settings:
+
+```bash
+curl -s http://localhost:8091/v1/images/edits \
+  -F model="Boogu/Boogu-Image-0.1-Edit-Turbo" \
+  -F image="@input.png" \
+  -F prompt="Change the style to a colored pencil drawing." \
+  -F num_inference_steps=4 \
+  -F guidance_scale=1.0 \
+  -F seed=42 \
+  | jq -r '.data[0].b64_json' | base64 -d > edited-turbo.png
+```
+
+Or via chat completions:
+
+```bash
+curl -s http://localhost:8091/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "messages": [
+      {"role": "user", "content": [
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,<BASE64>"}},
+        {"type": "text", "text": "Change the style to a colored pencil drawing."}
+      ]}
+    ],
+    "extra_body": {
+      "num_inference_steps": 4,
+      "guidance_scale": 1.0,
+      "seed": 42
+    }
+  }' | jq -r '.choices[0].message.content[0].image_url.url' | cut -d',' -f2- | base64 -d > edited-turbo.png
+```
+
+### Notes
+
+- **Pinned revision:** use `hotfix-1k-20260708`; upstream recommends the 1K
+  hotfix over the 1.5K variant for more stable results.
+- **Recommended settings:** `num_inference_steps=4` and
+  `guidance_scale=1.0`. Edit-Turbo is guidance-distilled, meaning the guidance
+  behavior is already baked into the distilled checkpoint. A scale of `1.0`
+  avoids applying additional classifier-free guidance; unlike the regular Edit
+  checkpoint, it should not be increased to `5.0`.
+- **Reference images:** the same single-reference and `align_res` behavior as
+  the regular Edit checkpoint applies.
+- **Known limitations:** the same single-GPU limitations as Base and Edit apply;
+  CPU offload, Cache-DiT, and multi-GPU parallelism are not yet validated.
@@ -1142,3 +1142,4 @@ vllm serve "${MODEL_ROOT}/FL2VA" \
 - [Supported models](../../docs/models/supported_models.md)
 - [Video API](../../docs/serving/videos_api.md)
 - [Diffusion parallelism](../../docs/user_guide/diffusion/parallelism/overview.md)
+- [Temporal VAE chunk callback design](../../docs/design/feature/minimax_h3_temporal_vae_chunks.md)
@@ -78,6 +78,7 @@ recipes/
 | [`Qwen/Qwen3-Omni.md`](./Qwen/Qwen3-Omni.md) | Online serving for multimodal chat | 1x A100 80GB |
 | [`Qwen/Qwen3-TTS.md`](./Qwen/Qwen3-TTS.md) | Text-to-speech serving (CustomVoice / VoiceDesign / Base) | 1x H100/A100 80GB / 1x AMD MI300X |
 | [`SenseNova/SenseNova-U1.md`](./SenseNova/SenseNova-U1.md) | Unified image generation and understanding | 1x or 2x H200 / 1x AMD MI300X |
+| [`SenseNova/SenseNova-U1.5.md`](./SenseNova/SenseNova-U1.5.md) | Unified image generation and understanding, with an 8-step distilled LoRA | 1x A800 80GB |
 | [`StabilityAI/Stable-Audio-Open.md`](./StabilityAI/Stable-Audio-Open.md) | Offline + online text-to-audio generation (Stable Audio Open) | 1x RTX 4090 24GB / 1x AMD MI300X |
 | [`Tencent/Covo-Audio-Chat.md`](./Tencent/Covo-Audio-Chat.md) | Online serving for audio chat | 1x A100 80GB |
 | [`Tencent/HunyuanImage-3.0-Instruct.md`](./Tencent/HunyuanImage-3.0-Instruct.md) | DiT-only text-to-image serving and benchmark, including ModelOpt mixed FP8/NVFP4 | 4x H100/H800 80GB / 2x B200 |
@@ -0,0 +1,159 @@
+# SenseNova-U1.5
+
+> Unified image generation and understanding, with an official 8-step distilled LoRA
+
+## Summary
+
+- Vendor: SenseNova
+- Model: `sensenova/SenseNova-U1.5-8B-MoT`
+- LoRA: `sensenova/SenseNova-U1.5-8B-MoT-LoRAs` (`SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors`)
+- Task: text2img, img2img, img2text (visual understanding), text2text (chat)
+- Mode: Offline inference, Online serving (OpenAI-compatible API)
+- Maintainer: Community
+
+## When to use this recipe
+
+U1.5 runs on the same `SenseNovaU1Pipeline` as U1 — the checkpoint keeps
+`model_type: neo_chat`, so it resolves without `--model-class-name`. Relative to U1 the
+config flips two fields, `use_pixel_head` to `true` (the flow-matching head becomes a
+`ConvDecoder`) and `noise_scale_max_value` from 8.0 to 16.0; both are already read by
+`SenseNovaU1Config`. The checkpoint is 13 shards / 50.2 GB on disk, 30.3 GB of that fp32, and
+loads to roughly 34 GB in bf16.
+
+Use this recipe for U1.5 specifically, including its distilled few-step LoRA. For U1 see
+[SenseNova-U1](SenseNova-U1.md).
+
+## Hardware Support
+
+### GPU
+
+#### 1x A800 80GB
+
+- 1x NVIDIA A800 80GB PCIe with an Intel Xeon Gold 6336Y, driver 595.84, Linux
+- Python 3.12.13, vLLM 0.28.0, torch 2.13.0+cu130, CUDA 13.0, diffusers 0.40.0,
+  transformers 5.14.1, flashinfer 0.6.16.post3, BF16, TP=1
+- Peak GPU memory: 34.3 GB at 1024x1024, 36.4 GB at 1536x2720
+
+##### Text-to-Image
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, warm natural light." \
+    --width 1024 --height 1024 \
+    --seed 42 --num-inference-steps 50 --cfg-scale 4.0 \
+    --extra-body '{"think": false, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
+    --output sensenova_u15_t2i.png
+```
+
+Think mode (`"think": true`) is recommended for higher image quality.
+
+##### Image-to-Image Editing
+
+```bash
+python examples/offline_inference/image_to_image/image_edit.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --prompt "Turn this into an oil painting" \
+    --image input.png --resolution 1024 \
+    --seed 42 --num-inference-steps 25 --cfg-scale 4.0 \
+    --extra-args '{"think": true, "img_cfg_scale": 1.0, "cfg_norm": "none", "timestep_shift": 3.0}' \
+    --output sensenova_u15_edit.png
+```
+
+##### 8-step distilled LoRA
+
+The distilled LoRA is fused into the generation tower at load time:
+
+```bash
+python examples/offline_inference/text_to_image/text_to_image.py \
+    --model sensenova/SenseNova-U1.5-8B-MoT \
+    --lora-path SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors --lora-backend distill \
+    --prompt "Close portrait of an elderly woman by a farmhouse window, warm natural light." \
+    --width 1024 --height 1024 \
+    --seed 42 --num-inference-steps 8 --cfg-scale 1.0 \
+    --extra-body '{"think": false, "cfg_norm": "none", "timestep_shift": 3.0, "t_eps": 0.02}' \
+    --output sensenova_u15_lora8.png
+```
+
+**Use `--cfg-scale 1.0` with this LoRA.** It is distilled with DMD and runs without
+classifier-free guidance; the default `4.0` applies guidance twice and produces a
+blown-out, posterised image.
+
+##### Online serving
+
+```bash
+vllm serve sensenova/SenseNova-U1.5-8B-MoT --omni --port 8091
+
+python examples/online_serving/sensenova_u1/openai_chat_client.py \
+    -s http://127.0.0.1:8091 -m img2text -i input.png -p "Describe this image."
+```
+
+`-s` takes the base URL; the client appends `/v1` itself.
+
+#### Measured latency (1x A800 80GB, 25 steps, median of 3 after a warmup)
+
+| Resolution | Step latency | Total |
+| --- | --- | --- |
+| 1024x1024 | 209.5 ms | 5.24 s |
+| 1536x1536 | 458.5 ms | 11.46 s |
+
+At 1536x2720 with 50 steps, end-to-end is 43.0 s.
+
+#### Measured latency (1x NVIDIA H200 139GB, single run, no warmup)
+
+Reported by @hsliuustc0106 against PR head `29c090d`, on one reserved CUDA device.
+
+- Python 3.12.13, vLLM 0.28.0, torch 2.13.0+cu130, CUDA 13.0, diffusers 0.40.0,
+  transformers 5.14.1, flashinfer 0.6.16.post3, BF16, TP=1
+- `sensenova/SenseNova-U1.5-8B-MoT` with `SenseNova-U1.5-8B-MoT-LoRA-8step.safetensors`,
+  seed 42, 1024x1024
+
+| Case | Stage latency | Peak GPU memory |
+| --- | --- | --- |
+| 50 steps, cfg 4.0, think off | 9017 ms | 34384 MiB |
+| 8 steps, cfg 1.0, no LoRA control | 624.00 ms, 77.85 ms/step | 34376 MiB |
+| 8 steps, cfg 1.0, distilled LoRA | 623.85 ms, 77.98 ms/step | 34364 MiB |
+| 50 steps, cfg 4.0, think on, `VLLM_OMNI_SENSENOVA_PAGED_DECODE=1` | 81622 ms | 34594 MiB |
+
+Each row runs the Text-to-Image command above with `--num-inference-steps` and
+`--cfg-scale` as listed and `think` set through `--extra-body`; the LoRA row adds
+`--lora-path` and `--lora-backend distill`.
+
+Single runs without a preceding warmup, so the think-on row carries the first-request
+compile rather than a steady state. The distilled LoRA fused into 168 parameters, and its
+image differs from the same-seed no-LoRA control in 1,048,574 of 1,048,576 pixels, MAE 58.62.
+
+```bash
+pytest -m "cpu and not cuda" tests/diffusion/models/sensenova_u1/ tests/diffusion/lora/test_loader.py -q
+pytest -m "cpu and not cuda" tests/config/test_environment_variables.py tests/diffusion/test_diffusion_worker.py -q
+pytest -m cuda tests/diffusion/models/sensenova_u1/test_sensenova_u1_attention_gqa.py \
+    tests/diffusion/models/sensenova_u1/test_sensenova_u1_paged_decode.py -q --tb=short
+```
+
+71, 13 and 9 passed. Serving reached `/health`; `img2text` returned a 1921-character
+description and `text2text` the expected answer.
+
+#### Verification
+
+```bash
+pytest -q tests/diffusion/models/sensenova_u1/
+```
+
+## Notes
+
+- Both `use_pixel_head` and `noise_scale_max_value` come from `config.json`; no flag is needed.
+- The checkpoint carries no `configuration_neo_chat.py`, so the loader logs a few
+  "does not appear to have a file named configuration_neo_chat.py" errors during startup and
+  then proceeds on the in-tree config. Generation is unaffected.
+- The LoRA targets the generation tower only (`*_mot_gen`); understanding-tower weights are
+  untouched.
+- Autoregressive decode (think, text-to-text and image-to-text) runs on a paged K/V cache under
+  a captured CUDA graph. It falls back to the ordinary cache when the device or the bundled
+  `flash_attn_varlen_func` cannot support it; set `VLLM_OMNI_SENSENOVA_PAGED_DECODE=0` to force
+  that fallback.
+- The first request after startup costs about 0.7 s more than the steady state whether the paged
+  path is on or off. Measured on one A800 with the inductor, triton and vLLM compile caches all
+  cleared, median of three runs: 718 ms above steady with the path on, 679 ms with it off.
+- Each request captures its own graphs, and a think request captures twice because the sequence
+  grows past the 512 bucket, so the capture cost is paid per request rather than once at
+  startup.
@@ -0,0 +1,42 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from pathlib import Path
+from shlex import split
+
+import pytest
+import yaml
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+AMD_MERGE_PIPELINE = Path(".buildkite/amd/test-amd-merge.yml")
+
+
+def _find_step(label: str) -> dict:
+    pipeline = yaml.safe_load(AMD_MERGE_PIPELINE.read_text(encoding="utf-8"))
+
+    def walk(steps: list[dict]) -> dict | None:
+        for step in steps:
+            if step.get("label") == label:
+                return step
+            if nested := walk(step.get("steps", [])):
+                return nested
+        return None
+
+    step = walk(pipeline.get("steps", []))
+    assert step is not None, f"missing AMD pipeline step: {label}"
+    return step
+
+
+def test_qwen3_tts_base_preserves_advanced_model_arguments() -> None:
+    step = _find_step("Qwen3-TTS Base E2E Test")
+    commands = step["commands"]
+
+    assert all("bash -c" not in command for command in commands)
+    pytest_command = next(command for command in commands if "pytest" in command)
+    argv = split(pytest_command)
+
+    marker_index = argv.index("-m")
+    run_level_index = argv.index("--run-level")
+    assert argv[marker_index + 1] == "advanced_model and cuda"
+    assert argv[run_level_index + 1] == "advanced_model"
@@ -41,6 +41,11 @@ from vllm_omni.engine.arg_utils import SHARED_FIELDS, internal_blacklist_keys

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]

+SINGLE_STAGE_PIPE_CFG = (StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
+# Validation strings to match against in post init
+NO_STAGES_MATCH_STR = "no stages"
+NO_TERMINAL_STAGE_MATCH_STR = "No terminal stage"
+

@pytest.fixture(autouse=True)
 def _stable_test_platform(monkeypatch):
@@ -618,6 +623,7 @@ class TestPipelineDiscovery:
        p = PipelineConfig(
            model_type="custom_collide",
            hf_architectures=("SomeCollidingArch",),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        assert p.hf_architectures == ("SomeCollidingArch",)

@@ -639,41 +645,42 @@ class TestStagePipelineConfig:


 class TestPipelineConfigNew:
+    def test_simple_init(self):
+        # Ensure pipeline config validates in post init.
+        PipelineConfig(
+            model_type="t",
+            model_arch="A",
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
+
    def test_frozen(self):
-        p = PipelineConfig(model_type="t", model_arch="A")
+        p = PipelineConfig(
+            model_type="t",
+            model_arch="A",
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
        with pytest.raises(AttributeError):
            p.model_type = "changed"

-    def test_validate_valid(self):
-        p = PipelineConfig(
-            model_type="t",
-            model_arch="A",
-            stages=(
-                StagePipelineConfig(stage_id=0, model_stage="a"),
-                StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,), final_output=True),
-            ),
-        )
-        assert p.validate() == []
-
    def test_validate_no_stages(self):
-        p = PipelineConfig(model_type="t", model_arch="A")
-        assert any("no stages" in e.lower() for e in p.validate())
+        with pytest.raises(ValueError, match=NO_STAGES_MATCH_STR):
+            PipelineConfig(model_type="t", model_arch="A")

    def test_validate_no_terminal_stage(self):
        """A pipeline with no ``final_output`` stage can never emit a result."""
-        p = PipelineConfig(
-            model_type="t",
-            model_arch="A",
-            stages=(
-                StagePipelineConfig(stage_id=0, model_stage="a"),
-                StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,)),
-            ),
-        )
-        assert any("no terminal stage" in e.lower() for e in p.validate())
+        with pytest.raises(ValueError, match=NO_TERMINAL_STAGE_MATCH_STR):
+            PipelineConfig(
+                model_type="t",
+                model_arch="A",
+                stages=(
+                    StagePipelineConfig(stage_id=0, model_stage="a"),
+                    StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,)),
+                ),
+            )

-    def test_validate_terminal_stage_need_not_be_last(self):
-        """The check is cycle-agnostic: any stage may carry ``final_output``."""
-        p = PipelineConfig(
+    def test_pipeline_can_have_final_output_in_any_stage(self):
+        """Ensure any stage may carry ``final_output``."""
+        PipelineConfig(
            model_type="t",
            model_arch="A",
            stages=(
@@ -681,15 +688,20 @@ class TestPipelineConfigNew:
                StagePipelineConfig(stage_id=1, model_stage="b", input_sources=(0,)),
            ),
        )
-        assert p.validate() == []


 class TestPipelineRegistration:
    def test_resolve_pipeline_prefers_deploy_pipeline_key(self, clean_pipeline_registry, tmp_path):
        deploy_key = "deploy_selected_pipeline"
        model_type_key = "hf_model_type_pipeline"
-        deploy_pipe = PipelineConfig(model_type=deploy_key)
-        model_type_pipe = PipelineConfig(model_type=model_type_key)
+        deploy_pipe = PipelineConfig(
+            model_type=deploy_key,
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
+        model_type_pipe = PipelineConfig(
+            model_type=model_type_key,
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
        register_pipeline(deploy_pipe)
        register_pipeline(model_type_pipe)

@@ -715,7 +727,11 @@ class TestPipelineRegistration:
        tmp_path,
    ):
        model_type_key = "registered_model_type_pipeline"
-        register_pipeline(PipelineConfig(model_type=model_type_key))
+        pipeline_cfg = PipelineConfig(
+            model_type=model_type_key,
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
+        register_pipeline(pipeline_cfg)

        deploy_path = tmp_path / "deploy.yaml"
        deploy_path.write_text("pipeline: missing_pipeline\n", encoding="utf-8")
@@ -750,6 +766,7 @@ class TestPipelineRegistration:
        pipe_cfg = PipelineConfig(
            model_type=pipeline_key,
            hf_architectures=("ArchitectureFallbackForTest",),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)

@@ -780,15 +797,18 @@ class TestPipelineRegistration:
            model_type="rejected_by_predicate",
            hf_architectures=(shared_arch,),
            hf_config_predicate=rejecting_predicate,
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        raises_cfg = PipelineConfig(
            model_type="raises_in_predicate",
            hf_architectures=(shared_arch,),
            hf_config_predicate=raising_predicate,
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        accept_cfg = PipelineConfig(
            model_type="accepted_by_predicate",
            hf_architectures=(shared_arch,),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(reject_cfg)
        register_pipeline(raises_cfg)
@@ -822,6 +842,7 @@ class TestPipelineRegistration:
            model_type="predicate_without_arch_match",
            hf_architectures=("DifferentArchitectureForTest",),
            hf_config_predicate=predicate,
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)

@@ -845,6 +866,7 @@ class TestPipelineRegistration:
        resolved_cfg = PipelineConfig(
            model_type="callable_resolved_pipeline",
            hf_architectures=("CallableArchitectureForTest",),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        seen_hf_configs = []

@@ -918,7 +940,10 @@ class TestPipelineRegistration:
    def test_pipeline_registration(self, clean_pipeline_registry):
        """Ensure that we can register and create a custom pipeline config."""
        new_model_type = "new_model_type"
-        pipe_cfg = PipelineConfig(model_type=new_model_type)
+        pipe_cfg = PipelineConfig(
+            model_type=new_model_type,
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )

        # Register the new PipelineConfig
        assert new_model_type not in OMNI_PIPELINES
@@ -961,7 +986,10 @@ class TestPipelineRegistration:
        def custom_resolver(
            hf_config: FakeConfig,
        ) -> PipelineConfig:
-            return PipelineConfig(model_type=resolved_type)
+            return PipelineConfig(
+                model_type=resolved_type,
+                stages=SINGLE_STAGE_PIPE_CFG,
+            )

        # Register the new PipelineConfig
        assert new_model_type not in OMNI_PIPELINES
@@ -1018,7 +1046,7 @@ class TestPipelineRegistration:
        pipe_cfg = PipelineConfig(
            model_type=pipeline_key,
            model_arch="DeployOnlyArch",
-            stages=(StagePipelineConfig(stage_id=0, model_stage="diffusion"),),
+            stages=SINGLE_STAGE_PIPE_CFG,
        )
        register_pipeline(pipe_cfg)

@@ -1051,7 +1079,11 @@ class TestPipelineRegistration:

    def test_structured_path_loads_explicit_deploy_config_once(self, clean_pipeline_registry, tmp_path):
        pipeline_key = "single_load_pipeline"
-        register_pipeline(PipelineConfig(model_type=pipeline_key))
+        pipeline_cfg = PipelineConfig(
+            model_type=pipeline_key,
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
+        register_pipeline(pipeline_cfg)
        deploy_path = tmp_path / "single_load.yaml"
        deploy_path.write_text(f"pipeline: {pipeline_key}\n", encoding="utf-8")

@@ -1081,6 +1113,7 @@ class TestPipelineRegistration:
        pipeline = PipelineConfig(
            model_type="pipeline_with_default",
            default_deploy_config_name=default_name,
+            stages=SINGLE_STAGE_PIPE_CFG,
        )

        with patch(
@@ -1100,8 +1133,16 @@ class TestPipelineRegistration:
            OmniServingCapability.COMPLETIONS,
            "pipeline_a blocks completions",
        )
-        pipe_a = PipelineConfig(model_type="detect_type", endpoint_restrictions=(restriction,))
-        pipe_b = PipelineConfig(model_type="override_type", endpoint_restrictions=())
+        pipe_a = PipelineConfig(
+            model_type="detect_type",
+            endpoint_restrictions=(restriction,),
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
+        pipe_b = PipelineConfig(
+            model_type="override_type",
+            endpoint_restrictions=(),
+            stages=SINGLE_STAGE_PIPE_CFG,
+        )
        register_pipeline(pipe_a)
        register_pipeline(pipe_b)

@@ -1641,6 +1682,7 @@ stages:
                    model_stage="ar",
                    execution_type=StageExecutionType.LLM_AR,
                    requires_multimodal_data=True,
+                    final_output=True,
                ),
            ),
        )
@@ -1661,6 +1703,7 @@ stages:
                    model_stage="ar",
                    execution_type=StageExecutionType.LLM_AR,
                    scheduler_cls=scheduler_cls,
+                    final_output=True,
                ),
            ),
        )
@@ -1906,7 +1949,6 @@ class TestQwen3OmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "Qwen3OmniMoeForConditionalGeneration"
        assert len(p.stages) == 3
-        assert p.validate() == []

    def test_thinker(self):
        p = resolve_pipeline_config(
@@ -1960,7 +2002,6 @@ class TestQwen2_5OmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "Qwen2_5OmniForConditionalGeneration"
        assert len(p.stages) == 3
-        assert p.validate() == []

    def test_thinker(self):
        p = resolve_pipeline_config("qwen2_5_omni")
@@ -2009,7 +2050,6 @@ class TestQwen3TTSPipeline:
        assert p is not None
        assert p.model_arch == "Qwen3TTSTalkerForConditionalGeneration"
        assert len(p.stages) == 2
-        assert p.validate() == []

    def test_talker_stage(self):
        p = resolve_pipeline_config("qwen3_tts")
@@ -2089,7 +2129,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 2
-        assert p.validate() == []

    def test_thinker_stage(self):
        p = resolve_pipeline_config("ming_flash_omni")
@@ -2146,7 +2185,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniTalkerForConditionalGeneration"
        assert len(p.stages) == 1
-        assert p.validate() == []

    def test_tts_stage(self):
        p = resolve_pipeline_config("ming_flash_omni_tts")
@@ -2203,7 +2241,6 @@ class TestMingFlashOmniPipeline:
        assert isinstance(p, PipelineConfig)
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 1
-        assert p.validate() == []

    def test_thinker_only_stage(self):
        p = resolve_pipeline_config("ming_flash_omni_thinker_only")
@@ -2242,7 +2279,6 @@ class TestMingFlashOmniPipeline:
        assert p is not None
        assert p.model_arch == "MingFlashOmniForConditionalGeneration"
        assert len(p.stages) == 2
-        assert p.validate() == []

    def test_image_thinker_stage(self):
        s = resolve_pipeline_config("ming_flash_omni_image").get_stage(0)
@@ -2363,6 +2399,21 @@ class TestBaseConfigInheritance:
        # CI overrides max_tokens
        assert s0["max_tokens"] == 150

+    def test_qwen3_omni_colocate_async_bounds_only_rocm_kv_cache(self):
+        ci_path = Path(get_deploy_config_path("ci/qwen3_omni_moe_colocate_async.yaml"))
+        pipeline = resolve_pipeline_config("qwen3_omni_moe_thinker_only")
+        assert isinstance(pipeline, PipelineConfig)
+
+        cuda = _apply_platform_overrides(load_deploy_config(ci_path), platform="cuda")
+        cuda_stage = merge_pipeline_deploy(pipeline, cuda)[0]
+        assert cuda_stage.yaml_engine_args["gpu_memory_utilization"] == 0.9
+        assert "kv_cache_memory_bytes" not in cuda_stage.yaml_engine_args
+
+        rocm = _apply_platform_overrides(load_deploy_config(ci_path), platform="rocm")
+        rocm_stage = merge_pipeline_deploy(pipeline, rocm)[0]
+        assert "gpu_memory_utilization" not in rocm_stage.yaml_engine_args
+        assert rocm_stage.yaml_engine_args["kv_cache_memory_bytes"] == 2 * 1024**3
+
    def test_pure_inheritance_overlay(self, tmp_path):
        """An overlay with only ``base_config`` inherits everything."""
        base = Path(get_deploy_config_path("qwen3_omni_moe.yaml"))
@@ -2824,6 +2875,54 @@ stages:
        assert omega_config.engine_args.max_model_len == 8192
        assert omega_config.engine_args.enforce_eager is True

+    def test_cli_attention_shorthand_replaces_yaml_structured_default(self):
+        stage = StageConfig(
+            stage_id=0,
+            model_stage="dit",
+            stage_type=StageType.DIFFUSION,
+            yaml_engine_args={
+                "diffusion_attention_config": {
+                    "default": {"backend": "FLASH_ATTN"},
+                    "per_role": {"cross": {"backend": "TORCH_SDPA"}},
+                },
+            },
+            runtime_overrides={"diffusion_attention_backend": "SAGE_ATTN"},
+        )
+
+        engine_args = stage.to_omegaconf().engine_args
+
+        assert engine_args.diffusion_attention_backend == "SAGE_ATTN"
+        assert "default" not in engine_args.diffusion_attention_config
+        assert engine_args.diffusion_attention_config.per_role.cross.backend == "TORCH_SDPA"
+
+    def test_cli_attention_shorthand_keeps_yaml_per_role_only_config(self):
+        stage = StageConfig(
+            stage_id=0,
+            model_stage="dit",
+            stage_type=StageType.DIFFUSION,
+            yaml_engine_args={"diffusion_attention_config": {"per_role": {"cross": {"backend": "TORCH_SDPA"}}}},
+            runtime_overrides={"diffusion_attention_backend": "SAGE_ATTN"},
+        )
+
+        engine_args = stage.to_omegaconf().engine_args
+
+        assert engine_args.diffusion_attention_backend == "SAGE_ATTN"
+        assert engine_args.diffusion_attention_config == {"per_role": {"cross": {"backend": "TORCH_SDPA"}}}
+
+    def test_cli_attention_config_replaces_yaml_shorthand(self):
+        stage = StageConfig(
+            stage_id=0,
+            model_stage="dit",
+            stage_type=StageType.DIFFUSION,
+            yaml_engine_args={"diffusion_attention_backend": "TORCH_SDPA"},
+            runtime_overrides={"diffusion_attention_config": {"default": {"backend": "SAGE_ATTN"}}},
+        )
+
+        engine_args = stage.to_omegaconf().engine_args
+
+        assert "diffusion_attention_backend" not in engine_args
+        assert engine_args.diffusion_attention_config.default.backend == "SAGE_ATTN"
+

 class TestSentinelDefaultPrecedence:
    """Caller-typed (non-None) values win over YAML; None values fall through
@@ -167,7 +167,7 @@ def test_inventory_matches_reviewed_snapshot_counts():
        EnvironmentVariableCategory.PUBLIC_OMNI: 24,
        EnvironmentVariableCategory.INHERITED_VLLM: 20,
        EnvironmentVariableCategory.PLATFORM_EXTERNAL: 27,
-        EnvironmentVariableCategory.MODEL_SPECIFIC: 56,
+        EnvironmentVariableCategory.MODEL_SPECIFIC: 57,
        EnvironmentVariableCategory.BENCHMARK_TRANSITIONAL: 20,
        EnvironmentVariableCategory.INTERNAL: 2,
    }
@@ -178,7 +178,7 @@ def test_inventory_matches_reviewed_snapshot_counts():
        if item.category is EnvironmentVariableCategory.MODEL_SPECIFIC
    )
    assert {disposition: disposition_counts[disposition] for disposition in ModelEnvironmentVariableDisposition} == {
-        ModelEnvironmentVariableDisposition.PROMOTE: 33,
+        ModelEnvironmentVariableDisposition.PROMOTE: 34,
        ModelEnvironmentVariableDisposition.REQUEST_SCOPE: 6,
        ModelEnvironmentVariableDisposition.EXTERNAL: 0,
        ModelEnvironmentVariableDisposition.INTERNALIZE: 11,
@@ -45,6 +45,7 @@ from vllm_omni.config.stage_config import (
    PipelineConfig,
    StageDeployConfig,
    StageExecutionType,
+    StagePipelineConfig,
    load_deploy_config,
    merge_pipeline_deploy,
 )
@@ -1039,7 +1040,10 @@ def test_from_pipeline_config_uses_hf_config_for_callable_resolver():


 def test_from_pipeline_config_accepts_pre_resolved_pipeline():
-    resolved_pipeline = PipelineConfig(model_type="callable_resolved_variant")
+    resolved_pipeline = PipelineConfig(
+        model_type="callable_resolved_variant",
+        stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),),
+    )

    omni_config = VllmOmniConfig.from_pipeline_config(resolved_pipeline)

@@ -1072,6 +1076,7 @@ def test_from_pipeline_config_default_deploy_name_ignores_cwd(monkeypatch, tmp_p
    pipeline = PipelineConfig(
        model_type="pipeline_with_default",
        default_deploy_config_name=default_name,
+        stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),),
    )
    loaded_paths = []

@@ -1,16 +1,22 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for out of tree registration to OMNI_PIPELINES."""

 import pytest
 from transformers import PretrainedConfig

 from vllm_omni.config.pipeline_registry import OMNI_PIPELINES, register_pipeline
-from vllm_omni.config.stage_config import PipelineConfig, pipeline_cfg_resolver
+from vllm_omni.config.stage_config import PipelineConfig, StagePipelineConfig, pipeline_cfg_resolver

 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]


+def build_fake_pipeline_config(model_type: str) -> PipelineConfig:
+    return PipelineConfig(
+        model_type=model_type, stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
+    )
+
+
@pytest.fixture
 def custom_resolver():
    """Build a reusable custom resolver for PipelineConfigs."""
@@ -22,7 +28,7 @@ def custom_resolver():
    def custom_resolver(
        hf_config: CustomConfigType,
    ) -> PipelineConfig:
-        return PipelineConfig(model_type="resolved_type")
+        return build_fake_pipeline_config("resolved_type")

    return custom_resolver

@@ -30,7 +36,7 @@ def custom_resolver():
 def test_register_pipeline_config(clean_pipeline_registry):
    """Ensure that we can register a custom pipeline config to OMNI_PIPELINES."""
    new_model_type = "new_model_type"
-    pipe_cfg = PipelineConfig(model_type=new_model_type)
+    pipe_cfg = build_fake_pipeline_config(new_model_type)
    assert new_model_type not in OMNI_PIPELINES
    register_pipeline(pipe_cfg)
    assert new_model_type in OMNI_PIPELINES
@@ -41,7 +47,7 @@ def test_register_pipeline_config_with_model_type(clean_pipeline_registry):
    """Ensure that we can register a custom pipeline config with an explicit model_type to OMNI_PIPELINES."""
    new_model_type = "new_model_type"
    unused_model_type = "foo"
-    pipe_cfg = PipelineConfig(model_type=unused_model_type)
+    pipe_cfg = build_fake_pipeline_config(unused_model_type)
    assert new_model_type not in OMNI_PIPELINES
    assert unused_model_type not in OMNI_PIPELINES

@@ -73,20 +79,3 @@ def test_minimax_h3_disaggregation_is_explicit_opt_in():
    pipeline = OMNI_PIPELINES["minimax_h3_disaggregated"]
    assert isinstance(pipeline, PipelineConfig)
    assert pipeline.model_type == "minimax_h3_disaggregated"
-
-
-def test_in_tree_pipelines_pass_topology_validation():
-    """Every statically registered in-tree pipeline must have a valid topology.
-
-    ``PipelineConfig.validate()`` only runs on ``register_pipeline``, which
-    in-tree pipelines never go through, so a shipped topology error would
-    otherwise stay invisible until serving hangs or misroutes.
-    Resolver entries are skipped: they need an ``hf_config`` to produce a
-    ``PipelineConfig``.
-    """
-    failures = {
-        model_type: errors
-        for model_type, pipeline in OMNI_PIPELINES.items()
-        if isinstance(pipeline, PipelineConfig) and (errors := pipeline.validate())
-    }
-    assert not failures, f"in-tree pipelines failed topology validation: {failures}"
@@ -7,7 +7,7 @@ End-to-end test for diffusion batching via AsyncOmni.
 This test fires multiple concurrent ``AsyncOmni.generate()`` calls for a
 diffusion model and validates that every caller receives its correct
 individual result.  When the underlying diffusion stage is configured with
-``batch_size > 1`` (via stage config or ``StageDiffusionClient``), the
+``max_num_seqs > 1``, the
 requests will be batched internally.

 Even without explicit batching config this test is useful for verifying
@@ -260,7 +260,7 @@ async def compare_single_vs_parallel(
 ) -> None:
    """Run the same prompts sequentially then in parallel and print a comparison."""

-    omni = AsyncOmni(model=model, max_num_seqs=batch_size, diffusion_batch_size=batch_size)
+    omni = AsyncOmni(model=model, max_num_seqs=batch_size)
    try:
        await warmup(omni, WARMUP_PROMPTS)
        single_time = await run_single(omni, prompts)
@@ -288,7 +288,7 @@ async def main(model: str, num_prompts: int, mode: str, batch_size: int = 1) ->
        await compare_single_vs_parallel(model, prompts, batch_size=batch_size)
        return

-    omni = AsyncOmni(model=model, max_num_seqs=batch_size, diffusion_batch_size=batch_size)
+    omni = AsyncOmni(model=model, max_num_seqs=batch_size)
    try:
        await warmup(omni, WARMUP_PROMPTS)

@@ -389,7 +389,7 @@ def test_diffusion_batching_async_concurrent(model_name: str):
    fired via asyncio.gather. Each request_id must appear in the results."""

    async def _inner():
-        omni = AsyncOmni(model=model_name, diffusion_batch_size=1)
+        omni = AsyncOmni(model=model_name, max_num_seqs=1)
        try:
            prompts = TEST_PROMPTS[:4]
            sp = _default_sampling_params()
@@ -430,7 +430,7 @@ def test_diffusion_batching_list_prompt_rejected(model_name: str):
    """

    async def _inner():
-        omni = AsyncOmni(model=model_name, diffusion_batch_size=4)
+        omni = AsyncOmni(model=model_name, max_num_seqs=4)
        try:
            prompts = TEST_PROMPTS[:4]
            sp = _default_sampling_params()
@@ -528,7 +528,7 @@ if __name__ == "__main__":
        "--batch-size",
        type=int,
        default=1,
-        help="Scheduler in-flight width (--max-num-seqs) and client diffusion_batch_size",
+        help="Scheduler in-flight width (--max-num-seqs)",
    )
    parser.add_argument(
        "--mode",
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 from collections import defaultdict

 import pytest
@@ -106,7 +109,7 @@ def make_dummy_lora_state_dict(
    """
    key_a = f"{base_key}.{lora_a_suffix}"
    key_b = f"{base_key}.{lora_b_suffix}"
-    key_bias = f"{base_key}.{lora_bias_suffix}" if bias else None
+    key_bias = f"{base_key}.{lora_bias_suffix}"

    if weight_a is None:
        weight_a = torch.randn(RANK, HEAD_DIM)
@@ -436,3 +439,54 @@ def test_wan_load_lora_weights_targets_both_transformers(mocker: MockerFixture):
    assert_params_equal(pipeline.transformer, original_high)
    assert_params_equal(pipeline.transformer_2, original_low)
    assert "adapter0" not in pipeline.lora_loaded
+
+
+def test_prepare_lora_delta_disambiguates_prefix_overlapping_stacked_names():
+    """A fused name can be a prefix of another fused name.
+
+    SenseNova-U1 carries both `.qkv_proj` (understanding) and `.qkv_proj_mot_gen`
+    (generation). Selecting stacked mappings by substring matched both rules for
+    the generation weight and concatenated six shards into a delta twice as tall
+    as the parameter.
+    """
+    param_to_weight_names = {
+        ".qkv_proj_mot_gen": [".q_proj_mot_gen", ".k_proj_mot_gen", ".v_proj_mot_gen"],
+        ".qkv_proj": [".q_proj", ".k_proj", ".v_proj"],
+    }
+    state_dict = {}
+    for part in ("q", "k", "v"):
+        for tower in ("_mot_gen", ""):
+            base = f"layers.0.self_attn.{part}_proj{tower}"
+            state_dict[f"{base}.lora_A.weight"] = torch.ones(RANK, HEAD_DIM)
+            state_dict[f"{base}.lora_B.weight"] = torch.ones(HEAD_DIM, RANK)
+
+    for base_key in ("layers.0.self_attn.qkv_proj_mot_gen", "layers.0.self_attn.qkv_proj"):
+        delta, _ = _prepare_lora_delta(state_dict, base_key, param_to_weight_names)
+        assert delta.shape == (3 * HEAD_DIM, HEAD_DIM), base_key
+
+
+@pytest.mark.parametrize(
+    "fused, shards",
+    [
+        # qwen_image_transformer.py stacked_params_mapping
+        (".to_qkv", (".to_q", ".to_k", ".to_v")),
+        (".add_kv_proj", (".add_q_proj", ".add_k_proj", ".add_v_proj")),
+        # wan2_2_transformer.py stacked_params_mapping
+        (".attn1.to_qkv", (".attn1.to_q", ".attn1.to_k", ".attn1.to_v")),
+        # wan2_2_s2v_transformer.py stacked_params_mapping
+        (".self_attn.to_qkv", (".self_attn.q", ".self_attn.k", ".self_attn.v")),
+    ],
+)
+def test_prepare_lora_delta_unchanged_for_existing_fused_layouts(fused, shards):
+    """None of the fused names shipped today overlap by prefix, so matching on the
+    tail must select exactly the shards substring matching selected."""
+    param_to_weight_names = {fused: list(shards)}
+    base_key = f"blocks.0{fused}"
+    state_dict = {}
+    for shard in shards:
+        state_dict[f"blocks.0{shard}.lora_A.weight"] = torch.ones(RANK, HEAD_DIM)
+        state_dict[f"blocks.0{shard}.lora_B.weight"] = torch.ones(HEAD_DIM, RANK)
+
+    delta, used_keys = _prepare_lora_delta(state_dict, base_key, param_to_weight_names)
+    assert delta.shape == (len(shards) * HEAD_DIM, HEAD_DIM)
+    assert len(used_keys) == 2 * len(shards)
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """L1 unit tests for the native Boogu-Image pipeline.

@@ -133,6 +133,51 @@ def test_constructor_strips_mllm_lm_head(boogu_pipeline, mock_dependencies):
    assert boogu_pipeline.mllm is mock_dependencies["inner_encoder"]


+def test_constructor_forwards_revision_to_all_component_loaders(mock_dependencies, mocker):
+    from vllm_omni.diffusion.models.boogu_image.pipeline_boogu_image import (
+        BooguImagePipeline,
+    )
+
+    revision = "hotfix-1k-20260708"
+    prefetch = mocker.patch(f"{_MODULE}.prefetch_subfolders")
+    scheduler_loader = mocker.patch(
+        f"{_MODULE}.FlowMatchEulerDiscreteScheduler.from_pretrained",
+        return_value=mock_dependencies["scheduler"],
+    )
+    mllm_loader = mocker.patch(
+        f"{_MODULE}.Qwen3VLForConditionalGeneration.from_pretrained",
+        return_value=mock_dependencies["mllm_wrapper"],
+    )
+    processor_loader = mocker.patch(
+        f"{_MODULE}.Qwen3VLProcessor.from_pretrained",
+        return_value=mock_dependencies["processor"],
+    )
+    vae_loader = mocker.patch(
+        f"{_MODULE}.AutoencoderKL.from_pretrained",
+        return_value=mock_dependencies["vae"],
+    )
+    od_config = OmniDiffusionConfig(
+        model="dummy-boogu",
+        revision=revision,
+        tf_model_config=TransformerConfig(params={}),
+        dtype=torch.float32,
+        num_gpus=1,
+    )
+
+    pipeline = BooguImagePipeline(od_config=od_config)
+
+    (source,) = pipeline.weights_sources
+    assert source.revision == revision
+    prefetch.assert_called_once_with(
+        "dummy-boogu",
+        ["scheduler", "vae", "mllm", "processor"],
+        local_files_only=True,
+        revision=revision,
+    )
+    for loader in (scheduler_loader, mllm_loader, processor_loader, vae_loader):
+        assert loader.call_args.kwargs["revision"] == revision
+
+
 def test_constructor_weights_sources(boogu_pipeline):
    (source,) = boogu_pipeline.weights_sources
    assert source.model_or_path == "dummy-boogu"
@@ -0,0 +1,195 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 attention head handling.
+
+The model used to expand K/V from 8 heads to 32 with ``repeat_interleave``
+before calling the attention backend, so the backend's own GQA dispatch never
+saw a GQA shape, and it passed an all-zeros mask during single-token decode,
+which only served to keep SDPA off its fused kernels. These pin both contracts:
+K/V must reach the backend uncompressed, decode must carry no mask, and the
+compressed path must compute the same thing the expanded one did.
+"""
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.attention.backends.sdpa import SDPAImpl
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import (
+    SenseNovaU1Attention,
+    SenseNovaU1Model,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+N_HEADS, N_KV_HEADS, HEAD_DIM, SEQ = 32, 8, 8, 6
+GROUPS = N_HEADS // N_KV_HEADS
+
+
+class _Recorder:
+    """Stands in for the Attention module and records what it was handed."""
+
+    def __init__(self):
+        self.calls: list[tuple[torch.Size, torch.Size, torch.Tensor | None]] = []
+
+    def __call__(self, query, key, value, attn_metadata):
+        mask = None if attn_metadata is None else attn_metadata.attn_mask
+        self.calls.append((query.shape, key.shape, mask))
+        return query
+
+
+class _AttnHost:
+    """Carries only the attributes ``_run_attn``/``_run_attn_bshd`` touch."""
+
+    num_kv_groups = GROUPS
+    _align_mask_dtype = staticmethod(SenseNovaU1Attention._align_mask_dtype)
+    _run_attn = SenseNovaU1Attention._run_attn
+    _run_attn_bshd = SenseNovaU1Attention._run_attn_bshd
+
+    def __init__(self):
+        self.attn = _Recorder()
+
+
+def test_bhsd_path_hands_the_backend_compressed_kv():
+    host = _AttnHost()
+    q = torch.randn(1, N_HEADS, SEQ, HEAD_DIM)
+    k = torch.randn(1, N_KV_HEADS, SEQ, HEAD_DIM)
+    host._run_attn(q, k, k.clone(), None)
+    _, key_shape, _ = host.attn.calls[-1]
+    # [B, S, H, D] after the transpose, so heads sit on dim 2.
+    assert key_shape[2] == N_KV_HEADS, f"K reached the backend with {key_shape[2]} heads"
+
+
+def test_bshd_path_hands_the_backend_compressed_kv():
+    host = _AttnHost()
+    q = torch.randn(1, SEQ, N_HEADS, HEAD_DIM)
+    k = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    host._run_attn_bshd(q, k, k.clone(), None)
+    _, key_shape, _ = host.attn.calls[-1]
+    assert key_shape[2] == N_KV_HEADS, f"K reached the backend with {key_shape[2]} heads"
+
+
+def test_compressed_kv_computes_what_expansion_computed():
+    """The backend must return the same tensor either way, or this is a bugfix
+    dressed up as a perf change."""
+    torch.manual_seed(0)
+    q = torch.randn(1, SEQ, N_HEADS, HEAD_DIM)
+    k = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    v = torch.randn(1, SEQ, N_KV_HEADS, HEAD_DIM)
+    impl = SDPAImpl(
+        num_heads=N_HEADS,
+        head_size=HEAD_DIM,
+        softmax_scale=HEAD_DIM**-0.5,
+        num_kv_heads=N_KV_HEADS,
+    )
+    compressed = impl._forward_impl(q, k, v)
+    expanded = impl._forward_impl(q, k.repeat_interleave(GROUPS, dim=2), v.repeat_interleave(GROUPS, dim=2))
+    torch.testing.assert_close(compressed, expanded, atol=1e-6, rtol=1e-6)
+
+
+class _MaskProbe(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.masks: list[torch.Tensor | None] = []
+
+    def forward(self, hidden_states, **kwargs):
+        self.masks.append(kwargs["attention_mask"]["full_attention"])
+        return hidden_states
+
+
+def _fake_rope(hidden_states, position_ids):
+    """The forward builds the three RoPE tables before the layer loop; the probe
+    ignores them, so any 2-tuple will do."""
+    return hidden_states, hidden_states
+
+
+class _ModelHost:
+    """Carries only the attributes ``SenseNovaU1Model.forward`` touches."""
+
+    forward = SenseNovaU1Model.forward
+
+    def __init__(self, probe):
+        self.layers = [probe]
+        self.norm = torch.nn.Identity()
+        self.norm_mot_gen = torch.nn.Identity()
+        self.rotary_emb = _fake_rope
+        self.rotary_emb_hw = _fake_rope
+
+
+def _run_model(seq_len: int):
+    probe = _MaskProbe()
+    _ModelHost(probe).forward(
+        inputs_embeds=torch.zeros(1, seq_len, 4),
+        indexes=torch.zeros(3, seq_len, dtype=torch.long),
+    )
+    return probe.masks[-1]
+
+
+def test_single_token_decode_carries_no_mask():
+    assert _run_model(1) is None, "decode still builds the all-zeros mask"
+
+
+def test_prefill_still_gets_a_causal_mask():
+    mask = _run_model(3)
+    assert mask is not None and mask.shape == (1, 1, 3, 3)
+    assert mask[0, 0, 0, 0] == 0.0
+    assert torch.isinf(mask[0, 0, 0, 1]) and mask[0, 0, 0, 1] < 0
+    assert mask[0, 0, 2, 1] == 0.0
+
+
+# ---------------------------------------------------------------------------
+# The decode path changes which SDPA kernel runs, so it is not bit-identical to
+# what it replaced. "Different" would be a regression if the new result were
+# worse, so pin the direction against an independent float64 reference. This
+# needs a fused GQA kernel, which only exists on CUDA.
+# ---------------------------------------------------------------------------
+
+
+def _float64_attention(q, k, v):
+    q64 = q.double()
+    k64 = k.double().repeat_interleave(GROUPS, dim=1)
+    v64 = v.double().repeat_interleave(GROUPS, dim=1)
+    scores = (q64 @ k64.transpose(-1, -2)) * (q.shape[-1] ** -0.5)
+    return torch.softmax(scores, dim=-1) @ v64
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required for fused GQA")
+@pytest.mark.cuda
+@pytest.mark.L4
+@pytest.mark.parametrize("kv_len", [271, 512, 2048])
+def test_decode_without_the_no_op_mask_is_closer_to_float64(kv_len):
+    """The old decode path expanded K/V and passed an all-zeros mask, which kept
+    SDPA off its fused kernels. Dropping the mask is only defensible if the
+    kernel it unlocks is at least as accurate."""
+    dev, dtype = "cuda", torch.bfloat16
+    heads, dim = 32, 128
+    new_errs: list[float] = []
+    old_errs: list[float] = []
+    trials = 8
+    for t in range(trials):
+        torch.manual_seed(2000 + t)
+        q = torch.randn(1, heads, 1, dim, device=dev, dtype=dtype)
+        k = torch.randn(1, heads // GROUPS, kv_len, dim, device=dev, dtype=dtype)
+        v = torch.randn(1, heads // GROUPS, kv_len, dim, device=dev, dtype=dtype)
+        ref = _float64_attention(q, k, v)
+
+        zeros_mask = torch.zeros(1, 1, 1, kv_len, device=dev, dtype=dtype)
+        old = torch.nn.functional.scaled_dot_product_attention(
+            q,
+            k.repeat_interleave(GROUPS, dim=1),
+            v.repeat_interleave(GROUPS, dim=1),
+            attn_mask=zeros_mask,
+        )
+        new = torch.nn.functional.scaled_dot_product_attention(q, k, v, enable_gqa=True)
+
+        new_errs.append((new.double() - ref).abs().mean().item())
+        old_errs.append((old.double() - ref).abs().mean().item())
+
+    # A single trial can tie either way depending on the GPU and the CUDA build,
+    # so gate on the aggregate: the maskless path must not be worse on average.
+    new_mean = sum(new_errs) / trials
+    old_mean = sum(old_errs) / trials
+    wins = sum(n <= o for n, o in zip(new_errs, old_errs))
+    assert new_mean <= old_mean, (
+        f"kv_len={kv_len}: maskless mean error {new_mean:.4e} is worse than masked {old_mean:.4e} "
+        f"({wins}/{trials} per-trial wins)"
+    )
@@ -0,0 +1,196 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 distilled-LoRA fusion.
+
+The checkpoints ship unfused ``q_proj_mot_gen`` / ``gate_proj`` keys while the
+pipeline runs fused layers, so each delta lands in its own row slice.
+"""
+
+import pytest
+import torch
+import torch.nn as nn
+from safetensors.torch import save_file
+
+from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import SenseNovaU1Pipeline
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+HIDDEN, KV, INTER, RANK, ALPHA = 8, 4, 16, 2, 4.0
+
+
+class _Layer(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.self_attn = nn.Module()
+        self.self_attn.qkv_proj_mot_gen = nn.Linear(HIDDEN, HIDDEN + 2 * KV, bias=False)
+        self.self_attn.o_proj_mot_gen = nn.Linear(HIDDEN, HIDDEN, bias=False)
+        self.mlp_mot_gen = nn.Module()
+        self.mlp_mot_gen.gate_up_proj = nn.Linear(HIDDEN, 2 * INTER, bias=False)
+        self.mlp_mot_gen.down_proj = nn.Linear(INTER, HIDDEN, bias=False)
+
+
+class _Stub(SenseNovaU1Pipeline):
+    """Only the parameter tree matters here, so skip the real __init__."""
+
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.language_model = nn.Module()
+        self.language_model.model = nn.Module()
+        self.language_model.model.layers = nn.ModuleList([_Layer()])
+
+
+def _triple(out: int, in_: int, fill: float) -> dict[str, torch.Tensor]:
+    return {
+        "lora_down.weight": torch.full((RANK, in_), fill, dtype=torch.float32),
+        "lora_up.weight": torch.full((out, RANK), 1.0, dtype=torch.float32),
+        "alpha": torch.tensor(ALPHA),
+    }
+
+
+def _write(tmp_path, entries: dict[str, dict[str, torch.Tensor]]):
+    flat = {f"{mod}.{leaf}": t for mod, parts in entries.items() for leaf, t in parts.items()}
+    path = tmp_path / "lora.safetensors"
+    save_file(flat, str(path))
+    return str(path)
+
+
+def _prefix(i: int = 0) -> str:
+    return f"language_model.model.layers.{i}"
+
+
+def test_qkv_shards_land_in_their_own_row_slices(tmp_path):
+    pipe = _Stub()
+    qkv = pipe.language_model.model.layers[0].self_attn.qkv_proj_mot_gen.weight
+    with torch.no_grad():
+        qkv.zero_()
+    entries = {
+        f"{_prefix()}.self_attn.q_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0),
+        f"{_prefix()}.self_attn.k_proj_mot_gen": _triple(KV, HIDDEN, 2.0),
+        f"{_prefix()}.self_attn.v_proj_mot_gen": _triple(KV, HIDDEN, 3.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    scale = ALPHA / RANK
+    assert torch.allclose(qkv[:HIDDEN], torch.full((HIDDEN, HIDDEN), scale * RANK * 1.0))
+    assert torch.allclose(qkv[HIDDEN : HIDDEN + KV], torch.full((KV, HIDDEN), scale * RANK * 2.0))
+    assert torch.allclose(qkv[HIDDEN + KV :], torch.full((KV, HIDDEN), scale * RANK * 3.0))
+
+
+def test_gate_and_up_land_in_their_own_halves(tmp_path):
+    pipe = _Stub()
+    gu = pipe.language_model.model.layers[0].mlp_mot_gen.gate_up_proj.weight
+    with torch.no_grad():
+        gu.zero_()
+    entries = {
+        f"{_prefix()}.mlp_mot_gen.gate_proj": _triple(INTER, HIDDEN, 1.0),
+        f"{_prefix()}.mlp_mot_gen.up_proj": _triple(INTER, HIDDEN, 5.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    scale = ALPHA / RANK
+    assert torch.allclose(gu[:INTER], torch.full((INTER, HIDDEN), scale * RANK * 1.0))
+    assert torch.allclose(gu[INTER:], torch.full((INTER, HIDDEN), scale * RANK * 5.0))
+
+
+def test_unfused_targets_are_added_whole(tmp_path):
+    pipe = _Stub()
+    o = pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight
+    before = o.detach().clone()
+    entries = {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)}
+    pipe.load_lora_weights(_write(tmp_path, entries))
+    assert torch.allclose(o - before, torch.full((HIDDEN, HIDDEN), (ALPHA / RANK) * RANK))
+
+
+def test_lora_that_matches_nothing_raises_instead_of_no_op(tmp_path):
+    """Without a loader the worker only warned and generated without the
+    adapter, so a wrong path produced a plausible image and no error."""
+    pipe = _Stub()
+    entries = {f"{_prefix()}.self_attn.does_not_exist": _triple(HIDDEN, HIDDEN, 1.0)}
+    with pytest.raises(ValueError, match="matched no parameter"):
+        pipe.load_lora_weights(_write(tmp_path, entries))
+
+
+def test_shape_mismatch_raises(tmp_path):
+    pipe = _Stub()
+    entries = {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN + 1, HIDDEN, 1.0)}
+    with pytest.raises(RuntimeError, match="must match the size"):
+        pipe.load_lora_weights(_write(tmp_path, entries))
+
+
+def test_multiple_files_rejected(tmp_path):
+    pipe = _Stub()
+    path = _write(tmp_path, {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)})
+    with pytest.raises(ValueError, match="exactly one"):
+        pipe.load_lora_weights([path, path])
+
+
+class _ShardedLayer(nn.Module):
+    """One rank of a two-way sharded fused QKV, with a vLLM-style weight loader.
+
+    The parameter holds half of each of q, k and v. The loader is handed the whole
+    layer's tensor and keeps only this rank's slice, which is what vLLM does.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.weight = nn.Parameter(torch.zeros(HIDDEN // 2 + KV, HIDDEN))
+        self.seen: list[tuple[str, tuple[int, ...]]] = []
+        self.weight.weight_loader = self._weight_loader
+
+    def _weight_loader(self, param, loaded_weight, shard_id):
+        self.seen.append((shard_id, tuple(loaded_weight.shape)))
+        sizes = {"q": HIDDEN // 2, "k": KV // 2, "v": KV // 2}
+        offsets = {"q": 0, "k": HIDDEN // 2, "v": HIDDEN // 2 + KV // 2}
+        rows = sizes[shard_id]
+        param.data[offsets[shard_id] : offsets[shard_id] + rows] = loaded_weight[:rows]
+
+
+class _ShardedStub(SenseNovaU1Pipeline):
+    def __init__(self):
+        nn.Module.__init__(self)
+        self.language_model = nn.Module()
+        self.language_model.model = nn.Module()
+        layer = nn.Module()
+        layer.self_attn = nn.Module()
+        layer.self_attn.qkv_proj_mot_gen = _ShardedLayer()
+        self.language_model.model.layers = nn.ModuleList([layer])
+
+
+def test_fused_delta_is_sharded_through_the_weight_loader(tmp_path):
+    """Under tensor parallelism the parameter is a rank-local shard while the
+    checkpoint delta spans the whole layer, so the delta has to go through the
+    layer's weight loader. Adding it directly raises a shape error."""
+    pipe = _ShardedStub()
+    layer = pipe.language_model.model.layers[0].self_attn.qkv_proj_mot_gen
+    entries = {
+        f"{_prefix()}.self_attn.q_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0),
+        f"{_prefix()}.self_attn.k_proj_mot_gen": _triple(KV, HIDDEN, 2.0),
+        f"{_prefix()}.self_attn.v_proj_mot_gen": _triple(KV, HIDDEN, 3.0),
+    }
+    pipe.load_lora_weights(_write(tmp_path, entries))
+
+    # the loader saw the whole layer's tensors, not a pre-sharded slice
+    assert layer.seen == [("q", (HIDDEN, HIDDEN)), ("k", (KV, HIDDEN)), ("v", (KV, HIDDEN))]
+
+    scale = ALPHA / RANK * RANK
+    got = layer.weight
+    assert torch.allclose(got[: HIDDEN // 2], torch.full((HIDDEN // 2, HIDDEN), scale * 1.0))
+    assert torch.allclose(got[HIDDEN // 2 : HIDDEN // 2 + KV // 2], torch.full((KV // 2, HIDDEN), scale * 2.0))
+    assert torch.allclose(got[HIDDEN // 2 + KV // 2 :], torch.full((KV // 2, HIDDEN), scale * 3.0))
+
+
+def test_fusion_does_not_retain_the_adapter_tensors(tmp_path):
+    """Fusion is one-way and this pipeline has no unload path, so holding the
+    fp32 state dict would keep about 1.5 GiB alive for the shipped checkpoint."""
+    pipe = _Stub()
+    path = _write(tmp_path, {f"{_prefix()}.self_attn.o_proj_mot_gen": _triple(HIDDEN, HIDDEN, 1.0)})
+    pipe.load_lora_weights(path, adapter_name="distill")
+
+    kept = pipe.lora_loaded["distill"]
+    assert not isinstance(kept, dict)
+    assert not isinstance(kept, torch.Tensor)
+
+    # the sentinel still makes a second load a no-op
+    before = pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight.detach().clone()
+    pipe.load_lora_weights(path, adapter_name="distill")
+    torch.testing.assert_close(pipe.language_model.model.layers[0].self_attn.o_proj_mot_gen.weight, before)
@@ -0,0 +1,800 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Tests for the paged decode cache and its CUDA graph.
+
+The cache exists so a decode step has static shapes and can be captured. That
+puts the burden on two things a normal test would not look at: the prefix must
+survive the hand-off from the ordinary cache, and every quantity that varies
+between steps must live in a device tensor. A Python-level index is correct
+eagerly and silently wrong once captured, because capture bakes it in -- which
+is exactly what the last test here pins.
+"""
+
+from types import SimpleNamespace
+
+import pytest
+import torch
+from vllm.lora.layers import BaseLayerWithLoRA
+
+import vllm_omni.diffusion.models.sensenova_u1.paged_decode as paged_decode
+from vllm_omni.diffusion.models.sensenova_u1.paged_decode import (
+    BLOCK_SIZE,
+    BUCKETS,
+    TAIL_STEP,
+    PagedDecodeCache,
+    _bucket_for,
+)
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import (
+    SenseNovaU1Attention,
+)
+from vllm_omni.platforms import current_omni_platform
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+LAYERS, KV_HEADS, HEAD_DIM = 2, 4, 8
+N_HEADS = KV_HEADS * 2
+
+
+class _Layer:
+    def __init__(self, keys=None, values=None):
+        self.keys, self.values = keys, values
+
+
+class _Cache:
+    """Stands in for the transformers cache: layers with [B, H, S, D] tensors."""
+
+    def __init__(self, layers):
+        self.layers = layers
+
+
+def _dyn_cache(prefix, device=None):
+    torch.manual_seed(0)
+    return _Cache(
+        [
+            _Layer(
+                torch.randn(1, KV_HEADS, prefix, HEAD_DIM, device=device),
+                torch.randn(1, KV_HEADS, prefix, HEAD_DIM, device=device),
+            )
+            for _ in range(LAYERS)
+        ]
+    )
+
+
+def test_bucket_rounds_up_and_is_block_aligned():
+    assert _bucket_for(1) == 512
+    assert _bucket_for(512) == 512
+    assert _bucket_for(513) == 1024
+    assert _bucket_for(100_000) % BLOCK_SIZE == 0
+
+
+def test_past_the_last_bucket_the_schedule_grows_in_steps():
+    """Every bucket change reallocates the cache and re-captures the graph, so
+    the tail of the schedule must not step by BLOCK_SIZE. A think edit with two
+    2048x2048 inputs runs to about 9,100 tokens, so this regime is reachable,
+    and a request decodes at most 1024 steps.
+    """
+    last = BUCKETS[-1]
+    assert _bucket_for(last + 1) == last + TAIL_STEP
+    assert _bucket_for(100_000) % BLOCK_SIZE == 0
+    for start in (last + 1, last + TAIL_STEP - 1, last + TAIL_STEP + 7):
+        bucket, grows = _bucket_for(start), 0
+        for n in range(start, start + 1025):
+            if _bucket_for(n) != bucket:
+                bucket = _bucket_for(n)
+                grows += 1
+        assert grows <= 1, f"start {start}: {grows} reallocations over 1024 decode steps"
+
+
+def test_prefix_survives_the_hand_off():
+    dyn = _dyn_cache(37)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    assert paged.length == 37
+    for i in range(LAYERS):
+        stored = paged.k[i].view(-1, KV_HEADS, HEAD_DIM)[:37]
+        torch.testing.assert_close(stored, dyn.layers[i].keys[0].transpose(0, 1))
+
+
+def test_write_back_restores_what_was_handed_over():
+    """Decode is paged but the stage after it reads the ordinary cache."""
+    dyn = _dyn_cache(21)
+    original = [layer.keys.clone() for layer in dyn.layers]
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    paged.to_dynamic_cache(dyn)
+    for before, layer in zip(original, dyn.layers):
+        torch.testing.assert_close(layer.keys, before)
+
+
+def test_growing_keeps_the_tokens_already_stored():
+    dyn = _dyn_cache(500)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    assert paged.bucket == 512
+    before = [k.view(-1, KV_HEADS, HEAD_DIM)[:500].clone() for k in paged.k]
+    gen = paged.generation
+    assert paged.grow(513) is True
+    assert paged.bucket == 1024
+    assert paged.generation > gen, "a reallocation must invalidate captured graphs"
+    for old, new in zip(before, paged.k):
+        torch.testing.assert_close(new.view(-1, KV_HEADS, HEAD_DIM)[:500], old)
+
+
+def test_a_failed_grow_leaves_the_cache_on_its_old_buffers(monkeypatch):
+    """A half-applied resize would strand a capture on freed storage.
+
+    ``generation`` is half of the key a captured graph is stored under, so
+    publishing new K without publishing V, the block table, the bucket and the
+    generation leaves the old key selecting a graph whose recorded addresses
+    have just been freed. A later request replays it. Injecting a failure into
+    the last allocation is the cheapest way to reach that window.
+    """
+    dyn = _dyn_cache(500)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    before_k, before_v = list(paged.k), list(paged.v)
+    before_table, before_bucket = paged.block_table, paged.bucket
+    before_generation, before_length = paged.generation, paged.length
+
+    real_zeros = torch.zeros
+    calls = []
+
+    def flaky(*args, **kwargs):
+        calls.append(1)
+        if len(calls) == 2 * LAYERS:  # the last of the two allocations per layer
+            raise RuntimeError("injected allocation failure")
+        return real_zeros(*args, **kwargs)
+
+    monkeypatch.setattr(torch, "zeros", flaky)
+    with pytest.raises(RuntimeError, match="injected"):
+        paged.grow(513)
+    monkeypatch.undo()
+
+    assert [id(t) for t in paged.k] == [id(t) for t in before_k], "K was published on its own"
+    assert [id(t) for t in paged.v] == [id(t) for t in before_v], "V was published on its own"
+    assert paged.block_table is before_table
+    assert paged.bucket == before_bucket
+    assert paged.generation == before_generation, "the graph key moved without new buffers"
+    assert paged.length == before_length
+
+
+def test_length_lives_in_device_tensors():
+    """Both the attended length and the write slot must be tensors.
+
+    A captured graph reads them at replay; a Python int would be frozen at
+    capture time and every replay would attend the wrong span and overwrite the
+    same slot.
+    """
+    dyn = _dyn_cache(10)
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, LAYERS, torch.device("cpu"), torch.float32)
+    paged.set_length(15)
+    assert torch.is_tensor(paged.seqused) and int(paged.seqused) == 15
+    assert torch.is_tensor(paged.pos) and int(paged.pos) == 14
+
+
+# ---------------------------------------------------------------------------
+# The capture contract. These need CUDA and the bundled flash-attention kernel.
+# ---------------------------------------------------------------------------
+
+cuda_only = pytest.mark.skipif(not torch.cuda.is_available(), reason="CUDA required")
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_paged_attention_matches_sdpa_over_the_same_buffer():
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 100
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    paged.set_length(n)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    scale = dim**-0.5
+    # write the last slot back as itself so attend() is a pure read
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v = torch.zeros_like(k)
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+    got = paged.attend(0, q, k, v, scale)
+    flat_k = paged.k[0].view(-1, kv_heads, dim)[:n]
+    flat_v = paged.v[0].view(-1, kv_heads, dim)[:n]
+    want = F.scaled_dot_product_attention(
+        q,
+        flat_k.unsqueeze(0).transpose(1, 2).contiguous(),
+        flat_v.unsqueeze(0).transpose(1, 2).contiguous(),
+        enable_gqa=True,
+        scale=scale,
+    ).transpose(1, 2)
+    torch.testing.assert_close(got.float(), want.float(), atol=2e-3, rtol=2e-3)
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_captured_graph_follows_a_later_length():
+    """One capture has to serve every length in the bucket.
+
+    If the attended span were a Python value the graph would keep attending the
+    capture-time length, and the run would still look fast.
+    """
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 200
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    v = torch.zeros_like(k)
+    scale = dim**-0.5
+
+    paged.set_length(n)
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+
+    side = torch.cuda.Stream()
+    side.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(side):
+        for _ in range(3):
+            paged.attend(0, q, k, v, scale)
+    torch.cuda.current_stream().wait_stream(side)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        out = paged.attend(0, q, k, v, scale)
+
+    shorter = n - 43
+    paged.set_length(shorter)
+    graph.replay()
+    torch.accelerator.synchronize()
+
+    flat_k = paged.k[0].view(-1, kv_heads, dim)[:shorter]
+    flat_v = paged.v[0].view(-1, kv_heads, dim)[:shorter]
+    want = F.scaled_dot_product_attention(
+        q,
+        flat_k.unsqueeze(0).transpose(1, 2).contiguous(),
+        flat_v.unsqueeze(0).transpose(1, 2).contiguous(),
+        enable_gqa=True,
+        scale=scale,
+    ).transpose(1, 2)
+    torch.testing.assert_close(out.float(), want.float(), atol=2e-3, rtol=2e-3)
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_captured_graph_writes_each_step_to_its_own_slot():
+    """The write slot has to be a device tensor, not a Python index.
+
+    ``test_a_captured_graph_follows_a_later_length`` pins the attended span.
+    The slot written is the other half: a Python index is frozen at capture, so
+    every replay would overwrite the capture-time slot and the cache would stop
+    growing -- correct eagerly, silently wrong once captured.
+    """
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim, n = 8, 2, 64, 100
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    torch.manual_seed(0)
+    dyn = _Cache(
+        [
+            _Layer(
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+                torch.randn(1, kv_heads, n, dim, device=dev, dtype=dt),
+            )
+        ]
+    )
+    paged = PagedDecodeCache.from_dynamic_cache(dyn, 1, dev, dt)
+    q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+    k = torch.zeros(1, kv_heads, 1, dim, device=dev, dtype=dt)
+    v = torch.zeros_like(k)
+    scale = dim**-0.5
+
+    # Capture on the step that writes slot n-1, restoring what is already there
+    # so the capture itself changes nothing.
+    paged.set_length(n)
+    k[0, :, 0] = dyn.layers[0].keys[0, :, n - 1]
+    v[0, :, 0] = dyn.layers[0].values[0, :, n - 1]
+    side = torch.cuda.Stream()
+    side.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(side):
+        for _ in range(3):
+            paged.attend(0, q, k, v, scale)
+    torch.cuda.current_stream().wait_stream(side)
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph):
+        paged.attend(0, q, k, v, scale)
+
+    kept = paged.k[0].view(-1, kv_heads, dim)[n - 1].clone()
+
+    # The next decode step: one token further, different K/V.
+    paged.set_length(n + 1)
+    k[0, :, 0] = 1.5
+    v[0, :, 0] = -2.5
+    graph.replay()
+    current_omni_platform.synchronize()
+
+    flat_k = paged.k[0].view(-1, kv_heads, dim)
+    flat_v = paged.v[0].view(-1, kv_heads, dim)
+    torch.testing.assert_close(flat_k[n].float(), torch.full_like(flat_k[n], 1.5).float())
+    torch.testing.assert_close(flat_v[n].float(), torch.full_like(flat_v[n], -2.5).float())
+    torch.testing.assert_close(flat_k[n - 1].float(), kept.float())
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_whole_decode_loop_matches_the_unpaged_path():
+    """Per-step kernel equivalence is not loop equivalence.
+
+    With the switch off a single-token step attends over a cache grown by
+    ``torch.cat`` and no mask; with it on the same step reads a bucket-sized
+    buffer through ``seqused_k``. Run the same sequence of steps through both,
+    crossing a bucket boundary on the way, and require the same argmax at every
+    step off a fixed projection -- which is the token id the pipeline would emit.
+    """
+    import torch.nn.functional as F
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import PagedDecodeCache
+
+    heads, kv_heads, dim = 8, 2, 64
+    prefill, steps = 500, 40  # crosses the 512 bucket at step 13
+    dev, dt = torch.device("cuda"), torch.bfloat16
+    scale = dim**-0.5
+    torch.manual_seed(7)
+
+    keys = torch.randn(1, kv_heads, prefill, dim, device=dev, dtype=dt)
+    values = torch.randn(1, kv_heads, prefill, dim, device=dev, dtype=dt)
+    paged = PagedDecodeCache.from_dynamic_cache(_Cache([_Layer(keys, values)]), 1, dev, dt)
+    ref_k, ref_v = keys.clone(), values.clone()
+    proj = torch.randn(dim * heads, 128, device=dev, dtype=torch.float32)
+
+    paged_ids, ref_ids, worst = [], [], 0.0
+    for step in range(steps):
+        q = torch.randn(1, heads, 1, dim, device=dev, dtype=dt)
+        k = torch.randn(1, kv_heads, 1, dim, device=dev, dtype=dt)
+        v = torch.randn(1, kv_heads, 1, dim, device=dev, dtype=dt)
+
+        if paged.length + 1 > paged.bucket:
+            paged.grow(paged.length + 1)
+        paged.set_length(paged.length + 1)
+        got = paged.attend(0, q, k, v, scale)
+
+        ref_k = torch.cat([ref_k, k], dim=2)
+        ref_v = torch.cat([ref_v, v], dim=2)
+        want = F.scaled_dot_product_attention(q, ref_k, ref_v, enable_gqa=True, scale=scale).transpose(1, 2)
+
+        worst = max(worst, (got.float() - want.float()).abs().max().item())
+        paged_ids.append(int((got.float().reshape(1, -1) @ proj).argmax()))
+        ref_ids.append(int((want.float().reshape(1, -1) @ proj).argmax()))
+
+    assert paged.bucket == 1024, "the loop never crossed a bucket boundary"
+    assert paged_ids == ref_ids, (
+        f"argmax diverges at step {next(i for i, (a, b) in enumerate(zip(paged_ids, ref_ids)) if a != b)}"
+    )
+    assert worst < 5e-2, f"largest per-step deviation {worst:.3e}"
+
+
+class _StubLM(torch.nn.Module):
+    """Just enough CUDA work inside the capture for a pool to be allocated."""
+
+    def __init__(self, device):
+        super().__init__()
+        self.w = torch.randn(16, 64, device=device)
+
+    def forward(self, input_ids, indexes, past_key_values=None, use_cache=False, paged_cache=None):
+        x = input_ids.float().expand(1, 16) @ self.w
+        return SimpleNamespace(logits=x.unsqueeze(0))
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_the_decode_graph_captures_into_the_shared_platform_pool(monkeypatch):
+    """``_decode_context`` builds a runner per request, so a private capture pool
+    hands every request its own arena and never gives it back. Measured on a
+    real model that was about 40 MiB of device memory per request, rising
+    linearly over twelve sequential think requests while the same twelve with
+    the switch off held flat.
+    """
+    from vllm.platforms import current_platform
+
+    from vllm_omni.diffusion.models.sensenova_u1.paged_decode import DecodeGraphRunner
+
+    pools = []
+    real_graph = torch.cuda.graph
+
+    def spy(graph, pool=None, **kwargs):
+        pools.append(pool)
+        return real_graph(graph, pool=pool, **kwargs)
+
+    monkeypatch.setattr(torch.cuda, "graph", spy)
+
+    dev = torch.device("cuda")
+    cache = PagedDecodeCache.from_dynamic_cache(_dyn_cache(8, dev), LAYERS, dev, torch.float32)
+    runner = DecodeGraphRunner(_StubLM(dev), cache, dev)
+    runner.step(0, 0)
+
+    assert pools, "no graph was captured"
+    assert pools[0] is not None, "captured into a fresh private pool"
+    assert pools[0] == current_platform.get_global_graph_pool()
+    assert not hasattr(runner, "_pool"), "the runner still keeps a pool of its own"
+
+
+# ---------------------------------------------------------------------------
+# The cache is only worth anything if production actually reaches it, and the
+# kwargs it needs are not in every wheel that exports the kernel. The tests
+# above would all still pass with the `forward_und` branch deleted, so these
+# drive the real dispatch instead of the cache object.
+# ---------------------------------------------------------------------------
+
+
+class _PagedProbe:
+    def __init__(self, out):
+        self.out = out
+        self.calls: list[tuple[int, torch.Size, float]] = []
+
+    def attend(self, layer_idx, query, key, value, scaling):
+        self.calls.append((layer_idx, query.shape, scaling))
+        return self.out
+
+
+class _AttnHost:
+    """Carries only what ``forward_und`` touches on the way to the branch."""
+
+    forward_und = SenseNovaU1Attention.forward_und
+    qkv_proj = q_norm = k_norm = q_norm_hw = k_norm_hw = None
+    layer_idx = 3
+    scaling = HEAD_DIM**-0.5
+
+    def __init__(self, seq, paged_out):
+        self._qkv = (
+            torch.randn(1, N_HEADS, seq, HEAD_DIM),
+            torch.randn(1, KV_HEADS, seq, HEAD_DIM),
+            torch.randn(1, KV_HEADS, seq, HEAD_DIM),
+        )
+        self.paged = _PagedProbe(paged_out)
+        self.sdpa_calls: list[torch.Tensor | None] = []
+
+    def _project_and_rope(self, *args, **kwargs):
+        return self._qkv
+
+    def _run_attn(self, query, key, value, mask):
+        self.sdpa_calls.append(mask)
+        return torch.zeros(1, N_HEADS, query.shape[2], HEAD_DIM)
+
+    def o_proj(self, hidden_states):
+        return hidden_states * 2, None
+
+
+def test_forward_und_sends_a_single_token_decode_to_the_paged_cache():
+    """Goes through the real dispatch, so deleting the branch turns this red."""
+    out = torch.randn(1, N_HEADS, 1, HEAD_DIM)
+    host = _AttnHost(seq=1, paged_out=out)
+    got = host.forward_und(torch.zeros(1, 1, 16), None, None, position_embeddings=None, paged_cache=host.paged)
+    assert len(host.paged.calls) == 1, "the paged hook was never reached"
+    assert host.sdpa_calls == [], "the same step also ran the SDPA path"
+    assert host.paged.calls[0][0] == host.layer_idx
+    torch.testing.assert_close(got, out.reshape(1, 1, -1).contiguous() * 2)
+
+
+@pytest.mark.parametrize(
+    "seq,mask",
+    [(2, None), (1, torch.zeros(1, 1, 1, 1))],
+    ids=["multi_token", "masked"],
+)
+def test_only_an_unmasked_single_token_step_takes_the_paged_path(seq, mask):
+    """Both other shapes are uncapturable, and the cache holds one row per step."""
+    host = _AttnHost(seq=seq, paged_out=torch.zeros(1, N_HEADS, seq, HEAD_DIM))
+    host.forward_und(torch.zeros(1, seq, 16), None, mask, position_embeddings=None, paged_cache=host.paged)
+    assert host.paged.calls == [], "a step that cannot be captured reached the paged cache"
+    assert len(host.sdpa_calls) == 1
+
+
+def test_a_kernel_missing_any_kwarg_we_pass_is_not_supported(monkeypatch):
+    """The call names every argument, so the probe has to cover every name.
+
+    An older wheel exports ``flash_attn_varlen_func`` without the paged kwargs,
+    and a future one may rename the standard ones. Either way the probe is what
+    keeps that install on SDPA instead of raising TypeError at the first decode
+    step, with the feature on by default.
+    """
+
+    def no_paged_kwargs(q, k, v, max_seqlen_q, cu_seqlens_q, max_seqlen_k, cu_seqlens_k, causal=False):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    def renamed_standard_args(q, k, v, max_q, cu_q, max_k, seqused_k=None, block_table=None, causal=False):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    for kernel in (no_paged_kwargs, renamed_standard_args):
+        monkeypatch.setattr(paged_decode, "_flash_varlen", lambda kernel=kernel: kernel)
+        assert paged_decode.paged_decode_supported(torch.device("cuda"), HEAD_DIM) is False
+
+    def current_kernel(
+        q,
+        k,
+        v,
+        max_seqlen_q,
+        cu_seqlens_q,
+        max_seqlen_k,
+        seqused_k=None,
+        block_table=None,
+        softmax_scale=None,
+        causal=False,
+    ):
+        raise AssertionError("must not be called")  # pragma: no cover
+
+    monkeypatch.setattr(paged_decode, "_flash_varlen", lambda: current_kernel)
+    assert paged_decode.paged_decode_supported(torch.device("cuda"), HEAD_DIM) is True
+
+
+def test_only_q_k_and_v_reach_the_kernel_positionally(monkeypatch):
+    """A positional standard argument binds to whatever a future wheel puts in
+    that slot, and the probe above only checks that the names exist."""
+    seen = {}
+
+    def recorder(*args, **kwargs):
+        seen["args"], seen["kwargs"] = args, kwargs
+        return torch.zeros(1, N_HEADS, HEAD_DIM)
+
+    monkeypatch.setattr(paged_decode, "_flash_varlen", lambda: recorder)
+    cache = PagedDecodeCache(1, KV_HEADS, HEAD_DIM, 8, torch.device("cpu"), torch.float32)
+    cache.set_length(4)
+    cache.attend(
+        0,
+        torch.zeros(1, N_HEADS, 1, HEAD_DIM),
+        torch.zeros(1, KV_HEADS, 1, HEAD_DIM),
+        torch.zeros(1, KV_HEADS, 1, HEAD_DIM),
+        1.0,
+    )
+    assert len(seen["args"]) == 3, "only q, k and v may be positional"
+    assert set(seen["kwargs"]) == {
+        "max_seqlen_q",
+        "cu_seqlens_q",
+        "max_seqlen_k",
+        "seqused_k",
+        "block_table",
+        "softmax_scale",
+        "causal",
+    }
+
+
+def test_the_kill_switch_returns_no_cache(monkeypatch):
+    """The recipe documents `VLLM_OMNI_SENSENOVA_PAGED_DECODE=0` as the way back
+    to the ordinary cache, so the switch needs a test rather than only prose."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+
+    monkeypatch.setenv("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "0")
+    # The env check returns before anything else on the pipeline is touched.
+    host = object.__new__(SenseNovaU1Pipeline)
+    assert SenseNovaU1Pipeline._decode_context(host, object()) is None
+
+
+def _pipeline_host(monkeypatch, device="cpu", language_model=None):
+    """A bare pipeline carrying only what ``_decode_context`` touches."""
+    from vllm_omni.diffusion.models.sensenova_u1 import pipeline_sensenova_u1 as pipe_mod
+
+    monkeypatch.setenv("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "1")
+    monkeypatch.setattr(pipe_mod, "paged_decode_supported", lambda dev, head_dim: True)
+    layer = SimpleNamespace(self_attn=SimpleNamespace(head_dim=HEAD_DIM))
+    host = object.__new__(pipe_mod.SenseNovaU1Pipeline)
+    # The pipeline is an nn.Module, and the language model assigned below is one
+    # too, so the bookkeeping has to exist before the assignment.
+    torch.nn.Module.__init__(host)
+    host.device = device
+    lm = language_model if language_model is not None else torch.nn.Module()
+    lm.model = SimpleNamespace(layers=[layer] * LAYERS)
+    host.language_model = lm
+    return pipe_mod, host
+
+
+def test_the_decode_context_is_reused_across_requests(monkeypatch):
+    """A capture binds the addresses it recorded, so a cache built per request
+    forces a capture per request. Two requests whose prefill fits the same
+    bucket have to come back with the same cache and the same runner, and with
+    the generation unchanged so the captures stay valid.
+    """
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    first = ctx(host, _dyn_cache(10))
+    assert first is not None
+    second_input = _dyn_cache(20)
+    second = ctx(host, second_input)
+    assert second[0] is first[0], "the cache was rebuilt for a request that fits it"
+    assert second[1] is first[1], "the runner was rebuilt, so its captures are gone"
+    assert second[0].generation == first[0].generation, "reuse must not invalidate a capture"
+    assert second[0].length == 20
+
+    # The reused buffers must carry the new request's prefill, not the old one.
+    flat = second[0].k[0].view(-1, KV_HEADS, HEAD_DIM)[:20]
+    torch.testing.assert_close(flat, second_input.layers[0].keys[0].transpose(0, 1))
+
+    # A prefill past the bucket cannot share the buffers.
+    bigger = ctx(host, _dyn_cache(BUCKETS[0] + 1))
+    assert bigger[0] is not first[0], "a prefill past the bucket must get new buffers"
+
+
+def test_releasing_the_captures_forces_a_rebuild(monkeypatch):
+    """Sleep level 2 discards the memory a capture recorded, so what the
+    pipeline reuses across requests has to be droppable."""
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    first = ctx(host, _dyn_cache(10))
+    assert ctx(host, _dyn_cache(10))[0] is first[0], "the fixture is not reusing to begin with"
+    pipe_mod.SenseNovaU1Pipeline.release_captured_graphs(host)
+    after = ctx(host, _dyn_cache(10))
+    assert after[0] is not first[0], "the cache survived the release"
+    assert after[1] is not first[1], "the runner survived the release"
+
+
+class _FakeLoRAWrapper(BaseLayerWithLoRA):
+    """Stands in for what ``replace_submodule`` leaves in the module tree."""
+
+    def __init__(self):
+        super().__init__()
+
+
+def test_no_capture_crosses_a_request_once_lora_wrappers_are_installed(monkeypatch):
+    """Base -> LoRA -> base. The reuse test cannot see an adapter change.
+
+    ``DiffusionLoRAManager`` replaces the decode path's linear layers with
+    ``BaseLayerWithLoRA`` and binds, rescales or resets slot 0 per request, and
+    the wrappers stay in the tree afterwards. A graph captured before an adapter
+    was bound would replay without its matmuls; one captured with it bound would
+    survive its removal. Neither moves the shape, dtype or bucket that decide
+    reuse, so the wrappers themselves have to end it.
+    """
+    pipe_mod, host = _pipeline_host(monkeypatch)
+    monkeypatch.setattr(pipe_mod, "DecodeGraphRunner", lambda lm, cache, dev: SimpleNamespace(cache=cache))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    base = ctx(host, _dyn_cache(10))
+    assert ctx(host, _dyn_cache(10))[0] is base[0], "the fixture is not reusing to begin with"
+
+    host.language_model.add_module("q_proj", _FakeLoRAWrapper())
+    adapted = ctx(host, _dyn_cache(10))
+    assert adapted[0] is not base[0], "an adapted request reused the base capture"
+    assert adapted[1] is not base[1], "an adapted request reused the base runner"
+
+    back = ctx(host, _dyn_cache(10))
+    assert back[0] is not adapted[0], "a later request reused the adapted capture"
+    assert back[1] is not adapted[1], "a later request reused the adapted runner"
+    assert getattr(host, "_paged_decode", None) is None, "an adapted capture was left on the pipeline"
+
+
+@cuda_only
+@pytest.mark.cuda
+@pytest.mark.L4
+def test_a_second_request_replays_the_first_capture(monkeypatch):
+    """Reuse is only worth anything if the capture survives it."""
+    dev = torch.device("cuda")
+    pipe_mod, host = _pipeline_host(monkeypatch, device="cuda", language_model=_StubLM(dev))
+    ctx = pipe_mod.SenseNovaU1Pipeline._decode_context
+
+    cache, runner = ctx(host, _dyn_cache(10, dev))
+    cache.set_length(cache.length + 1)
+    runner.step(0, 0)
+    assert runner.captures == 1
+
+    cache2, runner2 = ctx(host, _dyn_cache(12, dev))
+    assert cache2 is cache and runner2 is runner
+    cache2.set_length(cache2.length + 1)
+    runner2.step(0, 0)
+    assert runner.captures == 1, "the second request captured the graph again"
+
+
+class _Tok:
+    """The two tokenizer calls ``_generate_text`` makes."""
+
+    EOS = 7
+
+    def convert_tokens_to_ids(self, token):
+        return self.EOS
+
+    def decode(self, ids, skip_special_tokens=False):
+        return " ".join(str(i) for i in ids)
+
+
+class _Out:
+    def __init__(self, logits, past_key_values):
+        self.logits = logits
+        self.past_key_values = past_key_values
+
+
+class _TextHost:
+    """Carries only what ``_generate_text`` touches."""
+
+    def __init__(self, context):
+        self._context = context
+        self.seen: list[object] = []
+        self.tokenizer = _Tok()
+
+    def _decode_context(self, past_key_values):
+        return self._context
+
+    def _ar_step(self, next_token, t_idx, past_key_values, decode=None):
+        self.seen.append(decode)
+        # Stop the loop: argmax of this row is the EOS id.
+        logits = torch.full((1, 1, 8), -1.0)
+        logits[0, 0, _Tok.EOS] = 1.0
+        return _Out(logits, past_key_values)
+
+
+def test_text_decoding_uses_the_paged_context_too():
+    """Image-to-text and text-to-text run their own decode loop. It used to call
+    ``_ar_step`` without the context, so the paged path was documented but never
+    reached outside think."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+
+    context = object()
+    host = _TextHost(context)
+    prefix = torch.full((1, 1, 8), -1.0)
+    prefix[0, 0, 5] = 1.0  # not EOS, so one step runs
+    SenseNovaU1Pipeline._generate_text(host, prefix, past_key_values=None, t_idx=0)
+    assert host.seen == [context], f"text decoding ran with decode={host.seen}"
+
+
+class _WarmupReq:
+    """The engine's dummy request, which carries the reserved id."""
+
+    def __init__(self, request_id):
+        self.request_id = request_id
+        self.prompts = [{"prompt": "", "modalities": ["image"]}]
+
+
+def test_the_dummy_warmup_request_drives_one_decode_step(monkeypatch):
+    """The dummy request is a text-to-image run with think off, so it exercises
+    the prefill shape and never the decode one. Without this hook whatever the
+    deploy config compiles for decode lands on the first real request."""
+    from vllm_omni.diffusion.models.sensenova_u1.pipeline_sensenova_u1 import (
+        SenseNovaU1Pipeline,
+    )
+    from vllm_omni.diffusion.request import DUMMY_DIFFUSION_REQUEST_ID
+
+    calls: list[str] = []
+    host = object.__new__(SenseNovaU1Pipeline)
+    monkeypatch.setattr(host, "_warm_ar_decode", lambda: calls.append("warm"), raising=False)
+    monkeypatch.setattr(
+        SenseNovaU1Pipeline,
+        "_parse_request",
+        lambda self, req: (_ for _ in ()).throw(_StopForwardError()),
+        raising=False,
+    )
+
+    with pytest.raises(_StopForwardError):
+        SenseNovaU1Pipeline.forward(host, _WarmupReq(DUMMY_DIFFUSION_REQUEST_ID))
+    assert calls == ["warm"], "the dummy warmup request did not drive a decode step"
+
+    calls.clear()
+    with pytest.raises(_StopForwardError):
+        SenseNovaU1Pipeline.forward(host, _WarmupReq("a-real-request"))
+    assert calls == [], "a real request paid for the warmup"
+
+
+class _StopForwardError(Exception):
+    """Stops ``forward`` right after the warmup hook, so the test needs no model."""
@@ -0,0 +1,107 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""CPU tests for SenseNova-U1 RMSNorm and the U1.5 config surface.
+
+``Qwen3RMSNorm`` was a hand-rolled fp32 up/down-cast chain; it now calls the fused
+``F.rms_norm``. These pin the mathematics against an independent reference so a
+future kernel swap cannot silently change what the layer computes, and pin the two
+config fields that distinguish U1.5 from U1.
+"""
+
+import pytest
+import torch
+
+from vllm_omni.diffusion.models.sensenova_u1.sensenova_u1_transformer import Qwen3RMSNorm
+from vllm_omni.transformers_utils.configs.sensenova_u1 import SenseNovaU1Config
+
+pytestmark = [pytest.mark.core_model, pytest.mark.diffusion, pytest.mark.cpu]
+
+
+def _reference(x: torch.Tensor, weight: torch.Tensor, eps: float) -> torch.Tensor:
+    """RMSNorm written out in float64 so it is independent of the implementation."""
+    x64 = x.to(torch.float64)
+    var = x64.pow(2).mean(-1, keepdim=True)
+    return (weight.to(torch.float64) * (x64 * torch.rsqrt(var + eps))).to(x.dtype)
+
+
+@pytest.mark.parametrize("hidden", [8, 128, 4096])
+@pytest.mark.parametrize(
+    "dtype,atol",
+    [(torch.float32, 1e-5), (torch.bfloat16, 3e-2), (torch.float16, 4e-3)],
+)
+def test_rmsnorm_matches_float64_reference(hidden: int, dtype: torch.dtype, atol: float):
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(hidden, eps=1e-6).to(dtype)
+    with torch.no_grad():
+        norm.weight.copy_(torch.rand(hidden, dtype=dtype) + 0.5)
+    x = torch.randn(7, hidden, dtype=dtype)
+    got = norm(x)
+    want = _reference(x, norm.weight, norm.variance_epsilon)
+    assert got.dtype == dtype
+    assert got.shape == x.shape
+    torch.testing.assert_close(got.to(torch.float32), want.to(torch.float32), atol=atol, rtol=atol)
+
+
+def _cast_chain(x: torch.Tensor, weight: torch.Tensor, eps: float) -> torch.Tensor:
+    """The implementation this PR replaced, kept so its accuracy can be compared."""
+    var = x.to(torch.float32).pow(2).mean(-1, keepdim=True)
+    return weight * (x.to(torch.float32) * torch.rsqrt(var + eps)).to(weight.dtype)
+
+
+@pytest.mark.parametrize("hidden", [512, 4096])
+@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
+def test_rmsnorm_is_closer_to_float64_than_the_cast_chain(hidden: int, dtype: torch.dtype):
+    """The cast chain rounds to ``dtype`` before the weight multiply and so rounds
+    twice, while ``F.rms_norm`` rounds once. Restoring the cast chain fails here."""
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(hidden, eps=1e-6).to(dtype)
+    with torch.no_grad():
+        norm.weight.copy_(torch.rand(hidden, dtype=dtype) + 0.5)
+    x = torch.randn(64, hidden, dtype=dtype)
+
+    reference = _reference(x, norm.weight, norm.variance_epsilon).to(torch.float64)
+    scale = reference.abs().mean()
+
+    def relative_error(out: torch.Tensor) -> float:
+        return ((out.to(torch.float64) - reference).abs().mean() / scale).item()
+
+    fused = relative_error(norm(x))
+    legacy = relative_error(_cast_chain(x, norm.weight, norm.variance_epsilon))
+    assert fused < legacy, f"{dtype} hidden={hidden}: fused {fused:.3e} not better than {legacy:.3e}"
+
+
+def test_rmsnorm_is_scale_equivariant():
+    """RMSNorm divides by the RMS, so scaling the input must not change the output."""
+    torch.manual_seed(0)
+    norm = Qwen3RMSNorm(64, eps=1e-12)
+    x = torch.randn(4, 64)
+    torch.testing.assert_close(norm(x), norm(x * 7.5), atol=1e-4, rtol=1e-4)
+
+
+def test_rmsnorm_preserves_leading_dims():
+    norm = Qwen3RMSNorm(16, eps=1e-6)
+    for shape in [(16,), (3, 16), (2, 3, 16)]:
+        assert norm(torch.randn(*shape)).shape == shape
+
+
+def test_rmsnorm_weight_is_applied():
+    norm = Qwen3RMSNorm(4, eps=1e-6)
+    x = torch.randn(2, 4)
+    base = norm(x)
+    with torch.no_grad():
+        norm.weight.mul_(3.0)
+    torch.testing.assert_close(norm(x), base * 3.0, atol=1e-5, rtol=1e-5)
+
+
+def test_u15_config_fields_default_to_u1_values():
+    """U1.5 flips exactly these two; the defaults must stay U1's so old checkpoints
+    keep their behaviour when the keys are absent."""
+    cfg = SenseNovaU1Config()
+    assert cfg.use_pixel_head is False
+    assert cfg.noise_scale_max_value == 10.0
+
+
+def test_u15_config_fields_are_read_from_kwargs():
+    cfg = SenseNovaU1Config(use_pixel_head=True, noise_scale_max_value=16.0)
+    assert cfg.use_pixel_head is True
+    assert cfg.noise_scale_max_value == 16.0
@@ -179,6 +179,30 @@ def test_request_scoped_cache_dit_lifecycle_is_pipeline_opt_in():
    assert events == [backend]


+def test_release_captured_graphs_clears_runners_and_delegates_to_the_pipeline():
+    """Both halves in one place, so a pipeline that keeps captures is collected
+    without the caller knowing which pipelines have one."""
+    released = []
+    runner = object.__new__(DiffusionModelRunner)
+    runner.graph_runners = {"decode": object()}
+    runner.pipeline = SimpleNamespace(release_captured_graphs=lambda: released.append(True))
+
+    runner.release_captured_graphs()
+
+    assert runner.graph_runners == {}
+    assert released == [True]
+
+
+def test_release_captured_graphs_tolerates_a_pipeline_without_captures():
+    """Most pipelines keep none, and the runner may not carry `graph_runners`."""
+    runner = object.__new__(DiffusionModelRunner)
+    runner.pipeline = object()
+
+    runner.release_captured_graphs()
+
+    assert not hasattr(runner, "graph_runners")
+
+
 def _make_runner(cache_backend, cache_backend_name: str, enable_cache_dit_summary: bool = True):
    runner = object.__new__(DiffusionModelRunner)
    runner.vllm_config = object()
@@ -441,7 +441,6 @@ class TestPipelineStreamingOutputToEntrypoint:
                    model="test_model",
                    od_config=od_config,
                    metadata=metadata,
-                    batch_size=1,
                )

    @staticmethod
@@ -526,6 +525,7 @@ class TestSupportedPipelines:
        runner = object.__new__(DiffusionModelRunner)
        runner.vllm_config = _make_vllm_config()
        runner.od_config = SimpleNamespace(
+            max_num_seqs=1,
            enable_cpu_offload=False,
            enable_layerwise_offload=False,
            enforce_eager=True,
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
 Unit tests for DiffusionWorker class.
@@ -143,6 +143,21 @@ class TestDiffusionWorkerSleep:
        assert "buffer1" in mock_gpu_worker._sleep_saved_buffers
        assert "buffer2" in mock_gpu_worker._sleep_saved_buffers

+    def test_sleep_level_2_releases_captures_through_the_model_runner(self, mocker: MockerFixture, mock_gpu_worker):
+        """Level 2 discards the memory captures were recorded against, and the
+        runner owns the release. The worker asks it once and does not reach past
+        it into the pipeline."""
+        mock_platform = mocker.patch("vllm_omni.diffusion.worker.diffusion_worker.current_omni_platform")
+        mock_platform.get_free_memory.side_effect = [10 * 1024**3, 12 * 1024**3]
+        mock_platform.get_device_total_memory.return_value = 80 * 1024**3
+        mocker.patch("vllm_omni.diffusion.worker.diffusion_worker.get_process_gpu_memory", side_effect=[0, 0])
+        mock_gpu_worker.model_runner.pipeline.named_buffers = mocker.Mock(return_value=[])
+
+        mock_gpu_worker.sleep(level=2)
+
+        mock_gpu_worker.model_runner.release_captured_graphs.assert_called_once_with()
+        mock_gpu_worker.model_runner.pipeline.release_captured_graphs.assert_not_called()
+
    def test_sleep_memory_freed_validation(self, mocker: MockerFixture, mock_gpu_worker):
        """Test that sleep validates memory was actually freed."""
        mock_allocator_class = patch_cumem_allocator(mocker)
@@ -47,7 +47,7 @@ def client(mock_engine):
    with patch.object(InlineStageDiffusionClient, "_enrich_config"):
        od_config = MagicMock(spec=OmniDiffusionConfig)
        od_config.streaming_output = False
-        c = InlineStageDiffusionClient(model="test_model", od_config=od_config, metadata=metadata, batch_size=1)
+        c = InlineStageDiffusionClient(model="test_model", od_config=od_config, metadata=metadata)
        yield c
        c.shutdown()

@@ -234,5 +234,4 @@ def test_inline_client_requires_replica_id(mock_engine):
                model="test_model",
                od_config=od_config,
                metadata=metadata,
-                batch_size=1,
            )
@@ -1171,7 +1171,6 @@ class TestStageDiffusionClientErrorPropagation:
                metadata,
                "tcp://req",
                "tcp://resp",
-                batch_size=1,
            )

    @pytest.mark.asyncio
@@ -387,7 +387,6 @@ class TestPromptUpdateIntegration:
                    model="test_model",
                    od_config=od_config,
                    metadata=metadata,
-                    batch_size=1,
                )

    @staticmethod
@@ -22,7 +22,7 @@ from vllm.inputs import TokensPrompt

 from tests.helpers.clean import wait_for_gpu_memory_to_clear
 from tests.helpers.mark import hardware_test
-from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
+from tests.helpers.stage_config import get_deploy_config_path
 from vllm_omni.entrypoints.async_omni import AsyncOmni
 from vllm_omni.platforms import current_omni_platform

@@ -52,26 +52,7 @@ def _colocate_async_deploy() -> str:
    CI yaml plus CuMem on every stage fails engine-core startup in the L3
    merge job: Stage 1 resolves ``devices: "1"`` while only GPU 0 is visible.
    """
-    return modify_stage_config(
-        get_deploy_config_path("qwen3_omni_moe_thinking.yaml"),
-        updates={
-            "stages": {
-                0: {
-                    "enable_sleep_mode": True,
-                    "enforce_eager": True,
-                    "trust_remote_code": True,
-                    "tensor_parallel_size": 1,
-                    "devices": "0",
-                    "gpu_memory_utilization": 0.9,
-                    "max_num_seqs": 1,
-                    "max_model_len": 8192,
-                    "max_num_batched_tokens": 8192,
-                    "enable_prefix_caching": False,
-                    "skip_mm_profiling": True,
-                },
-            },
-        },
-    )
+    return get_deploy_config_path("ci/qwen3_omni_moe_colocate_async.yaml")


@pytest.mark.advanced_model
@@ -0,0 +1,4 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""MiniMax online-serving tests."""
@@ -0,0 +1,369 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""Shared request and server helpers for MiniMax-H3 L3 tests."""
+
+from __future__ import annotations
+
+import base64
+import concurrent.futures
+import io
+import json
+import os
+from collections.abc import Callable
+from pathlib import Path
+from typing import Any
+
+import av
+import requests
+
+from tests.helpers.assertions import assert_video_valid
+from tests.helpers.media import generate_synthetic_image
+from tests.helpers.runtime import OmniServerParams, OpenAIClientHandler
+
+# DLO starts worker processes and uses shared-memory queues.  Always select
+# spawn for these cases so a parent fork cannot inherit stale resource-tracker
+# state from a preceding server lifecycle.
+os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
+
+MODEL = os.environ.get("VLLM_TEST_MINIMAX_H3_MODEL", "MiniMaxAI/MiniMax-H3")
+WIDTH = 1344
+HEIGHT = 768
+FPS = 24
+NUM_INFERENCE_STEPS = 4
+# Keep a failed collective/preprocessing regression bounded well below the
+# per-file Buildkite timeout.  Healthy H100 runs complete in under two minutes
+# after startup; ten minutes leaves ample room for a cold compile.
+REQUEST_TIMEOUT_SECONDS = 600
+
+# FastH3's four-step preview is a load-time fusion adapter.  Its model-level
+# deltas cannot use the dynamic PEFT LoRA manager, so this test deliberately
+# skips on branches that predate PR #6714 instead of starting a base-H3 server.
+_REPO_ROOT = Path(__file__).resolve().parents[4]
+FASTH3_SUPPORTED = (_REPO_ROOT / "vllm_omni/diffusion/models/minimax_h3/fasth3.py").is_file()
+# The Turbo artifact is handled by the MiniMax-specific loader introduced in
+# PR #6550; the generic diffusion LoRA manager cannot parse that checkpoint.
+TURBO_SUPPORTED = (_REPO_ROOT / "vllm_omni/diffusion/models/minimax_h3/lora.py").is_file()
+FASTH3_WIDTH = 1024
+FASTH3_HEIGHT = 576
+FASTH3_DURATION = 4.4
+
+DLO_SERVER_ARGS = [
+    "--trust-remote-code",
+    "--num-gpus",
+    "2",
+    "--tensor-parallel-size",
+    "1",
+    "--data-parallel-size",
+    "2",
+    "--request-batch-max-wait-ms",
+    "500",
+    "--usp",
+    "1",
+    "--ring",
+    "1",
+    "--text-encoder-tp-size",
+    "1",
+    "--vae-patch-parallel-size",
+    "1",
+    "--vae-parallel-mode",
+    "tile",
+    "--vae-use-tiling",
+    "--enable-distributed-layerwise-offload",
+]
+
+_FL2VA_IMAGE = base64.b64decode(generate_synthetic_image(WIDTH, HEIGHT, seed=42)["base64"])
+# Keep the DLO/DP2 Ref2VA wave on the image-only path.  The separate video
+# reference path performs per-request ffmpeg/Qwen video preprocessing and has
+# repeatedly stalled this merge wave before the first pipeline log; it remains
+# covered by the Ref2VA contract/accuracy suites.
+_REF2VA_IMAGE = base64.b64decode(generate_synthetic_image(512, 288, seed=43)["base64"])
+
+
+def resolve_turbo_lora() -> str | None:
+    """Resolve the legacy Turbo LoRA artifact without network at collection."""
+    configured = os.environ.get("VLLM_TEST_MINIMAX_H3_TURBO_LORA")
+    if configured and Path(configured).is_file():
+        return configured
+
+    try:
+        from huggingface_hub import hf_hub_download
+
+        return hf_hub_download(
+            repo_id="lightx2v/Minimax-h3-Turbo",
+            filename="minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors",
+            local_files_only=True,
+        )
+    except Exception:
+        return None
+
+
+def resolve_fasth3_lora() -> str | None:
+    """Resolve the dense/data-free FastH3 artifact without network at collection."""
+    configured = os.environ.get("VLLM_TEST_MINIMAX_H3_FASTH3_LORA")
+    if configured and Path(configured).is_file():
+        return configured
+
+    try:
+        from huggingface_hub import hf_hub_download
+
+        return hf_hub_download(
+            repo_id="FastVideo/FastVideo-FastH3-4-step-Preview-v1-LoRA",
+            filename="dense-datafree/adapter_model.safetensors",
+            local_files_only=True,
+        )
+    except Exception:
+        # Buildkite downloads the artifact before pytest.  Local collection
+        # must stay offline-friendly when that optional cache is absent.
+        return None
+
+
+FASTH3_LORA = resolve_fasth3_lora()
+TURBO_LORA = resolve_turbo_lora()
+
+
+def _assert_audio_stream_present(video: bytes) -> None:
+    """Assert that the generated MP4 contains decodable audio samples."""
+    with av.open(io.BytesIO(video)) as container:
+        audio_streams = [stream for stream in container.streams if stream.type == "audio"]
+        assert audio_streams, "MiniMax-H3 MP4 has no audio stream"
+        audio_frame = next(container.decode(audio=0), None)
+        assert audio_frame is not None and audio_frame.samples > 0, "MiniMax-H3 MP4 audio stream is empty"
+
+
+def assert_h3_video(video: bytes, *, width: int, height: int) -> None:
+    """Check the common H3 video/audio response contract."""
+    assert_video_valid(video, width=width, height=height, fps=FPS)
+    _assert_audio_stream_present(video)
+
+
+def post_sync(
+    client: OpenAIClientHandler,
+    form_data: dict[str, str],
+    files: Any = None,
+) -> bytes:
+    """Submit one synchronous video request and return its MP4 body."""
+    response = requests.post(
+        f"{client.base_url.rstrip('/')}/v1/videos/sync",
+        data=form_data,
+        files=files,
+        headers={"Accept": "video/mp4"},
+        timeout=REQUEST_TIMEOUT_SECONDS,
+    )
+    response.raise_for_status()
+    assert response.headers.get("content-type", "").startswith("video/mp4")
+    assert response.content, "MiniMax-H3 returned an empty video body"
+    return response.content
+
+
+def _h3_form(task: str, seed: int) -> dict[str, str]:
+    return {
+        "model": MODEL,
+        "prompt": (
+            "A cinematic live-action scene with a clear subject moving naturally; "
+            "the atmosphere includes synchronized environmental sound."
+        ),
+        "width": str(WIDTH),
+        "height": str(HEIGHT),
+        "fps": str(FPS),
+        "num_inference_steps": str(NUM_INFERENCE_STEPS),
+        "flow_shift": "12",
+        "seed": str(seed),
+        "extra_params": json.dumps(
+            {
+                "task": task,
+                "duration": 4.0,
+                "aspect_ratio": "16:9",
+                "audio_flow_shift": 3.0,
+            },
+            separators=(",", ":"),
+        ),
+    }
+
+
+def run_fl2va(client: OpenAIClientHandler, seed: int) -> bytes:
+    return post_sync(
+        client,
+        _h3_form("fl2va", seed),
+        files=fl2va_files(),
+    )
+
+
+def run_ref2va(client: OpenAIClientHandler, seed: int) -> bytes:
+    return post_sync(
+        client,
+        _h3_form("ref2va", seed),
+        files=[("input_references", ("reference.jpg", io.BytesIO(_REF2VA_IMAGE), "image/jpeg"))],
+    )
+
+
+def fl2va_files() -> dict[str, tuple[str, io.BytesIO, str]]:
+    """Return a fresh multipart first-frame payload for an FL2VA request."""
+    return {"input_reference": ("first_frame.jpg", io.BytesIO(_FL2VA_IMAGE), "image/jpeg")}
+
+
+def run_dlo_wave(
+    client: OpenAIClientHandler,
+    request_fn: Callable[[OpenAIClientHandler, int], bytes],
+) -> list[bytes]:
+    """Run one complete two-request DLO/DP2 wave concurrently."""
+    with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
+        return list(executor.map(request_fn, (client, client), (2101, 2102)))
+
+
+def dlo_params(task_type: str) -> OmniServerParams:
+    return OmniServerParams(
+        model=MODEL,
+        server_args=[*DLO_SERVER_ARGS, "--task-type", task_type],
+        stage_init_timeout=1800,
+        init_timeout=1800,
+    )
+
+
+def turbo_params(lora_path: str | None) -> OmniServerParams:
+    """Build the two-card TP2 + DLO-no-AllGather Turbo layout."""
+    return OmniServerParams(
+        model=MODEL,
+        server_args=[
+            "--trust-remote-code",
+            "--task-type",
+            "fl2va",
+            "--num-gpus",
+            "2",
+            "--tensor-parallel-size",
+            "2",
+            "--usp",
+            "1",
+            "--ring",
+            "1",
+            "--text-encoder-tp-size",
+            "2",
+            "--vae-patch-parallel-size",
+            "2",
+            "--vae-parallel-mode",
+            "tile",
+            "--vae-use-tiling",
+            "--enable-distributed-layerwise-offload",
+            "--dlo-no-use-allgather",
+            "--lora-backend",
+            "peft",
+            "--lora-path",
+            lora_path or "",
+        ],
+        stage_init_timeout=1800,
+        init_timeout=1800,
+    )
+
+
+def turbo_form(seed: int) -> dict[str, str]:
+    """Build the official five-sigma-point Turbo FL2VA request."""
+    return {
+        "model": MODEL,
+        "prompt": (
+            "A man stands beside a yellow car at night; the car drives away "
+            "as he follows it with his eyes and begins singing sadly."
+        ),
+        "width": str(WIDTH),
+        "height": str(HEIGHT),
+        "fps": str(FPS),
+        # The public Turbo artifact has four denoiser evaluations but its API
+        # contract requests five sigma points.
+        "num_inference_steps": "5",
+        "flow_shift": "6",
+        "seed": str(seed),
+        "extra_params": json.dumps(
+            {"task": "fl2va", "duration": FASTH3_DURATION, "aspect_ratio": "16:9", "audio_flow_shift": 3.0},
+            separators=(",", ":"),
+        ),
+        "lora": json.dumps(
+            {"name": "h3-turbo-v1.0", "path": TURBO_LORA, "scale": 1.0},
+            separators=(",", ":"),
+        ),
+    }
+
+
+def fasth3_params(lora_path: str | None) -> OmniServerParams:
+    """Build the requested two-card FastH3 parallel layout.
+
+    HSDP owns DiT sharding (the model tensor-parallel degree stays at its
+    default 1), USP2 shards the sequence, text-encoder TP2 shards Qwen3-VL,
+    and VAE patch parallel 2 is the project's VPP2 setting.
+    """
+    return OmniServerParams(
+        model=MODEL,
+        server_args=[
+            "--trust-remote-code",
+            "--task-type",
+            "fl2va",
+            "--num-gpus",
+            "2",
+            "--usp",
+            "2",
+            "--ring",
+            "1",
+            "--use-hsdp",
+            "--hsdp-shard-size",
+            "2",
+            "--text-encoder-tp-size",
+            "2",
+            "--vae-patch-parallel-size",
+            "2",
+            "--vae-parallel-mode",
+            "tile",
+            "--vae-use-tiling",
+            "--lora-path",
+            lora_path or "",
+        ],
+        stage_init_timeout=1800,
+        init_timeout=1800,
+    )
+
+
+def fasth3_form(seed: int) -> dict[str, str]:
+    """Build the four-step T2VA request accepted by the FastH3 preview."""
+    return {
+        "model": MODEL,
+        "prompt": (
+            "A cinematic live-action scene with a clear subject moving naturally; "
+            "the atmosphere includes synchronized environmental sound."
+        ),
+        "width": str(FASTH3_WIDTH),
+        "height": str(FASTH3_HEIGHT),
+        "fps": str(FPS),
+        "num_inference_steps": "4",
+        "seed": str(seed),
+        # FastH3 is fused at startup, so a request-level `lora` field is both
+        # unnecessary and rejected by the load-time adapter contract.
+        "extra_params": json.dumps(
+            {
+                "task": "t2va",
+                "duration": FASTH3_DURATION,
+                "aspect_ratio": "16:9",
+            },
+            separators=(",", ":"),
+        ),
+    }
+
+
+__all__ = [
+    "FASTH3_LORA",
+    "FASTH3_SUPPORTED",
+    "FASTH3_DURATION",
+    "FASTH3_HEIGHT",
+    "FASTH3_WIDTH",
+    "FPS",
+    "MODEL",
+    "assert_h3_video",
+    "dlo_params",
+    "fasth3_form",
+    "fasth3_params",
+    "fl2va_files",
+    "post_sync",
+    "run_dlo_wave",
+    "run_fl2va",
+    "run_ref2va",
+    "TURBO_LORA",
+    "TURBO_SUPPORTED",
+    "turbo_form",
+    "turbo_params",
+]
@@ -0,0 +1,27 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""MiniMax-H3 FL2VA DLO + DP2 L3 online-serving case."""
+
+from __future__ import annotations
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OpenAIClientHandler
+
+from ._common import assert_h3_video, dlo_params, run_dlo_wave, run_fl2va
+
+pytestmark = [pytest.mark.core_model, pytest.mark.advanced_model, pytest.mark.diffusion, pytest.mark.slow]
+H100_TWO_CARD_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+
+
+@pytest.mark.parametrize(
+    "omni_server",
+    [pytest.param(dlo_params("fl2va"), id="minimax_h3_dlo_dp2_fl2va", marks=H100_TWO_CARD_MARKS)],
+    indirect=True,
+)
+def test_minimax_h3_dlo_dp2_fl2va(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+    """Validate a complete concurrent DLO/DP2 wave with two FL2VA jobs."""
+    for video in run_dlo_wave(openai_client, run_fl2va):
+        assert_h3_video(video, width=1344, height=768)
@@ -0,0 +1,27 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""MiniMax-H3 Ref2VA DLO + DP2 L3 online-serving case."""
+
+from __future__ import annotations
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OpenAIClientHandler
+
+from ._common import assert_h3_video, dlo_params, run_dlo_wave, run_ref2va
+
+pytestmark = [pytest.mark.core_model, pytest.mark.advanced_model, pytest.mark.diffusion, pytest.mark.slow]
+H100_TWO_CARD_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+
+
+@pytest.mark.parametrize(
+    "omni_server",
+    [pytest.param(dlo_params("ref2va"), id="minimax_h3_dlo_dp2_ref2va", marks=H100_TWO_CARD_MARKS)],
+    indirect=True,
+)
+def test_minimax_h3_dlo_dp2_ref2va(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+    """Validate a complete concurrent DLO/DP2 wave with two image-only Ref2VA jobs."""
+    for video in run_dlo_wave(openai_client, run_ref2va):
+        assert_h3_video(video, width=1344, height=768)
@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""MiniMax-H3 FastH3 four-step LoRA L3 online-serving case."""
+
+from __future__ import annotations
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OpenAIClientHandler
+
+from ._common import (
+    FASTH3_HEIGHT,
+    FASTH3_LORA,
+    FASTH3_SUPPORTED,
+    FASTH3_WIDTH,
+    assert_h3_video,
+    fasth3_form,
+    fasth3_params,
+    post_sync,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.advanced_model, pytest.mark.diffusion, pytest.mark.slow]
+H100_TWO_CARD_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+
+
+@pytest.mark.skipif(not FASTH3_SUPPORTED, reason="requires FastH3 support from MiniMax-H3 PR #6714")
+@pytest.mark.skipif(
+    FASTH3_LORA is None,
+    reason="set VLLM_TEST_MINIMAX_H3_FASTH3_LORA or populate the local HF cache",
+)
+@pytest.mark.parametrize(
+    "omni_server",
+    [
+        pytest.param(
+            fasth3_params(FASTH3_LORA),
+            id="minimax_h3_fasth3_lora_hsdp2_usp2_vpp2",
+            marks=H100_TWO_CARD_MARKS,
+        )
+    ],
+    indirect=True,
+)
+def test_minimax_h3_fasth3_lora_hsdp2_usp2_vpp2(
+    omni_server: OmniServer,
+    openai_client: OpenAIClientHandler,
+) -> None:
+    """Run the dense/data-free FastH3 adapter on two H100-class cards.
+
+    The preview adapter is T2VA-only and must be fused through ``--lora-path``;
+    it is not a request-switchable PEFT adapter. 1024x576 is intentionally
+    below the 1344x768 H3 reference shape to stay inside an 80 GiB H100 after
+    regional compile cache growth.
+    """
+    video = post_sync(openai_client, fasth3_form(seed=3101))
+    assert_h3_video(video, width=FASTH3_WIDTH, height=FASTH3_HEIGHT)
@@ -0,0 +1,56 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""MiniMax-H3 Turbo LoRA + DLO L3 online-serving case."""
+
+from __future__ import annotations
+
+import pytest
+
+from tests.helpers.mark import hardware_marks
+from tests.helpers.runtime import OmniServer, OpenAIClientHandler
+
+from ._common import (
+    TURBO_LORA,
+    TURBO_SUPPORTED,
+    assert_h3_video,
+    fl2va_files,
+    post_sync,
+    turbo_form,
+    turbo_params,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.advanced_model, pytest.mark.diffusion, pytest.mark.slow]
+H100_TWO_CARD_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+
+
+@pytest.mark.skipif(
+    not TURBO_SUPPORTED,
+    reason="requires the MiniMax-H3 Turbo loader from PR #6550",
+)
+@pytest.mark.skipif(
+    TURBO_LORA is None,
+    reason="set VLLM_TEST_MINIMAX_H3_TURBO_LORA or populate the local HF cache",
+)
+@pytest.mark.parametrize(
+    "omni_server",
+    [
+        pytest.param(
+            turbo_params(TURBO_LORA),
+            id="minimax_h3_dlo_turbo_lora_tp2",
+            marks=H100_TWO_CARD_MARKS,
+        )
+    ],
+    indirect=True,
+)
+def test_minimax_h3_dlo_turbo_lora_fl2va(
+    omni_server: OmniServer,
+    openai_client: OpenAIClientHandler,
+) -> None:
+    """Validate the official four-evaluation Turbo LoRA + DLO FL2VA request."""
+    video = post_sync(
+        openai_client,
+        turbo_form(seed=2201),
+        files=fl2va_files(),
+    )
+    assert_h3_video(video, width=1344, height=768)
@@ -1,15 +1,15 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
-Online serving tests for Boogu-Image-0.1-Edit (image-to-image via chat completions).
+Online serving tests for Boogu-Image-0.1-Edit and Edit-Turbo.

-The Edit checkpoint shares the ``BooguImagePipeline`` class with the Base
+Both Edit checkpoints share the ``BooguImagePipeline`` class with the Base
 text-to-image checkpoint; the editing (TI2I) path activates when the request
-carries a reference image. Text guidance rides on ``guidance_scale`` (upstream
-default 4.0), and image guidance rides on ``guidance_scale_2`` (upstream default
-1.0 = off; a value > 1 enables the double-guidance path). Only a single
-reference image is supported.
+carries a reference image. The regular Edit checkpoint uses text guidance via
+``guidance_scale`` and optional image guidance via ``guidance_scale_2``.
+Edit-Turbo is guidance-distilled and uses four steps with guidance scale 1.0.
+Only a single reference image is supported.

 - ``test_single_image_to_image_001``: one reference image, text guidance only.
 - ``test_double_guidance_001``: one reference image, text + image guidance.
@@ -40,6 +40,10 @@ pytestmark = [pytest.mark.diffusion, pytest.mark.slow]
 EDIT_PROMPT = "Change the style to a colored pencil drawing."
 SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})

+EDIT_MODEL = "Boogu/Boogu-Image-0.1-Edit"
+EDIT_TURBO_MODEL = "Boogu/Boogu-Image-0.1-Edit-Turbo"
+EDIT_TURBO_REVISION = "hotfix-1k-20260708"
+

 def _get_diffusion_feature_cases(model: str):
    """Return one ``default`` ``OmniServerParams`` row for ``model`` (no extra ``server_args``)."""
@@ -54,13 +58,44 @@ def _get_diffusion_feature_cases(model: str):
    ]


+def _get_single_edit_cases():
+    """Return the regular and distilled Edit server/request configurations."""
+    return [
+        pytest.param(
+            OmniServerParams(model=EDIT_MODEL),
+            {
+                "num_inference_steps": 2,
+                "guidance_scale": 5.0,
+            },
+            id="edit",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+        pytest.param(
+            OmniServerParams(
+                model=EDIT_TURBO_MODEL,
+                server_args=["--revision", EDIT_TURBO_REVISION],
+            ),
+            {
+                "num_inference_steps": 4,
+                "guidance_scale": 1.0,
+            },
+            id="edit-turbo-hotfix-1k",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+    ]
+
+
@pytest.mark.parametrize(
-    "omni_server",
-    _get_diffusion_feature_cases("Boogu/Boogu-Image-0.1-Edit"),
-    indirect=True,
+    ("omni_server", "sampling_params"),
+    _get_single_edit_cases(),
+    indirect=["omni_server"],
 )
-def test_single_image_to_image_001(omni_server: OmniServer, openai_client: OpenAIClientHandler):
-    """Single-reference text-guided edit smoke for ``Boogu/Boogu-Image-0.1-Edit``."""
+def test_single_image_to_image_001(
+    omni_server: OmniServer,
+    sampling_params: dict,
+    openai_client: OpenAIClientHandler,
+):
+    """Single-reference text-guided edit smoke for regular and Turbo Edit."""
    image_data_url = f"data:image/jpeg;base64,{generate_synthetic_image(512, 512)['base64']}"

    messages = dummy_messages_from_mix_data(image_data_url=image_data_url, content_text=EDIT_PROMPT)
@@ -69,9 +104,7 @@ def test_single_image_to_image_001(omni_server: OmniServer, openai_client: OpenA
        "model": omni_server.model,
        "messages": messages,
        "extra_body": {
-            "num_inference_steps": 2,
-            # Boogu text guidance (upstream default 4.0); > 1 enables CFG.
-            "guidance_scale": 5.0,
+            **sampling_params,
            "seed": 42,
        },
    }
@@ -81,7 +114,7 @@ def test_single_image_to_image_001(omni_server: OmniServer, openai_client: OpenA

@pytest.mark.parametrize(
    "omni_server",
-    _get_diffusion_feature_cases("Boogu/Boogu-Image-0.1-Edit"),
+    _get_diffusion_feature_cases(EDIT_MODEL),
    indirect=True,
 )
 def test_double_guidance_001(omni_server: OmniServer, openai_client: OpenAIClientHandler):
@@ -105,15 +138,20 @@ def test_double_guidance_001(omni_server: OmniServer, openai_client: OpenAIClien
    openai_client.send_diffusion_request(request_config)


-def _edits_request_config(model: str, *, guidance_scale_2: float | None = None) -> dict:
+def _edits_request_config(
+    model: str,
+    *,
+    num_inference_steps: int = 2,
+    guidance_scale: float = 5.0,
+    guidance_scale_2: float | None = None,
+) -> dict:
    """Multipart ``/v1/images/edits`` config: one synthetic reference + smoke-depth params."""
    image_bytes = base64.b64decode(generate_synthetic_image(512, 512)["base64"])
    data = {
        "model": model,
        "prompt": EDIT_PROMPT,
-        "num_inference_steps": 2,
-        # Boogu text guidance (upstream default 4.0); > 1 enables CFG.
-        "guidance_scale": 5.0,
+        "num_inference_steps": num_inference_steps,
+        "guidance_scale": guidance_scale,
        "seed": 42,
    }
    if guidance_scale_2 is not None:
@@ -133,19 +171,25 @@ def _assert_edits_response_has_image(responses) -> None:


@pytest.mark.parametrize(
-    "omni_server",
-    _get_diffusion_feature_cases("Boogu/Boogu-Image-0.1-Edit"),
-    indirect=True,
+    ("omni_server", "sampling_params"),
+    _get_single_edit_cases(),
+    indirect=["omni_server"],
 )
-def test_images_edits_endpoint_001(omni_server: OmniServer, openai_client: OpenAIClientHandler):
-    """Single-reference text-guided edit via the OpenAI-compatible ``/v1/images/edits`` endpoint."""
-    responses = openai_client.send_images_edits_http_request(_edits_request_config(omni_server.model))
+def test_images_edits_endpoint_001(
+    omni_server: OmniServer,
+    sampling_params: dict,
+    openai_client: OpenAIClientHandler,
+):
+    """Single-reference regular and Turbo Edit via ``/v1/images/edits``."""
+    responses = openai_client.send_images_edits_http_request(
+        _edits_request_config(omni_server.model, **sampling_params)
+    )
    _assert_edits_response_has_image(responses)


@pytest.mark.parametrize(
    "omni_server",
-    _get_diffusion_feature_cases("Boogu/Boogu-Image-0.1-Edit"),
+    _get_diffusion_feature_cases(EDIT_MODEL),
    indirect=True,
 )
 def test_images_edits_endpoint_double_guidance_001(omni_server: OmniServer, openai_client: OpenAIClientHandler):
@@ -1,8 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """
-L4 expansion coverage for ``Boogu/Boogu-Image-0.1-Edit`` (image editing / TI2I).
+L4 expansion coverage for Boogu-Image Edit and Edit-Turbo (image editing / TI2I).

 The nightly smoke module (``test_boogu_image_edit.py``) already covers single-image
 text-guided editing and the text+image double-guidance path. This expansion
@@ -16,11 +16,12 @@ module adds the two edit paths smoke does not exercise:
  ``height`` / ``width``: asserts the output resolution follows the reference
  (upstream ``align_res``, on by default for a single reference).

-The Edit checkpoint shares ``BooguImagePipeline`` with the Base checkpoint; the
-TI2I path activates when the request carries a reference image. Only a single
-reference image is supported. Cases stay smoke-depth (``num_inference_steps=2``).
-CPU offload, Cache-DiT, and multi-GPU parallelism are not yet supported for this
-model and are intentionally omitted.
+Both Edit checkpoints share ``BooguImagePipeline`` with the Base checkpoint;
+the TI2I path activates when the request carries a reference image. Only a
+single reference image is supported. The regular Edit cases stay smoke-depth
+(``num_inference_steps=2``), while Edit-Turbo uses its tuned four-step
+configuration. CPU offload, Cache-DiT, and multi-GPU parallelism are not yet
+supported for these models and are intentionally omitted.

 From ``tests/``::

@@ -41,6 +42,8 @@ os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
 pytestmark = [pytest.mark.diffusion, pytest.mark.slow]

 MODEL = "Boogu/Boogu-Image-0.1-Edit"
+EDIT_TURBO_MODEL = "Boogu/Boogu-Image-0.1-Edit-Turbo"
+EDIT_TURBO_REVISION = "hotfix-1k-20260708"
 EDIT_PROMPT = "Change the style to a colored pencil drawing."

 SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
@@ -57,6 +60,33 @@ def _get_diffusion_feature_cases(model: str):
    ]


+def _get_align_res_cases():
+    """Return regular and Turbo Edit configurations for align-res coverage."""
+    return [
+        pytest.param(
+            OmniServerParams(model=MODEL),
+            {
+                "num_inference_steps": 2,
+                "guidance_scale": 5.0,
+            },
+            id="edit",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+        pytest.param(
+            OmniServerParams(
+                model=EDIT_TURBO_MODEL,
+                server_args=["--revision", EDIT_TURBO_REVISION],
+            ),
+            {
+                "num_inference_steps": 4,
+                "guidance_scale": 1.0,
+            },
+            id="edit-turbo-hotfix-1k",
+            marks=SINGLE_CARD_FEATURE_MARKS,
+        ),
+    ]
+
+
@pytest.mark.parametrize("omni_server", _get_diffusion_feature_cases(MODEL), indirect=True)
 def test_image_only_guidance(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
    """Image-only guidance: text CFG off (1.0), image CFG on (2.0) -> image-only CFG branch."""
@@ -76,8 +106,16 @@ def test_image_only_guidance(omni_server: OmniServer, openai_client: OpenAIClien
    openai_client.send_diffusion_request(request_config)


-@pytest.mark.parametrize("omni_server", _get_diffusion_feature_cases(MODEL), indirect=True)
-def test_align_res_output_size(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
+@pytest.mark.parametrize(
+    ("omni_server", "sampling_params"),
+    _get_align_res_cases(),
+    indirect=["omni_server"],
+)
+def test_align_res_output_size(
+    omni_server: OmniServer,
+    sampling_params: dict,
+    openai_client: OpenAIClientHandler,
+) -> None:
    """Output resolution follows the reference image (upstream ``align_res``).

    No ``height`` / ``width`` is requested. The reference is 512x384 — both
@@ -92,8 +130,7 @@ def test_align_res_output_size(omni_server: OmniServer, openai_client: OpenAICli
        "model": omni_server.model,
        "messages": messages,
        "extra_body": {
-            "num_inference_steps": 2,
-            "guidance_scale": 5.0,
+            **sampling_params,
            "seed": 42,
        },
    }
@@ -138,6 +138,23 @@ def test_voice_003(omni_server, online_client) -> None:
    online_client.send_audio_speech_request(request_config)


+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
+def test_sample_rate_001(omni_server, online_client) -> None:
+    """The speech API resamples Qwen3-TTS output from native 24 kHz to 8 kHz."""
+    request_config = {
+        "model": omni_server.model,
+        "input": "This response should be encoded as eight kilohertz audio.",
+        "stream": False,
+        "response_format": "wav",
+        "sample_rate": 8000,
+        "task_type": "CustomVoice",
+        "voice": "vivian",
+    }
+
+    online_client.send_audio_speech_request(request_config)
+
+
@hardware_test(res={"cuda": "L4"}, num_cards=1)
@pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_language_001(omni_server, online_client) -> None:
@@ -11,7 +11,7 @@ import types

 import pytest

-from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec
+from vllm_omni.diffusion.data import AttentionConfig, AttentionSpec, normalize_omni_diffusion_kwargs
 from vllm_omni.engine import async_omni_engine as async_omni_engine_module
 from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
 from vllm_omni.engine.stage_init_utils import (
@@ -172,7 +172,6 @@ def _make_stage_runtime() -> StageRuntime:
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -196,7 +195,6 @@ def test_async_omni_engine_initialize_stages_passes_log_stats_to_runtime(monkeyp
    engine.model = "dummy-model"
    engine.config_path = "dummy-config"
    engine.single_stage_mode = False
-    engine.diffusion_batch_size = 1
    engine.async_chunk = False
    engine.tokenizer = None
    engine._single_stage_id_filter = None
@@ -282,7 +280,6 @@ def test_build_logical_stage_init_plans_handles_stage_without_devices(monkeypatc
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -371,7 +368,6 @@ def test_initialize_local_diffusion_replica_restores_device_visibility_after_loc
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -416,7 +412,6 @@ def test_initialize_local_diffusion_replica_passes_stage_init_timeout_and_inline
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=4,
        async_chunk=False,
    )

@@ -429,7 +424,6 @@ def test_initialize_local_diffusion_replica_passes_stage_init_timeout_and_inline
    def _capture_launch_diffusion_stage_replica(**kwargs):
        captured["stage_id"] = kwargs["metadata"].stage_id
        captured["stage_init_timeout"] = kwargs["stage_init_timeout"]
-        captured["batch_size"] = kwargs["batch_size"]
        captured["use_inline"] = kwargs["use_inline"]
        captured["omni_master_server"] = kwargs["omni_master_server"]
        return types.SimpleNamespace(), StageReplicaResources()
@@ -441,7 +435,6 @@ def test_initialize_local_diffusion_replica_passes_stage_init_timeout_and_inline
    assert captured == {
        "stage_id": 0,
        "stage_init_timeout": 302,
-        "batch_size": 4,
        "use_inline": expected_inline,
        "omni_master_server": None,
    }
@@ -511,21 +504,20 @@ def test_initialize_local_llm_replica_scopes_runtime_env(monkeypatch):
    assert os.environ[runtime_env_var] == "parent-value"


-def test_initialize_diffusion_stage_forwards_batch_size_without_writing_max_num_seqs(monkeypatch):
+def test_initialize_diffusion_stage_preserves_configured_max_num_seqs(monkeypatch):
    import vllm_omni.diffusion.stage_diffusion_client as client_mod
    import vllm_omni.engine.stage_init_utils as init_mod

-    od_config = types.SimpleNamespace(max_num_seqs=1)
+    od_config = types.SimpleNamespace(max_num_seqs=4)
    captured: dict[str, object] = {}
    monkeypatch.setattr(init_mod, "build_diffusion_config", lambda *args: od_config)

-    def _capture_client(model, config, metadata, stage_init_timeout, batch_size, use_inline):
+    def _capture_client(model, config, metadata, stage_init_timeout, use_inline):
        captured.update(
            model=model,
            config=config,
            metadata=metadata,
            stage_init_timeout=stage_init_timeout,
-            batch_size=batch_size,
            use_inline=use_inline,
        )
        return object()
@@ -539,27 +531,25 @@ def test_initialize_diffusion_stage_forwards_batch_size_without_writing_max_num_
        types.SimpleNamespace(),
        metadata,
        stage_init_timeout=12,
-        batch_size=4,
        use_inline=True,
    )

-    assert od_config.max_num_seqs == 1
+    assert od_config.max_num_seqs == 4
    assert captured == {
        "model": "dummy-model",
        "config": od_config,
        "metadata": metadata,
        "stage_init_timeout": 12,
-        "batch_size": 4,
        "use_inline": True,
    }


-def test_launch_diffusion_stage_replica_does_not_write_max_num_seqs_from_batch_size(monkeypatch):
+def test_launch_diffusion_stage_replica_preserves_configured_max_num_seqs(monkeypatch):
    import vllm_omni.diffusion.stage_diffusion_client as client_mod
    import vllm_omni.diffusion.stage_diffusion_proc as proc_mod
    import vllm_omni.engine.stage_engine_startup as startup_mod

-    od_config = types.SimpleNamespace(max_num_seqs=1, parallel_config=types.SimpleNamespace(world_size=1))
+    od_config = types.SimpleNamespace(max_num_seqs=4, parallel_config=types.SimpleNamespace(world_size=1))
    monkeypatch.setattr(startup_mod, "build_diffusion_config", lambda *args: od_config)
    monkeypatch.setattr(startup_mod, "acquire_device_locks", lambda *args: [])
    monkeypatch.setattr(
@@ -596,13 +586,12 @@ def test_launch_diffusion_stage_replica_does_not_write_max_num_seqs_from_batch_s
        stage_config=types.SimpleNamespace(),
        metadata=types.SimpleNamespace(stage_id=0),
        stage_init_timeout=12,
-        batch_size=4,
        use_inline=False,
        omni_master_server=omni_master_server,
    )

    assert result is sentinel_client
-    assert od_config.max_num_seqs == 1
+    assert od_config.max_num_seqs == 4
    assert resources.manager is proc_manager


@@ -621,7 +610,6 @@ def test_initialize_diffusion_stage_does_not_write_max_num_seqs(monkeypatch):
        types.SimpleNamespace(),
        metadata,
        stage_init_timeout=12,
-        batch_size=4,
        use_inline=True,
    )

@@ -672,7 +660,6 @@ def test_launch_diffusion_stage_replica_preserves_step_execution_max_num_seqs(mo
        stage_config=types.SimpleNamespace(),
        metadata=types.SimpleNamespace(stage_id=0),
        stage_init_timeout=12,
-        batch_size=4,
        use_inline=False,
        omni_master_server=omni_master_server,
    )
@@ -688,7 +675,6 @@ def test_stage_runtime_initializes_stage_pools(monkeypatch):
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -754,7 +740,6 @@ def test_stage_runtime_passes_log_stats_to_llm_replica_launch(monkeypatch):
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
        log_stats=True,
    )
@@ -798,7 +783,6 @@ def test_stage_runtime_passes_log_stats_to_output_processor(monkeypatch):
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
        log_stats=True,
    )
@@ -851,7 +835,6 @@ def test_build_logical_stage_init_plans_applies_replica_device_splits(monkeypatc
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -892,7 +875,6 @@ def test_initialize_stage_replicas_collects_results_by_stage_and_replica_id(monk
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=123,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -930,7 +912,6 @@ def test_remote_replicas_use_distinct_init_group_keys():
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=123,
-        diffusion_batch_size=1,
        async_chunk=False,
    )
    plan = _make_llm_plan(
@@ -956,7 +937,6 @@ def test_initialize_stages_cleans_up_successful_replicas_after_partial_multi_rep
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -993,7 +973,6 @@ def test_initialize_stages_cleans_up_late_successful_replicas_after_early_multi_
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )

@@ -1032,7 +1011,6 @@ def test_initialize_local_llm_replica_passes_stage_init_timeout_to_complete_stag
        model="dummy-model",
        config_path="dummy-config",
        stage_init_timeout=1,
-        diffusion_batch_size=1,
        async_chunk=False,
    )
    stage_init_timeout = 302
@@ -1817,6 +1795,41 @@ def test_resolve_stage_configs_preserves_stage_diffusion_attention(monkeypatch):
    assert stage_configs[0].engine_args.diffusion_attention_config is existing_attention


+def test_resolve_stage_configs_does_not_inject_over_stage_diffusion_attention_backend(monkeypatch):
+    import vllm_omni.engine.async_omni_engine as engine_mod
+
+    engine = object.__new__(AsyncOmniEngine)
+    stage_cfg = types.SimpleNamespace(
+        stage_type="diffusion",
+        engine_args=types.SimpleNamespace(
+            diffusion_attention_backend="TORCH_SDPA",
+            diffusion_attention_config=None,
+            lora_path=None,
+            lora_scale=None,
+            enable_sleep_mode=None,
+            quantization_config=None,
+        ),
+    )
+
+    monkeypatch.setattr(
+        engine_mod,
+        "load_and_resolve_stage_configs",
+        lambda *args, **kwargs: ("dummy-config", [stage_cfg], None),
+    )
+
+    _config_path, stage_configs = engine._resolve_stage_configs(
+        model="dummy-model",
+        kwargs={"diffusion_attention_backend": "FLASH_ATTN"},
+        trust_remote_code=False,
+    )
+
+    engine_args = stage_configs[0].engine_args
+    assert engine_args.diffusion_attention_backend == "TORCH_SDPA"
+    assert engine_args.diffusion_attention_config is None
+    normalized = normalize_omni_diffusion_kwargs(vars(engine_args))
+    assert normalized["diffusion_attention_config"].default.backend == "TORCH_SDPA"
+
+
 def test_resolve_stage_configs_does_not_inject_diffusion_attention_into_llm_stage(monkeypatch):
    import vllm_omni.engine.async_omni_engine as engine_mod

@@ -575,7 +575,6 @@ class TestSingleStageInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
            single_stage_id_filter=stage_id_filter,
            omni_master_address="127.0.0.1",
@@ -807,7 +806,6 @@ class TestSingleStageInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
        )
        mocker.patch.object(runtime_mod, "prepare_engine_environment")
@@ -860,7 +858,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -928,7 +925,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -950,7 +946,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -1002,7 +997,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=2,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -1068,7 +1062,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=4,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -1120,7 +1113,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=4,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -1141,7 +1133,7 @@ class TestSingleStageReplicaInitialization:
        runtime._init_visible_devices_baseline = "0"

        mocker.patch.object(runtime_mod, "inject_kv_stage_info")
-        od_config = SimpleNamespace(max_num_seqs=None, parallel_config=SimpleNamespace(world_size=1))
+        od_config = SimpleNamespace(max_num_seqs=4, parallel_config=SimpleNamespace(world_size=1))
        mocker.patch("vllm_omni.engine.stage_engine_startup.build_diffusion_config", return_value=od_config)
        mock_register = mocker.patch(
            "vllm_omni.engine.stage_engine_startup.register_stage_with_omni_master",
@@ -1178,7 +1170,7 @@ class TestSingleStageReplicaInitialization:
                os.environ[device_env_var] = prev_device_env

        assert result is sentinel_client
-        assert od_config.max_num_seqs is None
+        assert od_config.max_num_seqs == 4
        mock_register.assert_called_once_with(
            omni_master_address="127.0.0.1",
            omni_master_port=25000,
@@ -1203,7 +1195,6 @@ class TestSingleStageReplicaInitialization:
            request_address="tcp://127.0.0.1:26002",
            response_address="tcp://127.0.0.1:26003",
            proc_manager=mocker.ANY,
-            batch_size=4,
        )
        assert mock_from_addresses.call_args.kwargs["proc_manager"].proc is proc

@@ -1216,7 +1207,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=4,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -1258,7 +1248,6 @@ class TestSingleStageReplicaInitialization:
            model="fake-model",
            config_path="/fake/stages.yaml",
            stage_init_timeout=60,
-            diffusion_batch_size=4,
            async_chunk=False,
            single_stage_id_filter=None,
            omni_master_address="127.0.0.1",
@@ -133,7 +133,7 @@ def _stable_engine_arg_environment(monkeypatch, tmp_path):

    for worker_type in ("ar", "generation"):
        worker_module = types.ModuleType(f"test.{worker_type}")
-        worker_module.Worker = _TestWorker
+        setattr(worker_module, "Worker", _TestWorker)
        monkeypatch.setitem(sys.modules, worker_module.__name__, worker_module)


@@ -434,6 +434,63 @@ def test_typed_diffusion_engine_args_use_structured_diffusion_config(tmp_path):
    assert typed_args["diffusion_attention_config"].per_role["cross"].backend == "TORCH_SDPA"


+def test_engine_args_consume_stage_diffusion_attention_shorthand(tmp_path):
+    """A stage-level ``diffusion_attention_backend`` shorthand is folded into the structured config."""
+    pipeline, deploy, model = _engine_arg_inputs(tmp_path)
+    deploy.stages[2] = replace(
+        deploy.stages[2],
+        diffusion_attention_config=None,
+        diffusion_attention_backend="TORCH_SDPA",
+    )
+    legacy_stages, omni_config = _legacy_and_typed_stages(pipeline, deploy, model)
+
+    legacy_args = build_legacy_engine_args_dict(legacy_stages[2], model)
+    typed_args = build_engine_args_dict_from_omni_stage_config(
+        omni_config.stage_by_id(2),
+        model,
+    )
+
+    for engine_args in (legacy_args, typed_args):
+        assert engine_args.get("diffusion_attention_backend") is None
+        assert isinstance(engine_args["diffusion_attention_config"], AttentionConfig)
+        assert engine_args["diffusion_attention_config"].default.backend == "TORCH_SDPA"
+        od_config = OmniDiffusionConfig.from_kwargs(**engine_args)
+        assert od_config.diffusion_attention_config.default.backend == "TORCH_SDPA"
+
+
+@pytest.mark.parametrize(
+    "yaml_attention_config",
+    [
+        {"default": {"backend": "FLASH_ATTN"}, "per_role": {"cross": {"backend": "SAGE_ATTN"}}},
+        {"per_role": {"cross": {"backend": "SAGE_ATTN"}}},
+    ],
+    ids=["yaml-default", "yaml-per-role-only"],
+)
+def test_engine_args_apply_cli_attention_shorthand_over_yaml_config(tmp_path, yaml_attention_config):
+    pipeline, deploy, model = _engine_arg_inputs(tmp_path)
+    deploy.stages[2] = replace(deploy.stages[2], diffusion_attention_config=yaml_attention_config)
+    legacy_stages, omni_config = _legacy_and_typed_stages(
+        pipeline,
+        deploy,
+        model,
+        cli_overrides={"stage_2_diffusion_attention_backend": "TORCH_SDPA"},
+    )
+
+    legacy_args = build_legacy_engine_args_dict(legacy_stages[2], model)
+    typed_args = build_engine_args_dict_from_omni_stage_config(
+        omni_config.stage_by_id(2),
+        model,
+    )
+
+    for engine_args in (legacy_args, typed_args):
+        assert engine_args.get("diffusion_attention_backend") is None
+        attention_config = engine_args["diffusion_attention_config"]
+        assert attention_config.default.backend == "TORCH_SDPA"
+        assert attention_config.per_role["cross"].backend == "SAGE_ATTN"
+        od_config = OmniDiffusionConfig.from_kwargs(**engine_args)
+        assert od_config.diffusion_attention_config.default.backend == "TORCH_SDPA"
+
+
 def test_typed_engine_args_preserve_explicit_backend_default_overrides(tmp_path):
    pipeline, deploy, model = _engine_arg_inputs(tmp_path)
    deploy.enable_prefix_caching = False
@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Tests for audio output format handling in chat completions.

 Covers:
@@ -11,10 +12,15 @@ Covers:

 from __future__ import annotations

+from io import BytesIO
+
 import numpy as np
 import pytest
+import soundfile
+import torch
+import torchaudio

-from vllm_omni.entrypoints.openai.audio_utils_mixin import AudioMixin
+from vllm_omni.entrypoints.openai.audio_utils_mixin import AudioMixin, StreamingAudioResampler
 from vllm_omni.entrypoints.openai.protocol.audio import (
    DEFAULT_AUDIO_FORMAT,
    SUPPORTED_AUDIO_FORMATS,
@@ -122,6 +128,81 @@ class TestCreateAudio:
        decoded = base64.b64decode(response.audio_data)
        assert decoded[:4] == b"RIFF"

+    def test_resamples_wav_to_requested_output_rate(self, mixin, audio_tensor):
+        response = mixin.create_audio(
+            CreateAudio(
+                audio_tensor=audio_tensor,
+                sample_rate=24000,
+                output_sample_rate=8000,
+                response_format="wav",
+                speed=1.0,
+                base64_encode=False,
+            )
+        )
+
+        with soundfile.SoundFile(BytesIO(response.audio_data)) as audio_file:
+            assert audio_file.samplerate == 8000
+            assert audio_file.frames == 8000
+
+
+class TestStreamingAudioResampler:
+    @staticmethod
+    def _resample(waveform):
+        resampler = StreamingAudioResampler(24000, 8000)
+        return np.concatenate(
+            (
+                resampler.process(waveform),
+                resampler.process(np.empty(0), final=True),
+            )
+        )
+
+    def test_chunk_boundaries_do_not_change_output(self):
+        waveform = np.sin(np.linspace(0, 200 * np.pi, 24000, endpoint=False)).astype(np.float32)
+
+        expected = self._resample(waveform)
+
+        chunked = StreamingAudioResampler(24000, 8000)
+        pieces = [chunked.process(chunk) for chunk in np.split(waveform, [137, 2048, 9001, 17003])]
+        pieces.append(chunked.process(np.empty(0), final=True))
+        actual = np.concatenate(pieces)
+
+        assert actual.shape == (8000,)
+        np.testing.assert_allclose(actual, expected, atol=1e-6)
+
+        reference = torchaudio.functional.resample(torch.from_numpy(waveform), 24000, 8000).numpy()
+        np.testing.assert_allclose(actual[100:-100], reference[100:-100], atol=2e-3, rtol=1e-3)
+
+    def test_preserves_passband_signal(self):
+        samples = np.arange(24000, dtype=np.float32)
+        waveform = np.sin(2 * np.pi * 3000 * samples / 24000).astype(np.float32)
+
+        output = self._resample(waveform)
+
+        input_rms = np.sqrt(np.mean(waveform**2))
+        output_rms = np.sqrt(np.mean(output[100:-100] ** 2))
+        assert output_rms == pytest.approx(input_rms, rel=0.02)
+
+    @pytest.mark.parametrize("frequency", [6000, 10000])
+    def test_attenuates_aliasing_frequencies(self, frequency):
+        samples = np.arange(24000, dtype=np.float32)
+        waveform = np.sin(2 * np.pi * frequency * samples / 24000).astype(np.float32)
+
+        output = self._resample(waveform)
+
+        input_rms = np.sqrt(np.mean(waveform**2))
+        output_rms = np.sqrt(np.mean(output[100:-100] ** 2))
+        assert output_rms < input_rms * 0.01
+
+    def test_rejects_non_integer_downsampling_ratio(self):
+        with pytest.raises(ValueError, match="integer downsampling ratio"):
+            StreamingAudioResampler(24000, 16000)
+
+    def test_rejects_multichannel_audio(self):
+        resampler = StreamingAudioResampler(24000, 8000)
+
+        with pytest.raises(ValueError, match="only supports mono audio"):
+            resampler.process(np.zeros((2, 240), dtype=np.float32))
+

 class TestResolveAudioFormat:
    """Test _resolve_audio_format via the serving chat class."""
@@ -2953,8 +2953,10 @@ class TestStreamingResponse:
            usage_acc=None,
            tts_params=None,
            collect=None,
+            target_sample_rate=None,
        ):
            captured["tts_params"] = tts_params
+            captured["target_sample_rate"] = target_sample_rate
            yield b"\x00\x00"

        monkeypatch.setattr(speech_server, "_prepare_speech_generation", prepare)
@@ -2967,6 +2969,7 @@ class TestStreamingResponse:

        assert response.status_code == 200
        assert captured["tts_params"] is finalized_tts_params
+        assert captured["target_sample_rate"] is None

    def test_sse_streaming(self, streaming_app):
        """stream_format=sse without stream=True returns audio deltas as SSE."""
@@ -3292,6 +3295,28 @@ class TestMergeBatchItem:

        assert merged.non_streaming_mode is False

+    def test_sample_rate_batch_default_used(self):
+        """Batch-level sample_rate should be used when the item omits it."""
+        batch = BatchSpeechRequest(
+            items=[SpeechBatchItem(input="hi")],
+            sample_rate=8000,
+        )
+
+        merged = OmniOpenAIServingSpeech._merge_batch_item(batch, batch.items[0])
+
+        assert merged.sample_rate == 8000
+
+    def test_sample_rate_item_override_wins(self):
+        """Per-item sample_rate should override the batch-level default."""
+        batch = BatchSpeechRequest(
+            items=[SpeechBatchItem(input="hi", sample_rate=24000)],
+            sample_rate=8000,
+        )
+
+        merged = OmniOpenAIServingSpeech._merge_batch_item(batch, batch.items[0])
+
+        assert merged.sample_rate == 24000
+

 def test_streaming_speech_session_config_accepts_non_streaming_mode():
    config = StreamingSpeechSessionConfig(non_streaming_mode=True)
@@ -3299,6 +3324,33 @@ def test_streaming_speech_session_config_accepts_non_streaming_mode():
    assert config.non_streaming_mode is True


+@pytest.mark.parametrize("sample_rate", [8000, 24000])
+def test_speech_requests_accept_supported_sample_rates(sample_rate):
+    assert OpenAICreateSpeechRequest(input="hello", sample_rate=sample_rate).sample_rate == sample_rate
+
+
+@pytest.mark.parametrize("sample_rate", [0, -8000])
+def test_speech_request_rejects_non_positive_sample_rate(sample_rate):
+    with pytest.raises(ValidationError):
+        OpenAICreateSpeechRequest(input="hello", sample_rate=sample_rate)
+
+
+def test_sample_rate_uses_adapter_capabilities():
+    server = object.__new__(OmniOpenAIServingSpeech)
+    request = OpenAICreateSpeechRequest(input="hello", sample_rate=8000)
+
+    server._tts_model_type = "qwen3_tts"
+    assert server._validate_speech_sample_rate(request) is None
+
+    request.sample_rate = 16000
+    assert server._validate_speech_sample_rate(request) == (
+        "sample_rate=16000 is not supported by the current TTS model; supported rates: 8000, 24000"
+    )
+
+    server._tts_model_type = "fish_speech"
+    assert server._validate_speech_sample_rate(request) == ("sample_rate is not supported by the current TTS model")
+
+
 def test_streaming_speech_session_config_scopes_native_speed_to_http():
    with pytest.raises(
        ValidationError,
@@ -5027,6 +5079,78 @@ class TestTTSAsyncOffloading:
        assert "partial_audio" not in payload["error"]
        assert "action" not in payload["error"]

+    @pytest.mark.asyncio
+    async def test_generate_audio_chunks_resamples_pcm_to_8khz(self, qwen3_tts_server):
+        async def pcm_generator():
+            output = create_mock_audio_output_for_test()
+            output.multimodal_output.update(
+                audio=torch.linspace(-0.5, 0.5, 2400, dtype=torch.float32),
+                sr=24000,
+            )
+            yield output
+
+        chunks = [
+            chunk
+            async for chunk in qwen3_tts_server._generate_audio_chunks(
+                pcm_generator(),
+                "req-8khz",
+                target_sample_rate=8000,
+            )
+        ]
+
+        assert len(chunks) >= 2
+        assert len(b"".join(chunks)) == 800 * 2
+
+    @pytest.mark.asyncio
+    async def test_generate_audio_chunks_wav_header_uses_target_sample_rate(self, qwen3_tts_server):
+        async def pcm_generator():
+            output = create_mock_audio_output_for_test()
+            output.multimodal_output.update(
+                audio=torch.zeros(2400, dtype=torch.float32),
+                sr=24000,
+            )
+            yield output
+
+        chunks = [
+            chunk
+            async for chunk in qwen3_tts_server._generate_audio_chunks(
+                pcm_generator(),
+                "req-8khz-wav",
+                response_format="wav",
+                target_sample_rate=8000,
+            )
+        ]
+
+        assert struct.unpack("<I", chunks[0][24:28])[0] == 8000
+
+    @pytest.mark.asyncio
+    async def test_streaming_resampler_retains_first_chunk_sample_rate_metadata(self, qwen3_tts_server):
+        async def pcm_generator():
+            first = create_mock_audio_output_for_test()
+            first.multimodal_output.update(
+                audio=torch.zeros(10, dtype=torch.float32),
+                sr=24000,
+            )
+            yield first
+
+            second = create_mock_audio_output_for_test()
+            second.multimodal_output.update(audio=torch.zeros(230, dtype=torch.float32))
+            second.multimodal_output.pop("sr", None)
+            yield second
+
+        chunks = [
+            chunk
+            async for chunk in qwen3_tts_server._generate_audio_chunks(
+                pcm_generator(),
+                "req-retained-sr",
+                response_format="wav",
+                target_sample_rate=8000,
+            )
+        ]
+
+        assert chunks[0][:4] == b"RIFF"
+        assert struct.unpack("<I", chunks[0][24:28])[0] == 8000
+
    @pytest.mark.asyncio
    async def test_generate_audio_chunks_discards_ref_audio_artifact_warmup_on_close(self, qwen3_tts_server):
        async def pcm_generator():
@@ -46,6 +46,16 @@ class TestOmniBaseProfiler:
            stage_ids=None,
        )

+    def test_removed_diffusion_batch_size_fails_loudly(self, mock_engine, mocker: MockerFixture):
+        """Legacy diffusion_batch_size callers receive an actionable error."""
+        mocker.patch("vllm_omni.entrypoints.omni_base.AsyncOmniEngine", return_value=mock_engine)
+        mocker.patch("vllm_omni.entrypoints.omni_base.omni_snapshot_download", side_effect=lambda x: x)
+        mocker.patch("vllm_omni.entrypoints.omni_base.weakref.finalize")
+        from vllm_omni.entrypoints.omni_base import OmniBase
+
+        with pytest.raises(TypeError, match="diffusion_batch_size.*max_num_seqs"):
+            OmniBase(model="test-model", diffusion_batch_size=8)
+
    def test_start_profile_with_prefix(self, omni_base_instance, mock_engine):
        """Test that start_profile passes profile_prefix to collective_rpc."""
        omni_base_instance.start_profile(profile_prefix="test_trace")
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Unit tests for vllm_omni.entrypoints.utils module."""

 import os
@@ -420,6 +423,7 @@ class TestLoadAndResolveStageConfigs:
        """Ensure that kwargs survive default stage creation."""
        engine_backend = "vllm_omni.experimental.ar_diffusion.engine.ARDiffusionEngine"
        kwargs = {"dtype": torch.float32, "engine_backend": engine_backend}
+        kwargs["revision"] = "pinned-revision"
        mocker.patch("vllm_omni.entrypoints.utils.resolve_model_config_path", return_value=None)
        mocker.patch("vllm_omni.entrypoints.utils.load_stage_configs_from_model", return_value=([], None))

@@ -433,6 +437,7 @@ class TestLoadAndResolveStageConfigs:
        assert len(stage_configs) == 1
        assert "dtype" in stage_configs[0]["engine_args"]
        assert stage_configs[0]["engine_args"]["engine_backend"] == engine_backend
+        assert stage_configs[0]["engine_args"]["revision"] == "pinned-revision"

    def test_deploy_config_preserves_cli_overrides_and_replicas(self, tmp_path, mocker: MockerFixture):
        deploy_path = tmp_path / "qwen3_multi.yaml"
@@ -873,6 +873,13 @@ def assert_audio_speech_response(response: Any, request_config: dict[str, Any],
    elif req_fmt == "wav" and response.audio_format:
        assert req_fmt in response.audio_format

+    expected_sample_rate = request_config.get("sample_rate")
+    if expected_sample_rate is not None and req_fmt == "wav" and response.audio_bytes:
+        info = sf.info(io.BytesIO(response.audio_bytes))
+        assert info.samplerate == int(expected_sample_rate), (
+            f"Expected sample_rate={expected_sample_rate}, got {info.samplerate}"
+        )
+
    if run_level in {"advanced_model", "full_model"}:
        if req_fmt == "pcm" and response.audio_bytes:
            min_hnr_db = float(request_config.get("min_hnr_db", _MIN_PCM_SPEECH_HNR_DB))
@@ -1461,6 +1461,7 @@ class OnlineOmniClient:
            "seed",
            "instructions",
            "speed",
+            "sample_rate",
            "stream_format",
            "x_vector_only_mode",
        ):
@@ -381,6 +381,40 @@ _CI_OVERLAYS: dict[str, dict[str, Any]] = {
            },
        },
    },
+    "qwen3_omni_moe_colocate_async": {
+        "base_config": "qwen3_omni_moe_thinking.yaml",
+        "stages": [
+            {
+                "stage_id": 0,
+                "enable_sleep_mode": True,
+                "enforce_eager": True,
+                "trust_remote_code": True,
+                "tensor_parallel_size": 1,
+                "devices": "0",
+                "max_num_seqs": 1,
+                "max_model_len": 8192,
+                "max_num_batched_tokens": 8192,
+                "enable_prefix_caching": False,
+                "skip_mm_profiling": True,
+            },
+        ],
+        "platforms": {
+            "rocm": {
+                "stages": [
+                    {
+                        "stage_id": 0,
+                        # This control-plane test needs only 8K tokens. On MI300,
+                        # percentage sizing attempted a ~125 GiB KV allocation
+                        # after the 64.5 GiB model load and then failed while
+                        # allocating the cache. Keep CUDA's existing sizing and
+                        # give only ROCm a bounded cache with ample headroom.
+                        "gpu_memory_utilization": None,
+                        "kv_cache_memory_bytes": 2 * 1024**3,
+                    },
+                ],
+            },
+        },
+    },
    "qwen3_omni_moe_multi_replicas_4gpu": {
        "base_config": "qwen3_omni_moe.yaml",
        "async_chunk": True,
@@ -26,6 +26,7 @@ from vllm_omni.config.pipeline_registry import OMNI_PIPELINES
 from vllm_omni.config.stage_config import (
    PipelineConfig,
    StageExecutionType,
+    StagePipelineConfig,
    _apply_platform_overrides,
    load_deploy_config,
    merge_pipeline_deploy,
@@ -65,7 +66,9 @@ class TestRegistryDeclaration:
        assert not hasattr(pipeline, "max_native_duplex_sessions")

    def test_ordinary_pipeline_defaults_to_no_duplex_control(self) -> None:
-        pipeline = PipelineConfig(model_type="ordinary")
+        pipeline = PipelineConfig(
+            model_type="ordinary", stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
+        )
        assert pipeline.duplex_control_enabled is False
        assert pipeline.duplex_serving_adapter is None
        assert not hasattr(pipeline, "max_native_duplex_sessions")
@@ -80,10 +83,6 @@ class TestPipelineTopology:
        assert len(pipeline.stages) == 3
        assert [s.stage_id for s in pipeline.stages] == [0, 1, 2]

-    def test_topology_validates(self, pipeline: PipelineConfig) -> None:
-        # ``validate`` returns a list of structural errors; empty == valid.
-        assert pipeline.validate() == []
-
    def test_thinker_stage(self, pipeline: PipelineConfig) -> None:
        thinker = pipeline.get_stage(0)
        assert thinker is not None
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """L1 registration + config-shim tests for NemotronVoiceChat (CPU, no weights)."""

 import subprocess
@@ -59,7 +59,6 @@ def test_pipeline_registered_with_three_stages() -> None:
    assert code2wav.execution_type == StageExecutionType.LLM_GENERATION
    assert code2wav.final_output and code2wav.final_output_type == "audio"
    assert code2wav.input_sources == (1,)
-    assert pipeline.validate() == []


 def test_architectures_registered() -> None:
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Tests for TrackingArgumentParser and related utilities."""

 import argparse
@@ -35,7 +38,7 @@ _TEST_PIPELINE = PipelineConfig(
    stages=(
        StagePipelineConfig(stage_id=0, model_stage="stage0"),
        StagePipelineConfig(stage_id=1, model_stage="stage1"),
-        StagePipelineConfig(stage_id=2, model_stage="stage2"),
+        StagePipelineConfig(stage_id=2, model_stage="stage2", final_output=True),
    ),
 )

@@ -167,6 +167,7 @@ _MODEL_PROMOTE = (
    "VLLM_GEPARD_LOOKBACK_FRAMES",
    "VLLM_OMNI_FISH_KVCACHE_ATTN",
    "VLLM_OMNI_SANA_WM_STAGE1_TEXT_ENCODER",
+    "VLLM_OMNI_SENSENOVA_PAGED_DECODE",
 )

 _MODEL_REQUEST_SCOPE = (
@@ -47,6 +47,7 @@ from vllm_omni.config.stage_config import (
    build_stage_runtime_overrides,
    load_deploy_config,
    normalize_pipeline_cli_overrides,
+    reconcile_diffusion_attention_overrides,
 )
 from vllm_omni.diffusion.diffusion_kv.config import DiffusionKVCacheMode

@@ -1207,6 +1208,9 @@ def _stage_engine_values(
    if topology.omni_kv_config:
        engine["omni_kv_config"] = _copy_value(topology.omni_kv_config)
    if stage_cli_overrides:
+        if topology.execution_type == StageExecutionType.DIFFUSION:
+            # Mirror StageConfig.to_omegaconf so both projections resolve alike.
+            reconcile_diffusion_attention_overrides(engine, stage_cli_overrides)
        engine.update(_copy_value(stage_cli_overrides))
    _validate_stage_engine_override_ownership(
        topology.stage_id,
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Pipeline registry and factory for vllm-omni.

 ``OMNI_PIPELINES`` maps each ``model_type`` to either a ``PipelineConfig``
@@ -196,18 +196,13 @@ def register_pipeline(pipeline: PipelineConfig | PipelineResolverFunc, model_typ
    since resolvers can return multiple different PipelineConfigs depending on the
    consumed config.
    """
-    errors: list[str] = []
    if isinstance(pipeline, PipelineConfig):
-        errors = pipeline.validate()
        model_type = model_type if model_type is not None else pipeline.model_type
-    else:
-        if model_type is None:
-            raise ValueError("Model type must be explicitly provided when registering a pipeline resolver")
+    elif model_type is None:
+        raise ValueError("Model type must be explicitly provided when registering a pipeline resolver")

    if model_type in OMNI_PIPELINES:
-        errors.append(f"Model type {model_type} is already registered; the old mapping will be clobbered")
-    if errors:
-        logger.warning("Registration for pipeline of type %s produced the following issues: %s", model_type, errors)
+        logger.warning(f"Model type {model_type} is already registered; the old mapping will be clobbered")
    OMNI_PIPELINES[model_type] = pipeline


@@ -7,7 +7,7 @@ from __future__ import annotations
 import functools
 import re
 import warnings
-from collections.abc import Callable
+from collections.abc import Callable, Mapping
 from dataclasses import asdict, dataclass, field, fields
 from enum import Enum
 from pathlib import Path
@@ -161,6 +161,28 @@ def _apply_diffusion_parallel_runtime_overrides(
        engine_args["parallel_config"] = parallel_config_dict


+def reconcile_diffusion_attention_overrides(
+    engine_args: dict[str, Any],
+    runtime_overrides: Mapping[str, Any],
+) -> None:
+    """Apply CLI precedence across the two diffusion attention representations.
+
+    ``diffusion_attention_backend`` and ``diffusion_attention_config.default``
+    express the same setting and are rejected downstream when both are set, so
+    a CLI value in one form replaces the YAML value in the other.
+    """
+    if runtime_overrides.get("diffusion_attention_backend") is not None:
+        yaml_config = engine_args.get("diffusion_attention_config")
+        if isinstance(yaml_config, Mapping) and yaml_config.get("default") is not None:
+            remaining = {k: v for k, v in yaml_config.items() if k != "default"}
+            if remaining:
+                engine_args["diffusion_attention_config"] = remaining
+            else:
+                engine_args.pop("diffusion_attention_config")
+    if runtime_overrides.get("diffusion_attention_config") is not None:
+        engine_args.pop("diffusion_attention_backend", None)
+
+
 class StageType(str, Enum):
    """Type of processing stage in the Omni pipeline."""

@@ -302,6 +324,17 @@ class PipelineConfig:
    # Global CLI spelling -> (stage id, stage-local spelling).
    stage_cli_aliases: dict[str, tuple[int, str]] = field(default_factory=dict)

+    def __post_init__(self) -> None:
+        errors = self.get_validation_errors()
+
+        if errors:
+            # If we have multiple errors, put them on separate indented lines for readability
+            multi_sep = "\n\t"
+            error_str = multi_sep.join(errors)
+            if len(errors) > 1:
+                error_str = f"{multi_sep}{error_str}"
+            raise ValueError(f"PipelineConfig initialization failed with the following error(s): {error_str}")
+
    def get_stage(self, stage_id: int) -> StagePipelineConfig | None:
        """Look up a stage by its ID."""
        for stage in self.stages:
@@ -309,7 +342,7 @@ class PipelineConfig:
                return stage
        return None

-    def validate(self) -> list[str]:
+    def get_validation_errors(self) -> list[str]:
        """Return list of topology errors (empty if valid)."""
        errors: list[str] = []
        if not self.stages:
@@ -1081,6 +1114,7 @@ class StageConfig:

        if StageType(self.stage_type) == StageType.DIFFUSION:
            _apply_diffusion_parallel_runtime_overrides(engine_args, runtime_overrides)
+            reconcile_diffusion_attention_overrides(engine_args, runtime_overrides)

        # CLI overrides take precedence over YAML defaults
        for key, value in runtime_overrides.items():
@@ -1,7 +1,6 @@
 # Wan2.2 TI2V-5B deploy.
 #
 # Single GPU by default. Increase tensor_parallel_size and devices for A14B MoE variants.
-# Request-level batching uses scheduler max_num_seqs, not diffusion_batch_size.
 async_chunk: false
 trust_remote_code: true
 distributed_executor_backend: mp
@@ -48,7 +48,6 @@ class InlineStageDiffusionClient(StageClientBase):
        model: str,
        od_config: OmniDiffusionConfig,
        metadata: StageMetadata,
-        batch_size: int = 1,
    ) -> None:
        self.model = model
        self.od_config = od_config
@@ -63,7 +62,6 @@ class InlineStageDiffusionClient(StageClientBase):
        self.prompt_transform_func = None
        self.prompt_expand_func = None
        self.engine_input_source = metadata.engine_input_source
-        self.batch_size = batch_size

        self._enrich_config()
        self._engine = DiffusionEngine.make_engine(self.od_config)
@@ -79,10 +77,10 @@ class InlineStageDiffusionClient(StageClientBase):
        self._engine.executor.register_failure_callback(self._mark_engine_dead)

        logger.info(
-            "[InlineStageDiffusionClient] stage-%s [rep-%s] initialized inline (batch_size=%d)",
+            "[InlineStageDiffusionClient] stage-%s [rep-%s] initialized inline (max_num_seqs=%d)",
            self.stage_id,
            self.replica_id,
-            self.batch_size,
+            self.od_config.max_num_seqs,
        )

    def _enrich_config(self) -> None:
@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import os
 from collections import OrderedDict, defaultdict
 from collections.abc import Callable
@@ -41,7 +44,7 @@ def _prepare_lora_delta(
    lora_b_suffix: str = "lora_B.weight",
    lora_bias_suffix: str = "bias",
 ):
-    used_keys = set()
+    used_keys: set[str] = set()
    # stacked_params_mapping                       param_to_weight_names
    # [(".to_qkv", ".to_q.", "q")
    # (".to_qkv", ".to_k.", "k")  ========> {".to_qkv": [".to_q", ".to_k", ".to_v"]}
@@ -54,7 +57,10 @@ def _prepare_lora_delta(
    if param_to_weight_names is None:
        param_to_weight_names = defaultdict(list)
    for param_name, weight_names in param_to_weight_names.items():
-        if param_name not in base_key:
+        # Fused names can overlap by prefix (`.qkv_proj` is a substring of
+        # `.qkv_proj_mot_gen`), and a substring match stacks both mappings into
+        # one delta of twice the height.
+        if not base_key.endswith(param_name):
            continue
        is_stacked_param = True
        # handle lora_a_key and lora_b_key together
@@ -78,7 +84,7 @@ def _prepare_lora_delta(
                stacked_deltas.append(delta)
                used_keys.add(lora_a_key)
                used_keys.add(lora_b_key)
-        continue
+        break

    if is_stacked_param:
        return torch.concat(stacked_deltas), used_keys
@@ -200,6 +206,10 @@ def _apply_diffusers_lora_alpha_scaling(state_dict: dict[str, torch.Tensor]) ->
 class LoraLoaderMixin:
    transformer_name = "transformer"

+    # Set by the pipeline this mixin is combined with; annotated, not assigned.
+    transformer: torch.nn.Module
+    _lora_loaded: dict[str | None, dict[str, torch.Tensor]]
+
    # Lazy initialization to avoid MRO issues: __init__ may not be called
    # when mixin with nn.Module
    @property
@@ -373,6 +383,7 @@ class QwenImageLoraLoaderMixin(LoraLoaderMixin):

 class WanLoraLoaderMixin(LoraLoaderMixin):
    transformer_2_name = "transformer_2"
+    has_transformer_2: bool

    def load_lora_weights(
        self,
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project

 """Native vLLM-Omni pipeline for Boogu-Image-0.1.

@@ -220,7 +220,7 @@ class BooguImagePipeline(nn.Module, ProgressBarMixin, SupportsComponentDiscovery
            DiffusersPipelineLoader.ComponentSource(
                model_or_path=od_config.model,
                subfolder="transformer",
-                revision=None,
+                revision=od_config.revision,
                prefix="transformer.",
                fall_back_to_pt=True,
            )
@@ -233,10 +233,18 @@ class BooguImagePipeline(nn.Module, ProgressBarMixin, SupportsComponentDiscovery
        # See ``hub_prefetch.py`` for the transformers v5 multi-worker subfolder
        # race; prefetch the whole component set before any from_pretrained.
        boogu_subfolders = ["scheduler", "vae", "mllm", "processor"]
-        prefetch_subfolders(model, boogu_subfolders, local_files_only=local_files_only)
+        prefetch_subfolders(
+            model,
+            boogu_subfolders,
+            local_files_only=local_files_only,
+            revision=od_config.revision,
+        )

        self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
-            model, subfolder="scheduler", local_files_only=local_files_only
+            model,
+            subfolder="scheduler",
+            local_files_only=local_files_only,
+            revision=od_config.revision,
        )

        mllm = from_pretrained_with_prefetch(
@@ -246,6 +254,7 @@ class BooguImagePipeline(nn.Module, ProgressBarMixin, SupportsComponentDiscovery
            prefetch_list=boogu_subfolders,
            local_files_only=local_files_only,
            torch_dtype=od_config.dtype,
+            revision=od_config.revision,
        )
        # Upstream reuses the full VLM as an optional instruction rewriter and
        # encodes with its inner model (no ``lm_head``); the rewriter is not
@@ -255,7 +264,10 @@ class BooguImagePipeline(nn.Module, ProgressBarMixin, SupportsComponentDiscovery
        self.mllm = mllm.to(self._execution_device)

        self.processor = Qwen3VLProcessor.from_pretrained(
-            model, subfolder="processor", local_files_only=local_files_only
+            model,
+            subfolder="processor",
+            local_files_only=local_files_only,
+            revision=od_config.revision,
        )

        self.vae = from_pretrained_with_prefetch(
@@ -264,6 +276,7 @@ class BooguImagePipeline(nn.Module, ProgressBarMixin, SupportsComponentDiscovery
            subfolder="vae",
            prefetch_list=boogu_subfolders,
            local_files_only=local_files_only,
+            revision=od_config.revision,
        ).to(self._execution_device)

        self.transformer = BooguImageTransformer2DModel(od_config=od_config)
@@ -2223,7 +2223,7 @@ class Cosmos3OmniDiffusersPipeline(
        if prompt_suffix:
            prompt = f"{prompt.rstrip()} {prompt_suffix.lstrip()}".strip()
        if _is_rank_zero():
-            logger.info("Final prompt: '%s'", prompt)
+            logger.debug("Final prompt: '%s'", prompt)

        if negative_metadata_mode == "none":
            negative_dur_tmpl = None
@@ -0,0 +1,391 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Model-local coordination for MiniMax H3 temporal chunk decoding.
+
+This boundary owns released-code compatibility and reconciles preflight and
+callback failures at fixed VAE-group rendezvous points. Cropping,
+representation conversion, transport, and encoding remain runtime
+responsibilities.
+"""
+
+from __future__ import annotations
+
+from contextlib import AbstractContextManager
+from typing import Protocol
+
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+
+from .temporal_chunks import (
+    MiniMaxH3TemporalChunkCompatibilityError,
+    MiniMaxH3TemporalDecodePlan,
+    _TemporalChunkCallbackSignature,
+    decode_minimax_h3_temporal_chunks_compat,
+    prepare_minimax_h3_temporal_chunks_compat,
+)
+
+
+class MiniMaxH3VideoChunkCallback(_TemporalChunkCallbackSignature, Protocol):
+    """Consume one ordered, temporally committed H3 video chunk.
+
+    ``frames`` is an owned, contiguous float32 RGB tensor in ``BCTHW`` layout
+    and the ``[0, 1]`` value range. The callback runs synchronously on the
+    producer's current device stream. Before reading it on another stream, the
+    consumer must establish ordering from the producer stream (for example via
+    an event) and retain the allocation for that stream's lifetime. The
+    consumer may retain or mutate the tensor; it does not alias the returned
+    complete decode.
+    """
+
+
+class MiniMaxH3ChunkedDecodeUnsupportedError(RuntimeError):
+    """Raised when loaded MiniMax H3 code cannot honor the chunk contract."""
+
+
+class MiniMaxH3ChunkCallbackPeerError(RuntimeError):
+    """Raised on peer VAE ranks when the output-rank callback fails."""
+
+
+class _MiniMaxH3ChunkDecodeHost(Protocol):
+    """Narrow VAE capabilities required by the chunk coordinator."""
+
+    model: nn.Module
+
+    def _denormalize_latent(self, latent: torch.Tensor) -> torch.Tensor: ...
+
+    def _decode_tiling_context(
+        self,
+        latent: torch.Tensor,
+    ) -> AbstractContextManager: ...
+
+    def _normalize_decoded_frames(
+        self,
+        decoded: torch.Tensor,
+    ) -> torch.Tensor: ...
+
+
+class _MiniMaxH3ChunkDecodeCoordinator:
+    """Run source-gated chunk decode with group-local failure semantics.
+
+    The coordinator is a plain Python object. It caches only successful source
+    validation; request callbacks, tensors, metadata, and process groups remain
+    local to :meth:`decode`.
+
+    Errors raised inside the checkpoint decoder after native collectives begin
+    cannot be reconciled here: another rank may already be blocked inside that
+    collective. Resolving such a failure requires native collective handling or
+    supervisor-level timeout and worker-group teardown; the process group cannot
+    be assumed reusable.
+    """
+
+    def __init__(self) -> None:
+        self._sources_validated = False
+
+    @staticmethod
+    def _build_execution_signature(
+        latent: torch.Tensor,
+        plan: MiniMaxH3TemporalDecodePlan,
+    ) -> tuple[int, ...]:
+        """Describe every shape and plan field used by the temporal schedule."""
+
+        dtype_codes: dict[torch.dtype | None, int] = {
+            None: 0,
+            torch.float16: 1,
+            torch.bfloat16: 2,
+            torch.float32: 3,
+            torch.float64: 4,
+        }
+        return (
+            # Keep the layout versioned so adding a field cannot silently make
+            # an older and newer worker appear compatible.
+            1,
+            *[int(size) for size in latent.shape],
+            dtype_codes.get(latent.dtype, -1),
+            *[int(size) for size in plan.latent.shape],
+            dtype_codes.get(plan.latent.dtype, -1),
+            int(plan.num_chunks),
+            int(plan.total_frames),
+            int(plan.pad_frames),
+            int(plan.output_frames),
+            len(plan.chunk_frames),
+            *[int(frame_count) for frame_count in plan.chunk_frames],
+            dtype_codes.get(plan.temporal_dtype, -1),
+        )
+
+    @staticmethod
+    def _is_output_rank(
+        chunk_callback: MiniMaxH3VideoChunkCallback | None,
+        *,
+        group: dist.ProcessGroup | None,
+        device: torch.device,
+    ) -> bool:
+        if group is None:
+            return chunk_callback is not None
+
+        group_rank = dist.get_rank(group)
+        # A valid configuration has one contribution: rank 0 contributes 1.
+        # A non-zero owner contributes at least 2, and multiple owners sum to
+        # at least 3, so every rank makes the same decision from one reduction.
+        callback_owner = torch.tensor(
+            [group_rank + 1 if chunk_callback is not None else 0],
+            dtype=torch.int64,
+            device=device,
+        )
+        dist.all_reduce(callback_owner, op=dist.ReduceOp.SUM, group=group)
+        if int(callback_owner.item()) != 1:
+            raise ValueError("MiniMax-H3 chunk decode requires a callback on exactly VAE group rank 0")
+        return group_rank == 0
+
+    @staticmethod
+    def _synchronize_callback_error(
+        callback_error: BaseException | None,
+        *,
+        group: dist.ProcessGroup | None,
+        device: torch.device,
+    ) -> None:
+        if group is None:
+            if callback_error is not None:
+                raise callback_error
+            return
+
+        is_output_rank = dist.get_rank(group) == 0
+        failed = torch.tensor(
+            [int(is_output_rank and callback_error is not None)],
+            dtype=torch.int32,
+            device=device,
+        )
+        dist.broadcast(
+            failed,
+            src=dist.get_global_rank(group, 0),
+            group=group,
+        )
+        if int(failed.item()) == 0:
+            return
+        if is_output_rank:
+            assert callback_error is not None
+            raise callback_error
+        raise MiniMaxH3ChunkCallbackPeerError("MiniMax-H3 video chunk callback failed on this VAE group's output rank")
+
+    @staticmethod
+    def _synchronize_preflight_error(
+        preflight_error: Exception | None,
+        temporal_dtype: torch.dtype | None,
+        execution_signature: tuple[int, ...] | None,
+        *,
+        group: dist.ProcessGroup | None,
+        device: torch.device,
+    ) -> None:
+        if group is None:
+            if preflight_error is not None:
+                raise preflight_error
+            return
+        failed = torch.tensor(
+            [int(preflight_error is not None)],
+            dtype=torch.int32,
+            device=device,
+        )
+        dist.all_reduce(failed, op=dist.ReduceOp.MAX, group=group)
+        if int(failed.item()) == 0:
+            dtype_codes = {
+                None: 0,
+                torch.float16: 1,
+                torch.bfloat16: 2,
+                torch.float32: 3,
+            }
+            code = torch.tensor(
+                [dtype_codes.get(temporal_dtype, -1)],
+                dtype=torch.int32,
+                device=device,
+            )
+            lowest = code.clone()
+            highest = code.clone()
+            dist.all_reduce(lowest, op=dist.ReduceOp.MIN, group=group)
+            dist.all_reduce(highest, op=dist.ReduceOp.MAX, group=group)
+            if int(lowest.item()) == int(highest.item()) and int(lowest.item()) >= 0:
+                if execution_signature is None:
+                    raise RuntimeError("MiniMax-H3 temporal decode preflight produced no execution signature")
+
+                # The plan length is shape-dependent, so first agree on a
+                # common tensor size and then compare the complete signatures.
+                # This keeps all ranks in tensor collectives and catches a
+                # schedule mismatch before entering the decoder's collectives.
+                signature_length = torch.tensor(
+                    [len(execution_signature)],
+                    dtype=torch.int64,
+                    device=device,
+                )
+                max_signature_length = signature_length.clone()
+                dist.all_reduce(
+                    max_signature_length,
+                    op=dist.ReduceOp.MAX,
+                    group=group,
+                )
+                padded_signature = torch.full(
+                    (int(max_signature_length.item()),),
+                    -2,
+                    dtype=torch.int64,
+                    device=device,
+                )
+                padded_signature[: len(execution_signature)] = torch.tensor(
+                    execution_signature,
+                    dtype=torch.int64,
+                    device=device,
+                )
+                gathered_signatures = [torch.empty_like(padded_signature) for _ in range(dist.get_world_size(group))]
+                dist.all_gather(gathered_signatures, padded_signature, group=group)
+                if any(not torch.equal(padded_signature, peer) for peer in gathered_signatures):
+                    raise MiniMaxH3ChunkedDecodeUnsupportedError(
+                        "MiniMax-H3 VAE ranks built incompatible temporal decode plans"
+                    )
+                return
+            raise MiniMaxH3ChunkedDecodeUnsupportedError(
+                "MiniMax-H3 VAE ranks selected incompatible temporal concat dtypes"
+            )
+        if preflight_error is not None:
+            raise preflight_error
+        raise MiniMaxH3ChunkedDecodeUnsupportedError("A peer VAE rank rejected the MiniMax-H3 temporal chunk preflight")
+
+    def decode(
+        self,
+        host: _MiniMaxH3ChunkDecodeHost,
+        latent: torch.Tensor,
+        chunk_callback: MiniMaxH3VideoChunkCallback | None,
+        *,
+        group: dist.ProcessGroup | None,
+    ) -> torch.Tensor:
+        """Decode one latent while publishing committed normalized chunks."""
+
+        plan: MiniMaxH3TemporalDecodePlan | None = None
+        preflight_error: Exception | None = None
+        if latent.ndim != 5:
+            preflight_error = MiniMaxH3ChunkedDecodeUnsupportedError(
+                f"MiniMax-H3 video latent must be rank 5, got {tuple(latent.shape)}"
+            )
+        else:
+            try:
+                denormalized = host._denormalize_latent(latent)
+                plan = prepare_minimax_h3_temporal_chunks_compat(
+                    host.model,
+                    denormalized,
+                    validate_sources=not self._sources_validated,
+                )
+                self._sources_validated = True
+                del denormalized
+            except MiniMaxH3TemporalChunkCompatibilityError as exc:
+                preflight_error = MiniMaxH3ChunkedDecodeUnsupportedError(str(exc))
+            except Exception as exc:  # noqa: BLE001
+                # Every VAE rank must reach the preflight reduction even when
+                # one node's private planner or environment fails locally.
+                exc.__traceback__ = None
+                preflight_error = exc
+        self._synchronize_preflight_error(
+            preflight_error,
+            plan.temporal_dtype if plan is not None else None,
+            self._build_execution_signature(latent, plan) if plan is not None else None,
+            group=group,
+            device=latent.device,
+        )
+        assert plan is not None
+        is_output_rank = self._is_output_rank(
+            chunk_callback,
+            group=group,
+            device=latent.device,
+        )
+        callback_error: BaseException | None = None
+        next_chunk_index = 0
+        expected_total_chunks: int | None = None
+        next_frame_start = 0
+        saw_final = False
+
+        def publish(
+            decoded_chunk: torch.Tensor,
+            *,
+            chunk_index: int,
+            total_chunks: int,
+            frame_start: int,
+            is_final: bool,
+        ) -> None:
+            nonlocal callback_error, expected_total_chunks
+            nonlocal next_chunk_index, next_frame_start, saw_final
+            if not is_output_rank or callback_error is not None:
+                return
+            try:
+                if expected_total_chunks is None:
+                    expected_total_chunks = total_chunks
+                if (
+                    saw_final
+                    or total_chunks <= 0
+                    or total_chunks != expected_total_chunks
+                    or chunk_index != next_chunk_index
+                    or chunk_index >= total_chunks
+                    or frame_start != next_frame_start
+                    or is_final != (chunk_index + 1 == total_chunks)
+                ):
+                    raise RuntimeError(
+                        "MiniMax-H3 temporal backend emitted discontinuous "
+                        "chunk metadata: "
+                        f"index={chunk_index}/{next_chunk_index}, "
+                        f"total={total_chunks}/{expected_total_chunks}, "
+                        f"frame_start={frame_start}/{next_frame_start}, "
+                        f"final={is_final}, after_final={saw_final}"
+                    )
+                frames = host._normalize_decoded_frames(decoded_chunk)
+                frames = frames.clone(memory_format=torch.contiguous_format)
+                frame_count = int(frames.shape[2])
+                if frame_count <= 0:
+                    raise RuntimeError("MiniMax-H3 temporal backend emitted an empty chunk")
+                next_chunk_index += 1
+                next_frame_start += frame_count
+                saw_final = is_final
+                assert chunk_callback is not None
+                chunk_callback(
+                    frames,
+                    chunk_index=chunk_index,
+                    total_chunks=total_chunks,
+                    frame_start=frame_start,
+                    is_final=is_final,
+                )
+            except BaseException as exc:  # noqa: BLE001
+                # Keep a failed GPU chunk and its callback locals out of the
+                # traceback while peer ranks finish the remaining collectives.
+                exc.__traceback__ = None
+                callback_error = exc
+
+        # Do not catch and reduce exceptions from this region. A peer may
+        # already be blocked in one of the checkpoint's native tile
+        # collectives, in which case a new adapter-level collective would also
+        # deadlock. A native backend or supervisor must handle timeout and
+        # worker-group teardown.
+        with host._decode_tiling_context(latent):
+            decoded = decode_minimax_h3_temporal_chunks_compat(
+                host.model,
+                plan,
+                publish,
+            )
+        frames = host._normalize_decoded_frames(decoded)
+        if (
+            is_output_rank
+            and callback_error is None
+            and (not saw_final or next_chunk_index != expected_total_chunks or next_frame_start != int(frames.shape[2]))
+        ):
+            callback_error = RuntimeError(
+                "MiniMax-H3 temporal backend did not reconstruct the full "
+                "decode: "
+                f"final={saw_final}, "
+                f"chunks={next_chunk_index}/{expected_total_chunks}, "
+                f"frames={next_frame_start}/{frames.shape[2]}"
+            )
+        self._synchronize_callback_error(
+            callback_error,
+            group=group,
+            device=frames.device,
+        )
+        return frames
+
+
+__all__ = [
+    "MiniMaxH3ChunkCallbackPeerError",
+    "MiniMaxH3ChunkedDecodeUnsupportedError",
+    "MiniMaxH3VideoChunkCallback",
+]
@@ -0,0 +1,414 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Source-gated temporal chunk compatibility for the released MiniMax H3 VAE."""
+
+from __future__ import annotations
+
+import hashlib
+import importlib
+import inspect
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Protocol
+
+import torch
+import torch.nn as nn
+
+MINIMAX_H3_RELEASED_REVISION = "42ed227ee7df40d41602854ae760620d6eb651fe"
+# Compatibility pins detect upstream code drift; they are not a trust or
+# security boundary for remote code execution.
+MINIMAX_H3_RELEASED_BUNDLE_SHA256 = "30e64afbaa940696bf8bfe2c86003a4b1666c84b22fa3aa12403a3e6a03f705d"
+
+_RELEASED_TEMPORAL_CONFIG = {
+    "clip_length": 17,
+    "tokens_chunk_size": 5,
+    "token_overlap": 2,
+    "token_drop": 3,
+    "vae_ratio_t": 4,
+    "frame_pre_padding": 3,
+    "frame_overlap": 5,
+    "isolated_first_frame": False,
+    "isolated_last_frame": False,
+}
+
+
+class MiniMaxH3TemporalChunkCompatibilityError(RuntimeError):
+    """Raised when remote H3 code cannot use the local compatibility path."""
+
+
+class _TemporalChunkCallbackSignature(Protocol):
+    """Shared callable shape; each producer defines tensor semantics."""
+
+    def __call__(
+        self,
+        frames: torch.Tensor,
+        /,
+        *,
+        chunk_index: int,
+        total_chunks: int,
+        frame_start: int,
+        is_final: bool,
+    ) -> None: ...
+
+
+@dataclass(slots=True)
+class MiniMaxH3TemporalDecodePlan:
+    latent: torch.Tensor
+    num_chunks: int
+    total_frames: int
+    pad_frames: int
+    output_frames: int
+    chunk_frames: tuple[int, ...]
+    temporal_dtype: torch.dtype | None
+
+
+class _RawChunkAssembly:
+    def __init__(
+        self,
+        *,
+        callback: _TemporalChunkCallbackSignature,
+        plan: MiniMaxH3TemporalDecodePlan,
+    ) -> None:
+        self.callback = callback
+        self.plan = plan
+        self.frames: torch.Tensor | None = None
+        self.write_pos = 0
+        self.logical_frames = 0
+        self.dropped_frames = 0
+        self.chunk_index = 0
+        self.pending_final: tuple[torch.Tensor, int, int] | None = None
+        self.contract_error: RuntimeError | None = None
+
+    def write(self, part: torch.Tensor) -> None:
+        part_frames = int(part.shape[2])
+        if part_frames <= 0:
+            return
+        self.logical_frames += part_frames
+        remaining = self.plan.output_frames - self.write_pos
+        copy_frames = min(part_frames, max(0, remaining))
+        if copy_frames <= 0:
+            self.dropped_frames += part_frames
+            return
+        if self.chunk_index >= len(self.plan.chunk_frames):
+            self.contract_error = RuntimeError("MiniMax-H3 compatibility decode emitted an extra chunk")
+        else:
+            expected_frames = self.plan.chunk_frames[self.chunk_index]
+            if copy_frames != expected_frames and self.contract_error is None:
+                self.contract_error = RuntimeError(
+                    "MiniMax-H3 compatibility chunk size mismatch: "
+                    f"chunk={self.chunk_index}, frames={copy_frames}/{expected_frames}"
+                )
+
+        emitted = part[:, :, :copy_frames]
+        if self.frames is None:
+            output_shape = list(emitted.shape)
+            output_shape[2] = self.plan.output_frames
+            self.frames = torch.empty(
+                output_shape,
+                dtype=emitted.dtype,
+                device=emitted.device,
+            )
+        frame_start = self.write_pos
+        frame_stop = frame_start + copy_frames
+        self.frames[:, :, frame_start:frame_stop].copy_(emitted)
+        if frame_stop == self.plan.output_frames:
+            self.pending_final = (emitted, self.chunk_index, frame_start)
+        elif self.contract_error is None:
+            self.callback(
+                emitted,
+                chunk_index=self.chunk_index,
+                total_chunks=len(self.plan.chunk_frames),
+                frame_start=frame_start,
+                is_final=False,
+            )
+        self.chunk_index += 1
+        self.write_pos = frame_stop
+        self.dropped_frames += part_frames - copy_frames
+
+    def finish(self) -> torch.Tensor:
+        if self.contract_error is not None:
+            raise self.contract_error
+        if self.frames is None or self.pending_final is None:
+            raise RuntimeError("MiniMax-H3 compatibility decode produced no terminal chunk")
+        if (
+            self.logical_frames != self.plan.total_frames
+            or self.dropped_frames != self.plan.pad_frames
+            or self.write_pos != self.plan.output_frames
+            or self.chunk_index != len(self.plan.chunk_frames)
+        ):
+            raise RuntimeError(
+                "MiniMax-H3 compatibility frame plan mismatch: "
+                f"logical={self.logical_frames}/{self.plan.total_frames}, "
+                f"dropped={self.dropped_frames}/{self.plan.pad_frames}, "
+                f"written={self.write_pos}/{self.plan.output_frames}, "
+                f"chunks={self.chunk_index}/{len(self.plan.chunk_frames)}"
+            )
+
+        final_frames, final_index, final_start = self.pending_final
+        self.callback(
+            final_frames,
+            chunk_index=final_index,
+            total_chunks=len(self.plan.chunk_frames),
+            frame_start=final_start,
+            is_final=True,
+        )
+        return self.frames
+
+
+def _source_bundle_sha256(model: nn.Module) -> tuple[str | None, str | None]:
+    try:
+        source = inspect.getsourcefile(type(model))
+    except TypeError:
+        return None, None
+    if source is None:
+        return None, None
+    package_dir = Path(source).parent
+    files = sorted(path for path in package_dir.glob("*.py") if path.name != "__init__.py")
+    digest = hashlib.sha256()
+    try:
+        for path in files:
+            digest.update(path.name.encode())
+            digest.update(b"\0")
+            digest.update(path.read_bytes())
+            digest.update(b"\0")
+    except OSError:
+        return str(package_dir), None
+    return str(package_dir), digest.hexdigest()
+
+
+def _validate_released_sources(model: nn.Module) -> None:
+    source_dir, actual = _source_bundle_sha256(model)
+    if actual != MINIMAX_H3_RELEASED_BUNDLE_SHA256:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            "The vLLM-owned MiniMax-H3 chunk adapter supports only remote code "
+            f"from MiniMaxAI/MiniMax-H3@{MINIMAX_H3_RELEASED_REVISION}: "
+            f"bundle={actual!r} ({source_dir}). "
+            "The normal complete-output decode remains available."
+        )
+
+
+def _validate_released_config(model: nn.Module) -> None:
+    actual = {name: getattr(model, name, None) for name in _RELEASED_TEMPORAL_CONFIG}
+    if actual != _RELEASED_TEMPORAL_CONFIG:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            "The vLLM-owned MiniMax-H3 chunk adapter requires the released temporal "
+            f"configuration {_RELEASED_TEMPORAL_CONFIG}, found {actual}. "
+            "The normal complete-output decode remains available."
+        )
+
+
+def _validate_compat_contract(model: nn.Module, latent: torch.Tensor) -> None:
+    if latent.ndim != 5 or int(latent.shape[2]) <= 0:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            f"MiniMax-H3 video latent must have positive rank-5 time, got {tuple(latent.shape)}"
+        )
+    if not bool(getattr(model, "use_3d_conv", False)):
+        raise MiniMaxH3TemporalChunkCompatibilityError("MiniMax-H3 temporal chunk decode requires the 3D decoder")
+    if bool(getattr(model, "training", False)):
+        raise MiniMaxH3TemporalChunkCompatibilityError("MiniMax-H3 temporal chunk decode is inference-only")
+    required_methods = (
+        "_adaptive_decode",
+        "_decode_temporal_output_frame_plan",
+        "_decode_temporal_pad_frames",
+        "blend",
+    )
+    missing = [name for name in required_methods if not callable(getattr(model, name, None))]
+    if missing:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            "Loaded MiniMax-H3 VAE lacks required released-code methods: " + ", ".join(missing)
+        )
+
+
+def _truncate_chunk_frames(
+    logical_chunk_frames: list[int],
+    output_frames: int,
+) -> tuple[int, ...]:
+    remaining = output_frames
+    result: list[int] = []
+    for logical_frames in logical_chunk_frames:
+        emitted_frames = min(logical_frames, max(0, remaining))
+        if emitted_frames > 0:
+            result.append(emitted_frames)
+            remaining -= emitted_frames
+    if remaining != 0:
+        raise RuntimeError(f"MiniMax-H3 compatibility chunk plan cannot cover output: remaining={remaining}")
+    return tuple(result)
+
+
+def _build_decode_plan(
+    model: nn.Module,
+    latent: torch.Tensor,
+    temporal_dtype: torch.dtype | None,
+) -> MiniMaxH3TemporalDecodePlan:
+    pseudo_total_tokens = int(latent.shape[2]) + int(model.token_drop)  # type: ignore[attr-defined]
+    tokens_chunk_size = int(model.tokens_chunk_size)  # type: ignore[attr-defined]
+    remainder = pseudo_total_tokens % tokens_chunk_size
+    pad_tokens = 0 if remainder == 0 else tokens_chunk_size - remainder
+    pseudo_total_tokens += pad_tokens
+    num_chunks = pseudo_total_tokens // tokens_chunk_size - int(model.token_drop > 0)  # type: ignore[attr-defined]
+    if num_chunks <= 0:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            f"MiniMax-H3 temporal chunk plan is empty for latent T={latent.shape[2]}"
+        )
+
+    temporal_latent = latent
+    if pad_tokens:
+        pad = temporal_latent[:, :, -1:].repeat(1, 1, pad_tokens, 1, 1)
+        temporal_latent = torch.cat((temporal_latent, pad), dim=2)
+    total_frames, pad_frames, output_frames = model._decode_temporal_output_frame_plan(  # type: ignore[attr-defined]
+        temporal_latent,
+        None,
+        None,
+        num_chunks,
+        pad_tokens,
+    )
+    if output_frames <= 0:
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            f"MiniMax-H3 temporal chunk plan has {output_frames} output frames"
+        )
+
+    chunk_decoded_frames = tokens_chunk_size * int(model.vae_ratio_t)  # type: ignore[attr-defined]
+    split_count = int(model.token_drop > 0) + 1  # type: ignore[attr-defined]
+    logical_chunk_frames: list[int] = []
+    final_overlap_frames = 0
+    for temporal_index in range(num_chunks):
+        token_start = temporal_index * tokens_chunk_size
+        token_stop = token_start + tokens_chunk_size + int(model.token_overlap)  # type: ignore[attr-defined]
+        clip_tokens = max(
+            0,
+            min(token_stop, int(temporal_latent.shape[2])) - min(token_start, int(temporal_latent.shape[2])),
+        )
+        clip_frames = clip_tokens * int(model.vae_ratio_t)  # type: ignore[attr-defined]
+        for split_index in range(split_count):
+            frame_start = split_index * chunk_decoded_frames
+            frame_stop = min(frame_start + chunk_decoded_frames, clip_frames)
+            part_frames = max(
+                0,
+                frame_stop - frame_start - int(model.frame_pre_padding),  # type: ignore[attr-defined]
+            )
+            if split_index == 0:
+                if part_frames > 0:
+                    logical_chunk_frames.append(part_frames)
+            else:
+                final_overlap_frames = part_frames
+    if final_overlap_frames > 0:
+        logical_chunk_frames.append(final_overlap_frames)
+    if sum(logical_chunk_frames) != int(total_frames):
+        raise RuntimeError(
+            "MiniMax-H3 compatibility planner disagrees with released remote code: "
+            f"frames={sum(logical_chunk_frames)}/{total_frames}"
+        )
+    chunk_frames = _truncate_chunk_frames(logical_chunk_frames, int(output_frames))
+    return MiniMaxH3TemporalDecodePlan(
+        latent=temporal_latent,
+        num_chunks=num_chunks,
+        total_frames=int(total_frames),
+        pad_frames=int(pad_frames),
+        output_frames=int(output_frames),
+        chunk_frames=chunk_frames,
+        temporal_dtype=temporal_dtype,
+    )
+
+
+def resolve_minimax_h3_temporal_cat_dtype(model: nn.Module) -> torch.dtype | None:
+    module = importlib.import_module(type(model).__module__)
+    resolver = getattr(module, "_resolve_temporal_cat_dtype", None)
+    if resolver is None:
+        return None
+    if not callable(resolver):
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            "Loaded MiniMax-H3 VAE has a non-callable temporal dtype resolver"
+        )
+    try:
+        dtype = resolver()
+    except ValueError as exc:
+        raise MiniMaxH3TemporalChunkCompatibilityError(str(exc)) from None
+    if dtype is not None and not isinstance(dtype, torch.dtype):
+        raise MiniMaxH3TemporalChunkCompatibilityError(
+            f"Loaded MiniMax-H3 VAE returned invalid temporal dtype {dtype!r}"
+        )
+    return dtype
+
+
+def decode_minimax_h3_temporal_chunks_compat(
+    model: nn.Module,
+    plan: MiniMaxH3TemporalDecodePlan,
+    callback: _TemporalChunkCallbackSignature,
+) -> torch.Tensor:
+    """Mirror the released Apache-2.0 H3 temporal assembler with callbacks.
+
+    Keep this implementation and the source/config fingerprints above in sync
+    with ``MiniMaxAI/MiniMax-H3@42ed227e``. Unknown remote code fails closed
+    before this function is called.
+
+    Callback tensors are borrowed decoder-domain views and must not be retained
+    or mutated. The high-level adapter owns normalization and storage isolation.
+    """
+
+    chunk_decoded_frames = int(model.tokens_chunk_size) * int(model.vae_ratio_t)  # type: ignore[attr-defined]
+    split_count = int(model.token_drop > 0) + 1  # type: ignore[attr-defined]
+    assembly = _RawChunkAssembly(callback=callback, plan=plan)
+    overlap: torch.Tensor | None = None
+
+    for temporal_index in range(plan.num_chunks):
+        token_start = temporal_index * int(model.tokens_chunk_size)  # type: ignore[attr-defined]
+        token_stop = token_start + int(model.tokens_chunk_size) + int(model.token_overlap)  # type: ignore[attr-defined]
+        chunk_latent = plan.latent[:, :, token_start:token_stop]
+        decoded = model._adaptive_decode(chunk_latent)  # type: ignore[attr-defined]
+        if plan.temporal_dtype is not None and decoded.dtype != plan.temporal_dtype:
+            decoded = decoded.to(plan.temporal_dtype)
+        if decoded.device != plan.latent.device:
+            decoded = decoded.to(plan.latent.device)
+
+        for split_index in range(split_count):
+            frame_start = split_index * chunk_decoded_frames
+            frame_stop = min(frame_start + chunk_decoded_frames, int(decoded.shape[2]))
+            part = decoded[:, :, frame_start:frame_stop]
+            part = part[:, :, int(model.frame_pre_padding) :]  # type: ignore[attr-defined]
+            if split_index == 0:
+                if overlap is not None:
+                    part = model.blend(  # type: ignore[attr-defined]
+                        overlap,
+                        part,
+                        int(model.frame_overlap),  # type: ignore[attr-defined]
+                        dim=-3,
+                    )
+                    overlap = None
+                assembly.write(part)
+            else:
+                overlap = part.contiguous()
+            del part
+
+        if temporal_index == plan.num_chunks - 1 and overlap is not None:
+            assembly.write(overlap)
+            overlap = None
+        del decoded, chunk_latent
+
+    return assembly.finish()
+
+
+def prepare_minimax_h3_temporal_chunks_compat(
+    model: nn.Module,
+    latent: torch.Tensor,
+    *,
+    validate_sources: bool = True,
+) -> MiniMaxH3TemporalDecodePlan:
+    """Build a validated plan without entering decoder collectives."""
+
+    if validate_sources:
+        _validate_released_sources(model)
+    _validate_released_config(model)
+    _validate_compat_contract(model, latent)
+    temporal_dtype = resolve_minimax_h3_temporal_cat_dtype(model)
+    return _build_decode_plan(model, latent, temporal_dtype)
+
+
+__all__ = [
+    "MINIMAX_H3_RELEASED_REVISION",
+    "MINIMAX_H3_RELEASED_BUNDLE_SHA256",
+    "MiniMaxH3TemporalDecodePlan",
+    "MiniMaxH3TemporalChunkCompatibilityError",
+    "decode_minimax_h3_temporal_chunks_compat",
+    "prepare_minimax_h3_temporal_chunks_compat",
+    "resolve_minimax_h3_temporal_cat_dtype",
+]
@@ -21,12 +21,22 @@ from vllm.logger import init_logger
 from vllm_omni.diffusion.distributed.autoencoders.distributed_vae_executor import (
    DistributedVaeMixin,
 )
-from vllm_omni.diffusion.distributed.parallel_state import get_world_group
+from vllm_omni.diffusion.distributed.parallel_state import (
+    get_data_parallel_world_size,
+    get_world_group,
+    model_parallel_is_initialized,
+)
 from vllm_omni.diffusion.offloader.module_residency import (
    BoundedAllocatorCache,
    PinnedModuleStager,
 )

+from .chunked_decode import (
+    MiniMaxH3ChunkCallbackPeerError,
+    MiniMaxH3ChunkedDecodeUnsupportedError,
+    MiniMaxH3VideoChunkCallback,
+    _MiniMaxH3ChunkDecodeCoordinator,
+)
 from .ops import install_h3_vae_optimizations
 from .packed_tokens import minimax_h3_patchify_video_latent

@@ -153,6 +163,7 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
        self.use_slicing = False
        self.parallel_size = 1
        self.device_module = torch.get_device_module()
+        self._chunk_decode_coordinator: _MiniMaxH3ChunkDecodeCoordinator | None = None

    def load_to_device(self) -> None:
        if self._stager is not None:
@@ -265,6 +276,59 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):
    def is_distributed_enabled(self) -> bool:
        return self.parallel_size > 1 and dist.is_initialized()

+    def _denormalize_latent(self, latent: torch.Tensor) -> torch.Tensor:
+        channels = int(self.config_dict["latent_channels"])
+        mean = torch.tensor(
+            self.config_dict["latents_mean"],
+            device=latent.device,
+            dtype=latent.dtype,
+        ).view(1, channels, 1, 1, 1)
+        std = torch.tensor(
+            self.config_dict["latents_std"],
+            device=latent.device,
+            dtype=latent.dtype,
+        ).view(1, channels, 1, 1, 1)
+        return latent * std + mean
+
+    def _decode_tiling_context(self, latent: torch.Tensor) -> AbstractContextManager:
+        # The checkpoint hands rank r the tiles ``range(r, num_tiles, sp_size)``
+        # and then rejects an empty share inside the gather. A rank with no
+        # tiles raises and leaves the collective while the others block in it
+        # forever, so too few tiles hangs the whole stage rather than failing
+        # it. Tile count depends only on the latent shape, so every rank takes
+        # this branch together.
+        num_tiles = self._decoder_tile_count(latent)
+        if self.parallel_size > 1 and num_tiles < self.parallel_size:
+            logger.warning_once(
+                "MiniMax-H3 VAE decode splits into %d tile(s) but the tile group has "
+                "%d ranks; decoding rank-locally for this shape instead, which is "
+                "slower but avoids ranks without tiles hanging the collective.",
+                num_tiles,
+                self.parallel_size,
+            )
+            return self._rank_local_tiling()
+        return nullcontext()
+
+    def _normalize_decoded_frames(self, decoded: torch.Tensor) -> torch.Tensor:
+        frames = self.model.processor.revert_tensor(decoded)
+        if frames.ndim == 4:
+            frames = frames.unsqueeze(0).transpose(1, 2)
+        if frames.ndim != 5:
+            raise ValueError(f"unexpected decoded video shape {tuple(frames.shape)}")
+        return frames.float()
+
+    def _chunk_process_group(self) -> dist.ProcessGroup | None:
+        if not self.is_distributed_enabled():
+            return None
+        if model_parallel_is_initialized() and get_data_parallel_world_size() > 1:
+            raise MiniMaxH3ChunkedDecodeUnsupportedError(
+                "MiniMax-H3 temporal chunks do not yet support combining data parallelism with VAE patch parallelism"
+            )
+        group = self._native_parallel_state().get("sp_process_group")
+        if group is None or dist.get_world_size(group) != self.parallel_size:
+            raise RuntimeError("MiniMax-H3 VAE chunk decode has an invalid spatial-parallel group")
+        return group
+
    @torch.inference_mode()
    def encode_image(self, image: Image.Image) -> torch.Tensor:
        previous_parallel = self.model.parallel_tiling
@@ -357,44 +421,40 @@ class MiniMaxH3VideoVAE(nn.Module, DistributedVaeMixin):

    @torch.inference_mode()
    def decode_latent(self, latent: torch.Tensor) -> torch.Tensor:
-        channels = int(self.config_dict["latent_channels"])
-        mean = torch.tensor(
-            self.config_dict["latents_mean"],
-            device=latent.device,
-            dtype=latent.dtype,
-        ).view(1, channels, 1, 1, 1)
-        std = torch.tensor(
-            self.config_dict["latents_std"],
-            device=latent.device,
-            dtype=latent.dtype,
-        ).view(1, channels, 1, 1, 1)
-        # The checkpoint hands rank r the tiles ``range(r, num_tiles, sp_size)``
-        # and then rejects an empty share inside the gather. A rank with no
-        # tiles raises and leaves the collective while the others block in it
-        # forever, so too few tiles hangs the whole stage rather than failing
-        # it. Tile count depends only on the latent shape, so every rank takes
-        # this branch together.
-        num_tiles = self._decoder_tile_count(latent)
-        if self.parallel_size > 1 and num_tiles < self.parallel_size:
-            logger.warning_once(
-                "MiniMax-H3 VAE decode splits into %d tile(s) but the tile group has "
-                "%d ranks; decoding rank-locally for this shape instead, which is "
-                "slower but avoids ranks without tiles hanging the collective.",
-                num_tiles,
-                self.parallel_size,
-            )
-            tiling_context: AbstractContextManager = self._rank_local_tiling()
-        else:
-            tiling_context = nullcontext()
+        with self._decode_tiling_context(latent):
+            decoded = self.model.decode_base(self._denormalize_latent(latent))
+        return self._normalize_decoded_frames(decoded)

-        with tiling_context:
-            decoded = self.model.decode_base(latent * std + mean)
-        frames = self.model.processor.revert_tensor(decoded)
-        if frames.ndim == 4:
-            frames = frames.unsqueeze(0).transpose(1, 2)
-        if frames.ndim != 5:
-            raise ValueError(f"unexpected decoded video shape {tuple(frames.shape)}")
-        return frames.float()
+    @torch.inference_mode()
+    def decode_latent_with_chunks(
+        self,
+        latent: torch.Tensor,
+        chunk_callback: MiniMaxH3VideoChunkCallback | None,
+    ) -> torch.Tensor:
+        """Decode a video while publishing committed temporal chunks.
+
+        Every rank collaborating in the configured H3 VAE tile group must call
+        this method, with ``chunk_callback`` supplied only on group rank 0. A
+        callback failure is deferred until the compatibility decoder finishes its
+        remaining collectives, then propagated to every group rank. With
+        ``parallel_size=1`` the runtime designates each request's output owner
+        by supplying a callback there and ``None`` on replica peers.
+
+        This is a model-local producer seam before the typed media boundary in
+        RFC #6541 / PR #6615. Representation conversion, transport,
+        presentation policy, and encoding remain runtime-owned. Chunks retain
+        the VAE canvas dimensions; a runtime consumer must apply the same
+        requested-size crop as the complete H3 pipeline output.
+        """
+
+        if self._chunk_decode_coordinator is None:
+            self._chunk_decode_coordinator = _MiniMaxH3ChunkDecodeCoordinator()
+        return self._chunk_decode_coordinator.decode(
+            self,
+            latent,
+            chunk_callback,
+            group=self._chunk_process_group(),
+        )


 class MiniMaxH3AudioVAE(nn.Module):
@@ -516,4 +576,10 @@ class MiniMaxH3AudioVAE(nn.Module):
        return waveform.permute(1, 0, 2).contiguous().float()


-__all__ = ["MiniMaxH3AudioVAE", "MiniMaxH3VideoVAE"]
+__all__ = [
+    "MiniMaxH3AudioVAE",
+    "MiniMaxH3ChunkCallbackPeerError",
+    "MiniMaxH3ChunkedDecodeUnsupportedError",
+    "MiniMaxH3VideoChunkCallback",
+    "MiniMaxH3VideoVAE",
+]
@@ -0,0 +1,368 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Paged KV cache for SenseNova-U1 autoregressive decode, and a CUDA graph over it.
+
+Autoregressive decode dominates a think request -- 80% of the wall clock, at
+about 20 ms per token for 42 layers -- and almost all of the GPU idle in that
+window is Python/ATen dispatch rather than any CUDA call. A CUDA graph retires
+those dispatches, but capture needs static shapes, and the default cache grows
+K/V with ``torch.cat`` on every step.
+
+Padding the cache to a bucket and masking the tail does give static shapes, but
+the mask is what costs: measured on an A800 at real decode shapes, a masked
+bucket attention runs 7.96-11.55 ms per step against 1.03 ms for an exact-length
+unmasked one, which is more than the graph saves.
+
+The way out is a paged cache. ``flash_attn_varlen_func`` takes the used length
+as a *tensor* (``seqused_k``) alongside a ``block_table``, so the buffers stay
+bucket-sized and capturable while the kernel reads only the valid prefix. One
+captured graph then serves every length in the bucket.
+
+Scope, and when to delete this. The cache is model-local on purpose:
+``DiffusionKVCacheManager`` reserves once per scheduler request, and
+``ARDiffusionModelRunner`` lives under ``vllm_omni/experimental``. SenseNova runs
+its whole autoregressive loop inside one pipeline ``forward``, so the decode
+steps never surface to the scheduler and it cannot admit, pool, evict, reuse a
+prefix for, or continuously batch them. What is here is therefore one set of
+buffers behind an identity block table, reused by whichever request fits them,
+not general paged-KV support: it holds while the pipeline serves one sequence
+per forward, and is released when sleep discards the memory it captured
+against. Once the decode loop is
+driven by a scheduler-visible runner, or the pipeline handles more than one
+sequence per forward, delete this file and use the manager instead.
+"""
+
+from __future__ import annotations
+
+import inspect
+
+import torch
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+logger = init_logger(__name__)
+
+BLOCK_SIZE = 16
+# Growth schedule. Each new bucket costs one capture, so keep the list short and
+# let the last entry cover the tail rather than doubling forever.
+BUCKETS = (512, 1024, 2048, 4096, 8192)
+# Past the last bucket the schedule grows by this much at a time. A request
+# decodes at most `max_think_tokens` (1024) or `max_tokens` (512) steps, so a
+# step this size costs at most one reallocation per request. Doubling the last
+# bucket instead would also cost at most one, but a 9,100-token request would
+# then reserve 766 MiB it never uses, against 118 MiB here.
+TAIL_STEP = 2048
+
+
+# Every argument to the kernel is passed by keyword, so the probe covers all of
+# them. An older wheel can export `flash_attn_varlen_func` without the paged
+# ones, or under different names, and a version string would not say which;
+# probe the signature and fall back to SDPA when any is missing.
+_REQUIRED_KWARGS = (
+    "max_seqlen_q",
+    "cu_seqlens_q",
+    "max_seqlen_k",
+    "seqused_k",
+    "block_table",
+    "softmax_scale",
+    "causal",
+)
+
+
+def _flash_varlen():
+    """Import lazily: the bundled kernel is CUDA-only and optional."""
+    from vllm.vllm_flash_attn import flash_attn_varlen_func
+
+    return flash_attn_varlen_func
+
+
+def _accepts_paged_kwargs(fn) -> bool:
+    try:
+        params = inspect.signature(fn).parameters
+    except (TypeError, ValueError):  # pragma: no cover - no introspectable signature
+        return False
+    return all(name in params for name in _REQUIRED_KWARGS)
+
+
+def paged_decode_supported(device: torch.device, head_dim: int) -> bool:
+    if device.type != "cuda":
+        return False
+    if head_dim % 8 or head_dim > 256:
+        return False
+    try:
+        fn = _flash_varlen()
+    except Exception as exc:  # pragma: no cover - depends on the wheel
+        logger.debug("Paged decode unavailable: %s", exc)
+        return False
+    if not _accepts_paged_kwargs(fn):
+        logger.debug("Paged decode unavailable: flash_attn_varlen_func lacks %s", _REQUIRED_KWARGS)
+        return False
+    return True
+
+
+def dynamic_lora_wrappers_present(module) -> bool:
+    """True once ``DiffusionLoRAManager`` has wrapped layers under ``module``.
+
+    The manager replaces linear layers with ``BaseLayerWithLoRA`` and then
+    binds, rescales or resets slot 0 per request; the wrappers stay in the tree
+    afterwards. Capture records the module tree and the branches that ran during
+    it, so a graph taken before an adapter was bound would replay without the
+    adapter's matmuls and one taken with it bound would survive its removal --
+    both silently, and neither visible in the shape, dtype and bucket that decide
+    reuse. Nothing captured may outlive a request once these exist.
+
+    The distilled LoRA is not affected: it is fused one-way into the weights and
+    leaves no wrapper behind.
+    """
+    try:
+        from vllm.lora.layers import BaseLayerWithLoRA
+    except ImportError:  # pragma: no cover - depends on the wheel
+        return False
+    return any(isinstance(m, BaseLayerWithLoRA) for m in module.modules())
+
+
+def _bucket_for(length: int) -> int:
+    for b in BUCKETS:
+        if length <= b:
+            return b
+    # Past the schedule, grow in TAIL_STEP-sized steps. Rounding to a whole
+    # number of blocks here would reallocate and re-capture every BLOCK_SIZE
+    # tokens: a think edit with two 2048x2048 inputs runs to about 9,100 tokens
+    # and paid 41 reallocations and 42 captures in one request.
+    over = length - BUCKETS[-1]
+    return BUCKETS[-1] + ((over + TAIL_STEP - 1) // TAIL_STEP) * TAIL_STEP
+
+
+class PagedDecodeCache:
+    """Per-layer paged K/V for single-token decode.
+
+    Layout follows what ``flash_attn_varlen_func`` accepts with a block table:
+    ``(num_blocks, BLOCK_SIZE, kv_heads, head_dim)``. ``seqused`` holds the
+    number of valid tokens and is the only thing that changes between decode
+    steps, which is what lets a captured graph be replayed unchanged.
+    """
+
+    def __init__(self, num_layers, kv_heads, head_dim, length, device, dtype):
+        self.kv_heads = kv_heads
+        self.head_dim = head_dim
+        self.device = device
+        self.dtype = dtype
+        self.bucket = _bucket_for(length)
+        nblocks = self.bucket // BLOCK_SIZE
+        shape = (nblocks, BLOCK_SIZE, kv_heads, head_dim)
+        self.k = [torch.zeros(shape, device=device, dtype=dtype) for _ in range(num_layers)]
+        self.v = [torch.zeros(shape, device=device, dtype=dtype) for _ in range(num_layers)]
+        self.block_table = torch.arange(nblocks, device=device, dtype=torch.int32).unsqueeze(0)
+        self.cu_seqlens_q = torch.tensor([0, 1], device=device, dtype=torch.int32)
+        self.seqused = torch.zeros(1, device=device, dtype=torch.int32)
+        # The write slot has to live in a tensor, not a Python int. Under graph
+        # capture a Python index is baked in at capture time, so every replay
+        # would overwrite the same slot -- correct eagerly, silently wrong once
+        # captured.
+        self.pos = torch.zeros(1, device=device, dtype=torch.int64)
+        self._length = 0
+        # Bumped whenever the buffers are reallocated, so a graph runner can
+        # tell that its capture no longer points at live memory.
+        self.generation = 0
+
+    @property
+    def length(self) -> int:
+        return self._length
+
+    def set_length(self, n: int) -> None:
+        self._length = n
+        self.seqused.fill_(n)
+        self.pos.fill_(n - 1)
+
+    def reusable_for(self, num_layers, kv_heads, head_dim, dtype, length) -> bool:
+        """Can this cache serve another request without reallocating?
+
+        Reusing the buffers is what lets captured graphs survive across
+        requests: a capture binds the addresses it recorded, so a fresh cache
+        forces a fresh capture.
+        """
+        return (
+            len(self.k) == num_layers
+            and self.kv_heads == kv_heads
+            and self.head_dim == head_dim
+            and self.dtype == dtype
+            and _bucket_for(length) <= self.bucket
+        )
+
+    def load_prefix(self, cache) -> bool:
+        """Copy a prefill cache into the paged buffers.
+
+        Only the first ``seqused`` rows are ever read, so the tail is left as
+        whatever the previous request wrote. Returns False for an empty cache.
+        """
+        layer0 = cache.layers[0]
+        if layer0.keys is None:
+            return False
+        prefix = layer0.keys.shape[2]
+        for i in range(len(self.k)):
+            keys, values = cache.layers[i].keys, cache.layers[i].values
+            # [B, H, S, D] -> [S, H, D], then into the paged blocks
+            self._write_prefix(i, keys[0].transpose(0, 1), values[0].transpose(0, 1))
+        self.set_length(prefix)
+        return True
+
+    @classmethod
+    def from_dynamic_cache(cls, cache, num_layers, device, dtype, min_length=0):
+        """Allocate buffers for a prefill cache and copy it in."""
+        layer0 = cache.layers[0]
+        if layer0.keys is None:
+            return None
+        _, kv_heads, prefix, head_dim = layer0.keys.shape
+        obj = cls(num_layers, kv_heads, head_dim, max(prefix + 1, min_length), device, dtype)
+        obj.load_prefix(cache)
+        return obj
+
+    def _write_prefix(self, layer_idx, keys_shd, values_shd):
+        n = keys_shd.shape[0]
+        flat_k = self.k[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_v = self.v[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_k[:n].copy_(keys_shd)
+        flat_v[:n].copy_(values_shd)
+
+    def grow(self, length: int) -> bool:
+        """Move to the next bucket, keeping the tokens already stored.
+
+        Everything is built into locals first and published in one step. A
+        resize that failed part way -- an allocation for a later layer, say --
+        would otherwise leave new K beside old V under the ``generation`` a
+        captured graph is keyed on, so the next request would keep selecting
+        that graph and replay it over storage that had just been freed.
+        """
+        new_bucket = _bucket_for(length)
+        if new_bucket == self.bucket:
+            return False
+        nblocks = new_bucket // BLOCK_SIZE
+        shape = (nblocks, BLOCK_SIZE, self.kv_heads, self.head_dim)
+        old_len = self._length
+        new_k, new_v = [], []
+        for old_k, old_v in zip(self.k, self.v):
+            k = torch.zeros(shape, device=self.device, dtype=self.dtype)
+            v = torch.zeros(shape, device=self.device, dtype=self.dtype)
+            k.view(-1, self.kv_heads, self.head_dim)[:old_len].copy_(
+                old_k.view(-1, self.kv_heads, self.head_dim)[:old_len]
+            )
+            v.view(-1, self.kv_heads, self.head_dim)[:old_len].copy_(
+                old_v.view(-1, self.kv_heads, self.head_dim)[:old_len]
+            )
+            new_k.append(k)
+            new_v.append(v)
+        block_table = torch.arange(nblocks, device=self.device, dtype=torch.int32).unsqueeze(0)
+
+        # Publish. New buffers mean any graph captured against the old ones is
+        # stale, so the generation moves in the same step that hands them over.
+        self.k, self.v, self.block_table, self.bucket = new_k, new_v, block_table, new_bucket
+        self.generation += 1
+        self.set_length(old_len)
+        logger.debug("Paged decode grew to bucket %d at length %d (generation %d)", new_bucket, length, self.generation)
+        return True
+
+    def attend(self, layer_idx, query_bhsd, key_bhsd, value_bhsd, softmax_scale):
+        """Append this step's K/V and attend over the whole valid prefix.
+
+        Writes at ``length - 1`` because the caller has already advanced the
+        length for this step. Both the slot and the attended length come from
+        device tensors, so a captured graph reads whatever they hold at replay
+        rather than whatever they held at capture.
+        """
+        flat_k = self.k[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_v = self.v[layer_idx].view(-1, self.kv_heads, self.head_dim)
+        flat_k.index_copy_(0, self.pos, key_bhsd[0, :, 0].unsqueeze(0))
+        flat_v.index_copy_(0, self.pos, value_bhsd[0, :, 0].unsqueeze(0))
+        heads, head_dim = query_bhsd.shape[1], query_bhsd.shape[3]
+        q = query_bhsd[0].transpose(0, 1).reshape(1, heads, head_dim)
+        out = _flash_varlen()(
+            q,
+            self.k[layer_idx],
+            self.v[layer_idx],
+            max_seqlen_q=1,
+            cu_seqlens_q=self.cu_seqlens_q,
+            max_seqlen_k=self.bucket,
+            seqused_k=self.seqused,
+            block_table=self.block_table,
+            softmax_scale=softmax_scale,
+            causal=False,
+        )
+        return out.reshape(1, 1, heads, head_dim)
+
+    def to_dynamic_cache(self, cache) -> None:
+        """Write the decoded K/V back so the generation stage sees one cache.
+
+        Only decode is paged; the DiT stage that follows reads the ordinary
+        cache, so the two are reconciled once at the hand-off rather than kept
+        in sync on every step.
+        """
+        n = self._length
+        for i, layer in enumerate(cache.layers):
+            keys = self.k[i].view(-1, self.kv_heads, self.head_dim)[:n]
+            values = self.v[i].view(-1, self.kv_heads, self.head_dim)[:n]
+            layer.keys = keys.transpose(0, 1).unsqueeze(0).contiguous()
+            layer.values = values.transpose(0, 1).unsqueeze(0).contiguous()
+
+
+class DecodeGraphRunner:
+    """One captured decode step per KV bucket, replayed for every token.
+
+    The whole point of the paged cache is that everything the step reads which
+    varies -- the token, its position, and how much of the cache is live -- sits
+    in device tensors. So a single capture serves every step in a bucket: fill
+    the tensors, replay, read the logits out of the static output.
+
+    A capture is invalidated when the cache reallocates (tracked by
+    ``PagedDecodeCache.generation``), which happens once per bucket boundary.
+    """
+
+    def __init__(self, language_model, cache, device):
+        self.lm = language_model
+        self.cache = cache
+        self.device = device
+        self.input_ids = torch.zeros(1, 1, dtype=torch.long, device=device)
+        self.indexes = torch.zeros(3, 1, dtype=torch.long, device=device)
+        self._graphs: dict[tuple[int, int], tuple[torch.cuda.CUDAGraph, torch.Tensor]] = {}
+        self.captures = 0
+
+    def _forward(self):
+        return self.lm(
+            input_ids=self.input_ids,
+            indexes=self.indexes,
+            past_key_values=None,
+            use_cache=False,
+            paged_cache=self.cache,
+        ).logits
+
+    def _capture(self, key):
+        # Warm up on a side stream first so cuBLAS workspaces and any lazy
+        # allocation happen outside the capture. The warm-up recomputes this
+        # same step, writing the same K/V to the same slot, so it is idempotent.
+        side = torch.cuda.Stream()
+        side.wait_stream(torch.cuda.current_stream())
+        with torch.cuda.stream(side):
+            for _ in range(3):
+                self._forward()
+        torch.cuda.current_stream().wait_stream(side)
+
+        graph = torch.cuda.CUDAGraph()
+        # A private pool would hand every runner its own arena and never give it
+        # back, and a runner is rebuilt whenever the cache reallocates or a LoRA
+        # wrapper appears: measured, that grew device memory by about 40 MiB per
+        # request and never levelled off. The platform pool is shared, which is
+        # what every other captured path in the tree uses.
+        with torch.cuda.graph(graph, pool=current_platform.get_global_graph_pool()):
+            logits = self._forward()
+        self.captures += 1
+        logger.debug("Captured decode graph for bucket=%d generation=%d", self.cache.bucket, self.cache.generation)
+        self._graphs[key] = (graph, logits)
+        return self._graphs[key]
+
+    def step(self, token, t_index):
+        """Run one decode step. Returns the static logits tensor."""
+        self.input_ids[0, 0] = token
+        self.indexes[0, 0] = t_index
+        key = (self.cache.bucket, self.cache.generation)
+        entry = self._graphs.get(key) or self._capture(key)
+        entry[0].replay()
+        return entry[1]
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """SenseNova-U1 Pipeline for vLLM-Omni.

 SenseNova-U1 is a unified Qwen3-based model that uses Mixture-of-Tokenizers
@@ -35,6 +35,13 @@ from vllm.logger import init_logger
 from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
 from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
 from vllm_omni.diffusion.distributed.utils import get_local_device
+from vllm_omni.diffusion.lora.loader import (
+    LoraLoaderMixin,
+    _apply_diffusers_lora_alpha_scaling,
+    _load_lora_state_dict,
+    _prepare_lora_delta,
+    _remap_state_dict_keys,
+)
 from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
 from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
 from vllm_omni.diffusion.profiler.diffusion_pipeline_profiler import DiffusionPipelineProfilerMixin
@@ -44,7 +51,14 @@ from vllm_omni.transformers_utils.configs.sensenova_u1 import (
    SenseNovaU1Config,
 )

+from .paged_decode import (
+    DecodeGraphRunner,
+    PagedDecodeCache,
+    dynamic_lora_wrappers_present,
+    paged_decode_supported,
+)
 from .sensenova_u1_transformer import (
+    SenseNovaU1CausalLMOutput,
    SenseNovaU1ForCausalLM,
    clear_flash_kv_cache,
    create_block_causal_mask,
@@ -446,7 +460,13 @@ class SenseNovaU1DenoisingAdapter(nn.Module):
        return self.language_model(*args, **kwargs)


-class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipelineProfilerMixin, CFGParallelMixin):
+class SenseNovaU1Pipeline(
+    nn.Module,
+    SupportsComponentDiscovery,
+    DiffusionPipelineProfilerMixin,
+    CFGParallelMixin,
+    LoraLoaderMixin,
+):
    """SenseNova-U1 text-to-image and image-to-image pipeline for vllm-omni.

    Builds the full model graph internally:
@@ -709,8 +729,71 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
        sigma = shift * sigma / (1 + (shift - 1) * sigma)
        return 1 - sigma

-    def _ar_step(self, next_token, t_idx, past_key_values):
+    def _decode_context(self, past_key_values):
+        """Paged cache plus a captured graph for the AR loop, or ``None``.
+
+        Decode is where a think request spends most of its time, and every step
+        issues the same work at the same shapes, so it is worth capturing. The
+        cache has to be paged for that: it keeps the buffers a fixed size while
+        the kernel reads only ``seqused_k`` of them, which is what lets one
+        capture serve every step until the next bucket.
+
+        Both live on the pipeline and are reused across requests unless a LoRA
+        adapter is being served dynamically. A capture binds the addresses it
+        recorded, so a cache built per request forces a capture per request:
+        measured, that left about 40 MiB of device memory behind on every
+        request and re-captured every time. Reuse rests on the same thing the
+        paged path already rests on -- one sequence in flight per pipeline
+        forward.
+        """
+        if os.environ.get("VLLM_OMNI_SENSENOVA_PAGED_DECODE", "1") != "1":
+            return None
+        head_dim = self.language_model.model.layers[0].self_attn.head_dim
+        if not paged_decode_supported(torch.device(self.device), head_dim):
+            return None
+        if past_key_values is None or not past_key_values.layers:
+            return None
+        layer0 = past_key_values.layers[0]
+        if layer0.keys is None:
+            return None
+        _, kv_heads, prefix, kv_head_dim = layer0.keys.shape
+        device = torch.device(self.device)
+        num_layers = len(self.language_model.model.layers)
+        # Reuse is off once the LoRA manager holds the decode path: the test
+        # below sees shape, dtype and bucket, none of which move when an adapter
+        # is bound, rescaled or reset between requests.
+        reuse = not dynamic_lora_wrappers_present(self.language_model)
+        decode = getattr(self, "_paged_decode", None) if reuse else None
+        if decode is None or not decode[0].reusable_for(
+            num_layers, kv_heads, kv_head_dim, layer0.keys.dtype, prefix + 1
+        ):
+            cache = PagedDecodeCache(num_layers, kv_heads, kv_head_dim, prefix + 1, device, layer0.keys.dtype)
+            decode = (cache, DecodeGraphRunner(self.language_model, cache, device))
+            self._paged_decode = decode if reuse else None
+        if not decode[0].load_prefix(past_key_values):
+            return None
+        return decode
+
+    def release_captured_graphs(self) -> None:
+        """Drop the reused paged cache and the graphs captured against it.
+
+        Sleep level 2 discards the memory a capture recorded, so anything held
+        across requests has to go with it. The next request rebuilds both.
+        """
+        self._paged_decode = None
+
+    def _ar_step(self, next_token, t_idx, past_key_values, decode=None):
        """One autoregressive token step. Constructs single-token `indexes` explicitly."""
+        if decode is not None:
+            cache, runner = decode
+            if cache.length + 1 > cache.bucket:
+                cache.grow(cache.length + 1)
+            cache.set_length(cache.length + 1)
+            logits = runner.step(int(next_token), t_idx + 1)
+            # The caller reassigns its cache handle from this field; the paged
+            # tokens are merged into it once, at the hand-off below.
+            return SenseNovaU1CausalLMOutput(logits=logits, past_key_values=past_key_values)
+
        indexes = torch.tensor([[t_idx + 1], [0], [0]], dtype=torch.long, device=self.device)
        outputs = self.language_model(
            input_ids=next_token.unsqueeze(0),
@@ -725,24 +808,28 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
        think_end_token_id = self.tokenizer.convert_tokens_to_ids("</think>")
        think_token_ids = []
        next_token = torch.argmax(prefix_outputs.logits[:, -1, :], dim=-1)
+        decode = self._decode_context(past_key_values)

        for _ in range(max_think_tokens):
            token_item = next_token.item()
            if token_item == eos_token_id:
                break
            if token_item == think_end_token_id:
-                outputs = self._ar_step(next_token, t_idx, past_key_values)
+                outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
                past_key_values = outputs.past_key_values
                t_idx += 1
                think_token_ids.append(token_item)
                break

            think_token_ids.append(token_item)
-            outputs = self._ar_step(next_token, t_idx, past_key_values)
+            outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
            past_key_values = outputs.past_key_values
            t_idx += 1
            next_token = torch.argmax(outputs.logits[:, -1, :], dim=-1)

+        if decode is not None:
+            decode[0].to_dynamic_cache(past_key_values)
+
        # Append "\n\n<img>" tokens to cache
        append_ids = self.tokenizer(
            "\n\n<img>",
@@ -797,6 +884,10 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
        """Autoregressive text decoding seeded from prefix logits."""
        eos_token_id = self.tokenizer.convert_tokens_to_ids("<|im_end|>")
        generated_ids: list[int] = []
+        # Same paged decode as the think loop. No write-back at the end: the
+        # caller returns the decoded string and never reads the cache again,
+        # unlike the think path, which hands it to the generation stage.
+        decode = self._decode_context(past_key_values)

        logits = prefix_logits[:, -1, :]
        for _ in range(max_tokens):
@@ -810,7 +901,7 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
            if token_id == eos_token_id:
                break
            generated_ids.append(token_id)
-            outputs = self._ar_step(next_token, t_idx, past_key_values)
+            outputs = self._ar_step(next_token, t_idx, past_key_values, decode)
            past_key_values = outputs.past_key_values
            logits = outputs.logits[:, -1, :]
            t_idx += 1
@@ -1174,8 +1265,37 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
            layer.values = layer.values.expand(batch_size, *layer.values.shape[1:])
        prepare_flash_kv_cache(kv, current_len=token_hw, batch_size=batch_size)

+    @staticmethod
+    def _is_warmup_request(req) -> bool:
+        is_dummy_run = getattr(req, "is_dummy_run", None)
+        if callable(is_dummy_run):
+            return bool(is_dummy_run())
+        return OmniDiffusionRequest.is_dummy_run_request_id(getattr(req, "request_id", None))
+
+    def _warm_ar_decode(self) -> None:
+        """Run one single-token decode at startup so its compiled region is built.
+
+        The engine's dummy request is a text-to-image run with think off. It
+        exercises the prefill shape but never the decode one, so whatever the
+        deploy config compiles for decode lands on the first real request
+        instead of on readiness. Warmup is best effort and must not fail startup.
+        """
+        try:
+            lm = self.language_model
+            device = torch.device(self.device)
+            ids = torch.zeros(1, 2, dtype=torch.long, device=device)
+            idx = torch.zeros(3, 2, dtype=torch.long, device=device)
+            prefill = lm(input_ids=ids, indexes=idx, use_cache=True)
+            nxt = torch.zeros(1, 1, dtype=torch.long, device=device)
+            nidx = torch.tensor([[2], [0], [0]], dtype=torch.long, device=device)
+            lm(input_ids=nxt, indexes=nidx, past_key_values=prefill.past_key_values, use_cache=True)
+        except Exception as exc:  # pragma: no cover - warmup is best effort
+            logger.warning("Autoregressive decode warmup skipped: %s", exc)
+
    @torch.inference_mode()
    def forward(self, req: DiffusionRequestBatch) -> DiffusionOutput:
+        if self._is_warmup_request(req):
+            self._warm_ar_decode()
        p = self._parse_request(req)

        input_images = self._extract_input_images(p.first_prompt)
@@ -1408,21 +1528,129 @@ class SenseNovaU1Pipeline(nn.Module, SupportsComponentDiscovery, DiffusionPipeli
    # Weight loading
    # -----------------------------------------------------------------------

+    # Shard name -> fused parameter, used by weight loading and LoRA loading.
+    # More specific _mot_gen patterns FIRST to avoid substring ambiguity
+    # (e.g. `.q_proj` is a substring of `.q_proj_mot_gen`).
+    stacked_params_mapping: ClassVar[list[tuple[str, str, str | int]]] = [
+        (".qkv_proj_mot_gen", ".q_proj_mot_gen", "q"),
+        (".qkv_proj_mot_gen", ".k_proj_mot_gen", "k"),
+        (".qkv_proj_mot_gen", ".v_proj_mot_gen", "v"),
+        (".qkv_proj", ".q_proj", "q"),
+        (".qkv_proj", ".k_proj", "k"),
+        (".qkv_proj", ".v_proj", "v"),
+        # MLP gate/up fused into MergedColumnParallelLinear
+        (".gate_up_proj", ".gate_proj", 0),
+        (".gate_up_proj", ".up_proj", 1),
+    ]
+
+    def load_lora_weights(
+        self,
+        pretrained_model_name_or_path: str | list[str],
+        adapter_name: str | None = None,
+    ) -> None:
+        """Fuse a distilled few-step LoRA into the weights.
+
+        The checkpoints ship kohya `lora_down`/`lora_up`/`alpha` names, renamed
+        here to the Diffusers names before `load_lora_into_module` routes each
+        delta into its slice of the fused projections.
+        """
+        if adapter_name in self.lora_loaded:
+            return
+
+        paths = (
+            [pretrained_model_name_or_path]
+            if isinstance(pretrained_model_name_or_path, str)
+            else list(pretrained_model_name_or_path)
+        )
+        if len(paths) != 1:
+            raise ValueError(f"Expected exactly one distilled LoRA file, got {len(paths)}: {paths}")
+
+        state_dict = _load_lora_state_dict(paths[0])
+        state_dict = _remap_state_dict_keys(
+            state_dict,
+            [(".lora_down.weight", ".lora_A.weight"), (".lora_up.weight", ".lora_B.weight")],
+        )
+        # Fold alpha and the low-rank product in fp32; scaling B in bf16 would
+        # round once before the matmul and once after.
+        state_dict = {k: v.to(torch.float32) for k, v in state_dict.items()}
+        state_dict = _apply_diffusers_lora_alpha_scaling(state_dict)
+
+        applied = self._add_lora_deltas(state_dict)
+        if not applied:
+            raise ValueError(f"Distilled LoRA {paths[0]} matched no parameter of this pipeline.")
+        logger.info("Fused distilled LoRA %s into %d parameters", paths[0], applied)
+        # Fusion is one-way and there is no unload path, so keep only a sentinel;
+        # the fp32 state dict is ~1.5 GiB and would stay resident for the process.
+        self.lora_loaded[adapter_name] = paths[0]
+
+    def _fused_shards(self, base_key: str) -> list[tuple[str, str | int]] | None:
+        """Unfused source names and shard ids for a fused parameter, else None."""
+        for param_name, weight_name, shard_id in self.stacked_params_mapping:
+            if base_key.endswith(param_name):
+                return [
+                    (base_key[: -len(param_name)] + w, sid)
+                    for p, w, sid in self.stacked_params_mapping
+                    if p == param_name
+                ]
+        return None
+
+    def _add_lora_deltas(self, state_dict: dict[str, torch.Tensor]) -> int:
+        """Add every matching delta into its parameter, sharding through vLLM.
+
+        The delta read from the checkpoint spans the whole layer. A parameter of
+        a parallel layer only holds this rank's slice of it, so the delta goes
+        through the layer's own weight loader -- into a zeroed copy of the
+        parameter, which is then added -- instead of being added directly.
+
+        Each parameter is fused in place as it is reached, so a failure part way
+        through leaves the earlier ones fused and the pipeline has to be
+        reloaded. Deferring the writes would mean holding every delta at once,
+        and the deltas are full rank: 15.09 GiB for the 8B checkpoint, against
+        the 192 MiB one parameter costs here. Fusion runs at load time, before
+        the pipeline serves anything, so a partially fused model is never
+        reachable from a request.
+        """
+        applied = 0
+        for name, param in self.named_parameters():
+            if not name.endswith(".weight"):
+                continue
+            base_key = name[: -len(".weight")]
+            shards = self._fused_shards(base_key)
+
+            deltas: list[tuple[torch.Tensor, str | int | None]] = []
+            for src, shard_id in shards or [(base_key, None)]:
+                delta, _ = _prepare_lora_delta(state_dict, src, None)
+                if delta is None:
+                    deltas = []
+                    break
+                deltas.append((delta, shard_id))
+            if not deltas:
+                continue
+
+            weight_loader = getattr(param, "weight_loader", None)
+            buffer = torch.zeros_like(param.data)
+            if weight_loader is None:
+                # Not a parallel layer: the delta is already rank-local.
+                buffer = torch.cat([d for d, _ in deltas]).to(device=param.device, dtype=param.dtype)
+            else:
+                saved = param.data
+                try:
+                    param.data = buffer
+                    for delta, shard_id in deltas:
+                        delta = delta.to(device=param.device, dtype=param.dtype)
+                        if shard_id is None:
+                            weight_loader(param, delta)
+                        else:
+                            weight_loader(param, delta, shard_id)
+                finally:
+                    param.data = saved
+            with torch.no_grad():
+                param.data.add_(buffer)
+            applied += 1
+        return applied
+
    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
-        stacked_params_mapping = [
-            # More specific _mot_gen patterns FIRST to avoid substring
-            # ambiguity (e.g. `.q_proj` is a substring of `.q_proj_mot_gen`).
-            (".qkv_proj_mot_gen", ".q_proj_mot_gen", "q"),
-            (".qkv_proj_mot_gen", ".k_proj_mot_gen", "k"),
-            (".qkv_proj_mot_gen", ".v_proj_mot_gen", "v"),
-            (".qkv_proj", ".q_proj", "q"),
-            (".qkv_proj", ".k_proj", "k"),
-            (".qkv_proj", ".v_proj", "v"),
-            # MLP gate/up fused into MergedColumnParallelLinear
-            (".gate_up_proj", ".gate_proj", 0),
-            (".gate_up_proj", ".up_proj", 1),
-        ]
-        self.stacked_params_mapping = stacked_params_mapping
+        stacked_params_mapping = self.stacked_params_mapping

        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """Qwen3 LLM with Mixture-of-Tokenizers (MoT) for SenseNova-U1.

 Ported from the sensenova_u1 package with vllm tensor-parallel support:
@@ -16,6 +16,7 @@ from dataclasses import dataclass

 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 from cache_dit import ForwardPattern
 from transformers.cache_utils import DynamicCache
 from vllm.logger import init_logger
@@ -163,6 +164,18 @@ class Qwen3RotaryEmbedding(nn.Module):
        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)


+def _build_3d_rope(config) -> tuple[Qwen3RotaryEmbedding, Qwen3RotaryEmbedding]:
+    """The t and h/w rotary modules: t uses half head_dim, h/w a quarter each."""
+    t_config = copy.deepcopy(config)
+    t_config.head_dim = config.head_dim // 2
+
+    hw_config = copy.deepcopy(config)
+    hw_config.head_dim = config.head_dim // 4
+    hw_config.rope_theta = config.rope_theta_hw
+    hw_config.max_position_embeddings = config.max_position_embeddings_hw
+    return Qwen3RotaryEmbedding(config=t_config), Qwen3RotaryEmbedding(config=hw_config)
+
+
 # ---------------------------------------------------------------------------
 # Norms
 # ---------------------------------------------------------------------------
@@ -175,11 +188,9 @@ class Qwen3RMSNorm(nn.Module):
        self.variance_epsilon = eps

    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
+        # F.rms_norm keeps the fp32 accumulation internal and rounds once; the
+        # cast chain it replaces rounded to bf16 before the weight multiply.
+        return F.rms_norm(hidden_states, self.weight.shape, self.weight, self.variance_epsilon)


 # ---------------------------------------------------------------------------
@@ -303,23 +314,16 @@ class SenseNovaU1Attention(nn.Module):
        self.q_norm_hw_mot_gen = Qwen3RMSNorm(self.head_dim // 2, eps=config.rms_norm_eps)
        self.k_norm_hw_mot_gen = Qwen3RMSNorm(self.head_dim // 2, eps=config.rms_norm_eps)

-        # 3D RoPE: t uses half head_dim, h/w each use quarter head_dim
-        t_config = copy.deepcopy(config)
-        t_config.head_dim = config.head_dim // 2
-        self.rotary_emb = Qwen3RotaryEmbedding(config=t_config)
-
-        hw_config = copy.deepcopy(config)
-        hw_config.head_dim = config.head_dim // 4
-        hw_config.rope_theta = config.rope_theta_hw
-        hw_config.max_position_embeddings = config.max_position_embeddings_hw
-        self.rotary_emb_hw = Qwen3RotaryEmbedding(config=hw_config)
-
        self.attn = Attention(
            num_heads=self.num_heads,
            head_size=self.head_dim,
            causal=False,
            softmax_scale=self.scaling,
-            num_kv_heads=self.num_heads,
+            # Keep K/V at their real head count. The SDPA backend already picks
+            # a fused GQA kernel when the runtime shape allows one and expands
+            # K/V itself when it does not; expanding here hides the GQA shape
+            # from that check, so it never fires.
+            num_kv_heads=self.num_kv_heads,
            prefix=f"{prefix}.attn",
        )
        self.attn.attention = self.attn.sdpa_fallback
@@ -339,10 +343,6 @@ class SenseNovaU1Attention(nn.Module):
        attention_mask: torch.Tensor | None,
    ) -> torch.Tensor:
        """Run unified attention with [B, H, S, D] inputs. Returns [B, S, H, D]."""
-        if self.num_kv_groups > 1:
-            n = self.num_kv_groups
-            key_bhsd = key_bhsd.repeat_interleave(n, dim=1)
-            value_bhsd = value_bhsd.repeat_interleave(n, dim=1)
        q = query_bhsd.transpose(1, 2).contiguous()
        k = key_bhsd.transpose(1, 2).contiguous()
        v = value_bhsd.transpose(1, 2).contiguous()
@@ -358,16 +358,16 @@ class SenseNovaU1Attention(nn.Module):
        attention_mask: torch.Tensor | None,
    ) -> torch.Tensor:
        """Run unified attention with [B, S, H, D] inputs. Returns [B, S, H, D]."""
-        if self.num_kv_groups > 1:
-            n = self.num_kv_groups
-            key_bshd = key_bshd.repeat_interleave(n, dim=2)
-            value_bshd = value_bshd.repeat_interleave(n, dim=2)
        attention_mask = self._align_mask_dtype(attention_mask, query_bshd)
        attn_metadata = AttentionMetadata(attn_mask=attention_mask) if attention_mask is not None else None
        return self.attn(query_bshd, key_bshd, value_bshd, attn_metadata)

-    def _project_and_rope(self, hidden_states, indexes, qkv_proj, q_norm, k_norm, q_norm_hw, k_norm_hw):
-        """Project Q/K/V via the given QKVParallelLinear and apply 3D RoPE."""
+    def _project_and_rope(self, hidden_states, position_embeddings, qkv_proj, q_norm, k_norm, q_norm_hw, k_norm_hw):
+        """Project Q/K/V via the given QKVParallelLinear and apply 3D RoPE.
+
+        The three tables come from the model, which builds them once per forward
+        because every layer is handed the same `indexes`.
+        """
        input_shape = hidden_states.shape[:-1]  # (B, S)
        qkv, _ = qkv_proj(hidden_states)
        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
@@ -376,9 +376,7 @@ class SenseNovaU1Attention(nn.Module):
        k = k.view(*input_shape, self.num_kv_heads, self.head_dim)
        v = v.view(*input_shape, self.num_kv_heads, self.head_dim)

-        cos_t, sin_t = self.rotary_emb(hidden_states, indexes[0].unsqueeze(0))
-        cos_h, sin_h = self.rotary_emb_hw(hidden_states, indexes[1].unsqueeze(0))
-        cos_w, sin_w = self.rotary_emb_hw(hidden_states, indexes[2].unsqueeze(0))
+        (cos_t, sin_t), (cos_h, sin_h), (cos_w, sin_w) = position_embeddings

        value_states = v.transpose(1, 2)  # [B, H, S, D]
        try:
@@ -433,13 +431,23 @@ class SenseNovaU1Attention(nn.Module):
        input_shape = hidden_states.shape[:-1]
        query_states, key_states, value_states = self._project_and_rope(
            hidden_states,
-            indexes,
+            kwargs["position_embeddings"],
            self.qkv_proj,
            self.q_norm,
            self.k_norm,
            self.q_norm_hw,
            self.k_norm_hw,
        )
+        # Single-token decode with a paged cache: append this step's K/V and let
+        # the kernel read the valid prefix via `seqused_k`. Shapes stay fixed,
+        # which is what makes the step capturable as a CUDA graph.
+        paged = kwargs.get("paged_cache")
+        if paged is not None and attention_mask is None and query_states.shape[2] == 1:
+            attn_output = paged.attend(self.layer_idx, query_states, key_states, value_states, self.scaling)
+            attn_output = attn_output.reshape(*input_shape, -1).contiguous()
+            attn_output, _ = self.o_proj(attn_output)
+            return attn_output
+
        update_cache = kwargs.get("update_cache", True)
        if past_key_values is not None:
            if update_cache:
@@ -460,7 +468,7 @@ class SenseNovaU1Attention(nn.Module):
        input_shape = hidden_states.shape[:-1]
        query_states, key_states, value_states = self._project_and_rope(
            hidden_states,
-            indexes,
+            kwargs["position_embeddings"],
            self.qkv_proj_mot_gen,
            self.q_norm_mot_gen,
            self.k_norm_mot_gen,
@@ -659,6 +667,10 @@ class SenseNovaU1Model(nn.Module):
        )
        self.norm = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.norm_mot_gen = Qwen3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        # Every layer receives the same `indexes`, so the tables are the same for
+        # all of them: build them once per forward here and thread them down,
+        # rather than giving every layer its own pair.
+        self.rotary_emb, self.rotary_emb_hw = _build_3d_rope(config)

    def forward(
        self,
@@ -690,15 +702,25 @@ class SenseNovaU1Model(nn.Module):
            past_len = past_key_values.get_seq_length() if past_key_values else 0
            seq_len = inputs_embeds.shape[1]
            total_len = past_len + seq_len
-            mask = torch.zeros(1, 1, seq_len, total_len, device=inputs_embeds.device)
            if seq_len > 1:
+                mask = torch.zeros(1, 1, seq_len, total_len, device=inputs_embeds.device)
                causal = torch.tril(torch.ones(seq_len, seq_len, device=inputs_embeds.device))
                mask[:, :, :, past_len:] = torch.where(causal == 1, 0.0, float("-inf"))
+            else:
+                # A single query token attends to every cached key, so the mask
+                # this branch used to build was all zeros. Passing it changed no
+                # result, only which SDPA kernel was reachable.
+                mask = None
            causal_mask_mapping = {"full_attention": mask}
        else:
            causal_mask_mapping = attention_mask

        hidden_states = inputs_embeds
+        position_embeddings = (
+            self.rotary_emb(hidden_states, indexes[0].unsqueeze(0)),
+            self.rotary_emb_hw(hidden_states, indexes[1].unsqueeze(0)),
+            self.rotary_emb_hw(hidden_states, indexes[2].unsqueeze(0)),
+        )
        for layer in self.layers:
            hidden_states = layer(
                hidden_states,
@@ -706,6 +728,7 @@ class SenseNovaU1Model(nn.Module):
                exist_und=exist_und,
                exist_gen=exist_gen,
                indexes=indexes,
+                position_embeddings=position_embeddings,
                attention_mask=causal_mask_mapping,
                past_key_values=past_key_values,
                use_cache=use_cache,
@@ -49,14 +49,13 @@ def create_diffusion_client(
    od_config: OmniDiffusionConfig,
    metadata: StageMetadata,
    stage_init_timeout: int,
-    batch_size: int = 1,
    use_inline: bool = False,
 ) -> Any:
    """Factory to create either an inline or out-of-process diffusion client."""
    if use_inline:
        from vllm_omni.diffusion.inline_stage_diffusion_client import InlineStageDiffusionClient

-        return InlineStageDiffusionClient(model, od_config, metadata, batch_size=batch_size)
+        return InlineStageDiffusionClient(model, od_config, metadata)
    proc_manager = StageDiffusionProcManager(
        model=model,
        od_config=od_config,
@@ -67,7 +66,6 @@ def create_diffusion_client(
        request_address=proc_manager.addresses.inputs[0],
        response_address=proc_manager.addresses.outputs[0],
        proc_manager=proc_manager,
-        batch_size=batch_size,
    )


@@ -91,14 +89,12 @@ class StageDiffusionClient(StageClientBase):
        response_address: str,
        *,
        proc_manager: StageDiffusionProcManager | None = None,
-        batch_size: int = 1,
    ) -> None:
        self._initialize_client(
            metadata,
            request_address,
            response_address,
            proc_manager=proc_manager,
-            batch_size=batch_size,
        )

    @classmethod
@@ -109,7 +105,6 @@ class StageDiffusionClient(StageClientBase):
        response_address: str,
        *,
        proc_manager: StageDiffusionProcManager | None = None,
-        batch_size: int = 1,
    ) -> StageDiffusionClient:
        """Create a client for an already-running diffusion subprocess."""
        return cls(
@@ -117,7 +112,6 @@ class StageDiffusionClient(StageClientBase):
            request_address,
            response_address,
            proc_manager=proc_manager,
-            batch_size=batch_size,
        )

    def _initialize_client(
@@ -127,7 +121,6 @@ class StageDiffusionClient(StageClientBase):
        response_address: str,
        *,
        proc_manager: StageDiffusionProcManager | None = None,
-        batch_size: int,
    ) -> None:
        self._set_stage_metadata(metadata)
        self._proc_manager = proc_manager
@@ -144,11 +137,10 @@ class StageDiffusionClient(StageClientBase):
            self._start_proc_monitor()

        logger.info(
-            "[StageDiffusionClient] stage-%s [rep-%s] initialized (owns_process=%s, batch_size=%d)",
+            "[StageDiffusionClient] stage-%s [rep-%s] initialized (owns_process=%s)",
            self.stage_id,
            self.replica_id,
            self._proc_manager is not None,
-            batch_size,
        )

    def _set_stage_metadata(self, metadata: StageMetadata) -> None:
@@ -519,6 +519,23 @@ class DiffusionModelRunner(OmniConnectorModelRunnerMixin):
        if self.prompt_embed_cache is not None:
            self.prompt_embed_cache.clear()

+    def release_captured_graphs(self) -> None:
+        """Drop every CUDA graph held for this model, wherever it is kept.
+
+        Sleep level 2 discards the memory a capture was recorded against, so a
+        graph that outlives it replays over freed storage. The runner owns
+        compilation and execution resources, so it owns the release: a pipeline
+        that captures graphs of its own implements ``release_captured_graphs``
+        and is collected here, instead of every caller having to know which
+        pipelines have one.
+        """
+        runners = getattr(self, "graph_runners", None)
+        if runners is not None:
+            runners.clear()
+        release = getattr(self.pipeline, "release_captured_graphs", None)
+        if callable(release):
+            release()
+
    def get_prompt_embed_cache_stats(self) -> dict | None:
        """Return hit/miss statistics for the prompt-embedding cache, if enabled.

@@ -786,9 +786,8 @@ class DiffusionWorker:
        usage_before = allocator.get_current_usage()

        if level == 2 and self.model_runner is not None:
-            if hasattr(self.model_runner, "graph_runners"):
-                self.model_runner.graph_runners.clear()
-                logger.info(f"[Worker {self.rank}] CUDA Graphs cleared.")
+            self.model_runner.release_captured_graphs()
+            logger.info(f"[Worker {self.rank}] CUDA Graphs cleared.")
            model = self.model_runner.pipeline
            self._sleep_saved_buffers = {name: buffer.cpu().clone() for name, buffer in model.named_buffers()}

@@ -129,7 +129,6 @@ class AsyncOmniEngine:
        model: str,
        stage_init_timeout: int = 300,
        init_timeout: int = 600,
-        diffusion_batch_size: int = 1,
        single_stage_mode: bool = False,
        transfer_emitter: Any = None,
        prom_metrics: Any = None,
@@ -140,7 +139,6 @@ class AsyncOmniEngine:
    ) -> None:
        self.model = model
        self.tokenizer = tokenizer
-        self.diffusion_batch_size = diffusion_batch_size
        # Cached by get_diffusion_od_config().
        self._diffusion_od_config_view: Any = None
        startup_timeout = int(init_timeout)
@@ -320,7 +318,6 @@ class AsyncOmniEngine:
            config_path=self.config_path,
            single_stage_mode=self.single_stage_mode,
            stage_init_timeout=stage_init_timeout,
-            diffusion_batch_size=self.diffusion_batch_size,
            async_chunk=self.async_chunk,
            tokenizer=self.tokenizer,
            single_stage_id_filter=self._single_stage_id_filter,
@@ -1352,8 +1349,8 @@ class AsyncOmniEngine:
                    or kwargs.get("fastvideo_vsa_topk") is not None
                ):
                    has_stage_attention = (
-                        hasattr(cfg.engine_args, "diffusion_attention_config")
-                        and cfg.engine_args.diffusion_attention_config is not None
+                        getattr(cfg.engine_args, "diffusion_attention_config", None) is not None
+                        or getattr(cfg.engine_args, "diffusion_attention_backend", None) is not None
                    )
                    if not has_stage_attention:
                        cfg.engine_args.diffusion_attention_config = parse_attention_config(
@@ -1546,7 +1546,6 @@ def launch_diffusion_stage_replica(
    stage_config: Any,
    metadata: Any,
    stage_init_timeout: int,
-    batch_size: int,
    use_inline: bool,
    replica_id: int = 0,
    omni_master_server: OmniMasterServer | None = None,
@@ -1565,7 +1564,6 @@ def launch_diffusion_stage_replica(
            stage_config,
            metadata,
            stage_init_timeout=stage_init_timeout,
-            batch_size=batch_size,
            use_inline=use_inline,
        )
        return client, StageReplicaResources()
@@ -1624,7 +1622,6 @@ def launch_diffusion_stage_replica(
            request_address=proc_manager.addresses.inputs[0],
            response_address=proc_manager.addresses.outputs[0],
            proc_manager=proc_manager,
-            batch_size=batch_size,
        )
        return client, StageReplicaResources(
            manager=proc_manager,
@@ -1253,9 +1253,19 @@ def _finalize_engine_args_dict(
    if is_diffusion:
        from vllm_omni.diffusion.data import parse_attention_config

-        if engine_args_dict.get("diffusion_attention_config") is not None:
+        # Fold the attention shorthand into the structured config so only one
+        # representation reaches OmniDiffusionConfig.from_kwargs.
+        attention_backend = engine_args_dict.pop("diffusion_attention_backend", None)
+        fastvideo_vsa_topk = engine_args_dict.pop("fastvideo_vsa_topk", None)
+        if (
+            engine_args_dict.get("diffusion_attention_config") is not None
+            or attention_backend is not None
+            or fastvideo_vsa_topk is not None
+        ):
            engine_args_dict["diffusion_attention_config"] = parse_attention_config(
-                engine_args_dict["diffusion_attention_config"],
+                engine_args_dict.get("diffusion_attention_config"),
+                attention_backend=attention_backend,
+                fastvideo_vsa_topk=fastvideo_vsa_topk,
            )
    else:
        resolve_worker_cls(engine_args_dict)
@@ -1813,7 +1823,6 @@ def initialize_diffusion_stage(
    stage_cfg: Any,
    metadata: StageMetadata,
    stage_init_timeout: int,
-    batch_size: int = 1,
    use_inline: bool = False,
 ) -> Any:
    """Build a diffusion stage client.
@@ -1823,15 +1832,12 @@ def initialize_diffusion_stage(
        stage_cfg: Stage configuration.
        metadata: Extracted stage metadata.
        stage_init_timeout: Timeout in seconds for stage initialization handshake
-        batch_size: Client-side request batch width. Does not set scheduler
-            ``max_num_seqs``; pass ``--max-num-seqs`` or stage YAML for that.
-            Forwarded to ``StageDiffusionClient``.
        use_inline: If True, uses the inline diffusion client instead of subprocess.
    """
    from vllm_omni.diffusion.stage_diffusion_client import create_diffusion_client

    od_config = build_diffusion_config(model, stage_cfg, metadata)
-    return create_diffusion_client(model, od_config, metadata, stage_init_timeout, batch_size, use_inline)
+    return create_diffusion_client(model, od_config, metadata, stage_init_timeout, use_inline)


 def _stage_declares_cfg_pairs(model_config: Any) -> bool:
@@ -87,7 +87,6 @@ class StageRemoteFactoryContext:
    base_metadata: Any
    vllm_config: Any | None = None
    executor_class: type | None = None
-    diffusion_batch_size: int = 1


 def _build_load_balancer_factory(policy: str) -> Callable[[], LoadBalancer]:
@@ -124,7 +123,6 @@ class StageRuntime:
        config_path: str,
        *,
        stage_init_timeout: int,
-        diffusion_batch_size: int,
        async_chunk: bool,
        tokenizer: str | None = None,
        log_stats: bool = False,
@@ -133,7 +131,6 @@ class StageRuntime:
        self._model = model
        self._config_path = config_path
        self._stage_init_timeout = stage_init_timeout
-        self._diffusion_batch_size = diffusion_batch_size
        self._async_chunk = async_chunk
        self._tokenizer = tokenizer
        self._log_stats = log_stats
@@ -672,7 +669,6 @@ class StageRuntime:
                    stage_config=plan.stage_cfg,
                    metadata=plan.metadata,
                    stage_init_timeout=stage_init_timeout,
-                    batch_size=self._diffusion_batch_size,
                    use_inline=plan.num_replicas == 1
                    and bool(self._num_stages == 1 or inline_diffusion or custom_pipeline_args),
                    replica_id=plan.replica_id,
@@ -681,10 +677,9 @@ class StageRuntime:
                )

            logger.info(
-                "[StageRuntime] Stage %s replica %s initialized (diffusion, batch_size=%d)",
+                "[StageRuntime] Stage %s replica %s initialized (diffusion)",
                plan.metadata.stage_id,
                plan.replica_id,
-                self._diffusion_batch_size,
            )
            return cast(StagePoolClient, client)
        except Exception:
@@ -768,7 +763,6 @@ class DistStageRuntime(StageRuntime):
        config_path: str,
        *,
        stage_init_timeout: int,
-        diffusion_batch_size: int,
        async_chunk: bool,
        single_stage_id_filter: int | None,
        omni_master_address: str,
@@ -785,7 +779,6 @@ class DistStageRuntime(StageRuntime):
            model=model,
            config_path=config_path,
            stage_init_timeout=stage_init_timeout,
-            diffusion_batch_size=diffusion_batch_size,
            async_chunk=async_chunk,
            tokenizer=tokenizer,
            log_stats=log_stats,
@@ -914,7 +907,6 @@ class DistStageRuntime(StageRuntime):
            base_metadata=metadata,
            vllm_config=plan.stage_vllm_config,
            executor_class=plan.executor_class,
-            diffusion_batch_size=self._diffusion_batch_size,
        )
        return self._create_remote_replica_client(ctx, plan.replica_id)

@@ -977,7 +969,6 @@ class DistStageRuntime(StageRuntime):
                base_metadata=template.metadata,
                vllm_config=template.stage_vllm_config,
                executor_class=template.executor_class,
-                diffusion_batch_size=self._diffusion_batch_size,
            )
        return contexts

@@ -1045,7 +1036,6 @@ class DistStageRuntime(StageRuntime):
                metadata,
                request_address=resources.addresses.inputs[0],
                response_address=resources.addresses.outputs[0],
-                batch_size=ctx.diffusion_batch_size,
            )
            logger.info(
                "[DistStageRuntime] Remote diffusion replica attached stage=%d replica=%d",
@@ -1109,7 +1099,6 @@ def create_stage_runtime(
    *,
    single_stage_mode: bool,
    stage_init_timeout: int,
-    diffusion_batch_size: int,
    async_chunk: bool,
    tokenizer: str | None = None,
    # Distributed-only params:
@@ -1131,7 +1120,6 @@ def create_stage_runtime(
            model=model,
            config_path=config_path,
            stage_init_timeout=stage_init_timeout,
-            diffusion_batch_size=diffusion_batch_size,
            async_chunk=async_chunk,
            tokenizer=tokenizer,
            log_stats=log_stats,
@@ -1148,7 +1136,6 @@ def create_stage_runtime(
        model=model,
        config_path=config_path,
        stage_init_timeout=stage_init_timeout,
-        diffusion_batch_size=diffusion_batch_size,
        async_chunk=async_chunk,
        tokenizer=tokenizer,
        log_stats=log_stats,
@@ -187,7 +187,9 @@ class OmniBase(PDDisaggregationMixin):
        # override the deploy YAML's ``async_chunk: true`` default.
        async_chunk = kwargs.get("async_chunk")
        output_modalities = kwargs.pop("output_modalities", None)
-        diffusion_batch_size: int = kwargs.pop("diffusion_batch_size", 1)
+
+        if "diffusion_batch_size" in kwargs:
+            raise TypeError("`diffusion_batch_size` has been removed in Omni/AsyncOmni. Use `max_num_seqs`.")

        if "log_requests" in kwargs:
            raise TypeError("`log_requests` has been removed in Omni/AsyncOmni. Use `log_stats`.")
@@ -213,7 +215,6 @@ class OmniBase(PDDisaggregationMixin):
            model=model,
            init_timeout=init_timeout,
            stage_init_timeout=stage_init_timeout,
-            diffusion_batch_size=diffusion_batch_size,
            transfer_emitter=self.transfer_metrics,
            prom_metrics=self.prom_metrics,
            log_stats=log_stats,
Author	SHA1	Message	Date
hsliuustc0106	81de3f3308	[Bugfix][MiniMax-H3] Apply pre-commit formatting Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-03 08:44:26 +08:00
hsliuustc0106	f9c9de8d16	[Bugfix][MiniMax-H3] Reconcile temporal decode plans across VAE ranks Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com> (cherry picked from commit ac4ec9c0fcf3394e8476d3d2d913d7c86f5df8e3)	2026-09-03 08:20:00 +08:00
hsliuustc0106	ff2e33a329	[Docs][MiniMax-H3] Document temporal VAE chunks Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-02 07:08:40 +08:00
hsliuustc0106	0b5b0979d8	[Bugfix][MiniMax-H3] Reconcile chunk callback ownership Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-01 23:11:29 +08:00
hsliuustc0106	f04ff76c54	[Model][MiniMax-H3] Isolate temporal chunk coordination Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-01 23:11:29 +08:00
hsliuustc0106	b220df8397	[Model][MiniMax-H3] Support released VAE chunks directly Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-01 23:11:28 +08:00
hsliuustc0106	520e058b95	[Model][MiniMax-H3] Delegate temporal VAE chunks Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-01 23:11:27 +08:00
DanaerLeeandHongsheng Liu	0288c3f56e	[Model] Support SenseNova-U1.5-8B-MoT and its distilled 8-step LoRA (#6516 ) Signed-off-by: MrlixiangWE <mrdanaer@gmail.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-09-01 22:53:53 +08:00
andyluo7andHongsheng Liu	477083a103	[CI][ROCm] Fix deterministic AMD L3 blockers (#6884 ) Signed-off-by: andyluo7 <andy.luo@amd.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-09-01 22:53:00 +08:00
Rahul SteigerandRahul Steiger	e1b2f589c1	[Bugfix][Cosmos3] Avoid logging prompts at INFO level (#6913 ) Signed-off-by: Rahul Steiger <rsteiger@nvl72d233-T16.cm.cluster> Co-authored-by: Rahul Steiger <rsteiger@nvl72d233-T16.cm.cluster>	2026-09-01 22:52:15 +08:00
Hongsheng Liuandhsliu_ustc	44c8d1c449	docs: add concise API server endpoint guide (#6114 ) Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com> Co-authored-by: hsliu_ustc <hsliu_ustc@noreply.gitcode.com>	2026-09-01 20:34:09 +08:00
Andy Zhou	2298359f67	[Bugfix][Diffusion] Preserve stage attention backend (#6645 ) Signed-off-by: AndyZhou952 <jzhoubc@connect.ust.hk>	2026-09-01 20:18:44 +08:00
Hongsheng Liu	d1c8beeb18	[Doc] Update README and docs for v0.28.0 (#6858 ) Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>	2026-09-01 19:50:26 +08:00
WeiQing Chen	e2b83db683	[Test] Add MiniMax-H3 DLO DP2 variants and Turbo LoRA L3 test (#6556 ) Signed-off-by: david6666666 <530634352@qq.com>	2026-09-01 10:48:52 +00:00
AuFlowandHongsheng Liu	49c6fc1540	[Frontend] Add configurable output sample rate to Speech API (#6553 ) Signed-off-by: 郑倩倩 <2458983186@qq.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-09-01 17:25:09 +08:00
Alex Brooks	a65e89cbd4	[Core] Validate PipelineConfig in `__post_init__` (#6829 ) Signed-off-by: Alex Brooks <albrooks@redhat.com>	2026-09-01 11:17:57 +08:00
Yuxuan RhoandHongsheng Liu	7992c2213e	[Config] Remove diffusion_batch_size and use max_num_seqs instead. (#6484 ) Signed-off-by: LyxWxj <yuxuanlou621@foxmail.com> Signed-off-by: Yuxuan Rho <104440472+LyxWxj@users.noreply.github.com> Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>	2026-09-01 01:02:54 +00:00
XRay2016	8e26dce1f4	[Diffusion] Add Boogu Edit-Turbo support docs and tests (#6701 ) Signed-off-by: xRay2016 <1150722393@qq.com>	2026-09-01 08:41:23 +08:00