Update vllm_omni/diffusion/models/flux2_klein/flux2_klein_transformer.py

Signed-off-by: Hongsheng Liu <liuhongsheng4@huawei.com>

.buildkite/cuda/test-merge.yml

@@ -169,7 +169,6 @@ steps:
           - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
           - vllm_omni/deploy/minicpmo_4_5.yaml
           - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
-          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
           - vllm_omni/experimental/fullduplex/
           - vllm_omni/model_executor/models/minicpmo_4_5/
           - vllm_omni/model_executor/models/step_audio2/

.buildkite/cuda/test-nightly.yml

@@ -83,7 +83,7 @@ steps:
           - export BENCHMARK_DIR=tests/dfx/perf/results
           - |
             set +e
-            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json -m "H100 and full_model"
             EXIT=$$?
             buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
             exit $$EXIT
@@ -375,7 +375,7 @@ steps:
       - label: ":full_moon: Diffusion · Quantization Test with L4"
         timeout_in_minutes: 60
         commands:
-          - pip install "vllm-gguf-plugin>=0.0.3"
+          - pip install "vllm-gguf-plugin==0.0.4"
           - pytest -sv tests/diffusion/quantization -m 'full_model and cuda and L4' --run-level "full_model"
         mirror_hardwares: l4_1
 

.buildkite/cuda/test-ready.yml

@@ -125,32 +125,8 @@ steps:
           - vllm_omni/outputs/
         commands:
           - timeout 20m pytest -s -v tests/e2e/features/custom_pipeline/ -m "core_model"
-        agents:
-          queue: "gpu_1_queue"
-        plugins:
-          - docker#v5.2.0:
-              image: public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-              always-pull: true
-              propagate-environment: true
-              # vLLM 0.27 added a pre-flight capacity check in
-              # distributed/device_communicators/shm_broadcast.py. Without this
-              # the container gets Docker's 64 MiB default and every test in
-              # this step dies at engine start with
-              #   RuntimeError: Insufficient space in /dev/shm:
-              #                 240 MiB required, 64 MiB available
-              # (9 failed in 8.7s on build #2949). vLLM 0.26 had no such check
-              # and allocated regardless, so this step passed on main while
-              # being under-provisioned the whole time — 0.27 turned a silent
-              # risk into an explicit error.
-              #
-              # This is the only inline docker plugin block in .buildkite/cuda/;
-              # every other step inherits common/ci_mirror_hardwares.yml, which
-              # already sets 8gb. Matching that rather than inventing a value.
-              shm-size: "8gb"
-              environment:
-                - "HF_HOME=/fsx/hf_cache"
-              volumes:
-                - "/fsx/hf_cache:/fsx/hf_cache"
+        mirror_hardwares: l4_1
+
 
       - label: "RLHF · VeRL-Omni E2E Test"
         source_file_dependencies:
@@ -232,7 +208,6 @@ steps:
           - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
           - vllm_omni/deploy/minicpmo_4_5.yaml
           - tests/e2e/online_serving/helpers/minicpmo_4_5_duplex.py
-          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
           - vllm_omni/experimental/fullduplex/
           - vllm_omni/model_executor/models/minicpmo_4_5/
           - vllm_omni/model_executor/models/step_audio2/

.buildkite/cuda/test-weekly.yml

@@ -74,6 +74,19 @@ steps:
             exit $$EXIT
         mirror_hardwares: h100_2
 
+      - label: ":full_moon: Omni · Perf Test · Async Chunk · Random"
+        key: weekly-omni-performance-async-chunk-random
+        timeout_in_minutes: 180
+        commands:
+          - export BENCHMARK_DIR=tests/dfx/perf/results
+          - |
+            set +e
+            pytest -s -v tests/dfx/perf/scripts/run_benchmark.py --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json -m "H100 and slow"
+            EXIT=$$?
+            buildkite-agent artifact upload "tests/dfx/perf/results/*.json"
+            exit $$EXIT
+        mirror_hardwares: h100_2
+
       - label: ":full_moon: Omni · Perf Test · Multi-Replica"
         key: weekly-omni-performance-multi-replicas
         timeout_in_minutes: 180

.claude/skills/add-diffusion-model/SKILL.md

@@ -105,7 +105,7 @@ For custom loop structures, follow `vllm_omni/diffusion/models/progress_bar.py`
 
 ## Path B: Custom/Private Repo Model
 
-For models without a diffusers pipeline — weights in custom format, model code in a private repo. Real examples: DreamID-Omni, BAGEL, HunyuanImage3. See [references/custom-model-patterns.md](references/custom-model-patterns.md) for concrete custom-model integration patterns.
+For models without a diffusers pipeline — weights in custom format, model code in a private repo. Real examples: BAGEL, HunyuanImage3. See [references/custom-model-patterns.md](references/custom-model-patterns.md) for concrete custom-model integration patterns.
 
 ### B1. Understand the reference repo
 
@@ -126,9 +126,9 @@ This is the key design decision for custom models. Follow these placement rules:
 | **Pipeline orchestration** (init, forward, denoise loop) | `vllm_omni/diffusion/models/<name>/pipeline_<name>.py` | Always required |
 | **Custom transformer/backbone** (ported and adapted to vllm-omni) | `vllm_omni/diffusion/models/<name>/<name>_transformer.py` or similar | `wan2_2.py`, `fusion.py`, `bagel_transformer.py` |
 | **Custom sub-models** (VAE, fusion, autoencoder) | `vllm_omni/diffusion/models/<name>/` as separate files | `autoencoder.py`, `fusion.py` |
-| **External dependency code** (original repo utilities) | **External repo**, installed via download script or pip | `dreamid_omni` package via git clone |
+| **External dependency code** (original repo utilities) | **External repo**, installed via download script or pip | original-repo package via git clone |
 | **Hardcoded model configs** | Module-level dicts in pipeline file | `VIDEO_CONFIG`, `AUDIO_CONFIG` dicts |
-| **Download/setup script** | `examples/offline_inference/<name>/download_<name>.py` | `download_dreamid_omni.py` |
+| **Download/setup script** | `examples/offline_inference/<name>/download_<name>.py` | `download_<name>.py` |
 | **Custom `model_index.json`** | Generated by download script, placed at model root | Minimal: `{"_class_name": "YourPipeline", ...}` |
 
 ### B3. Handle external dependencies
@@ -157,7 +157,7 @@ Copy the essential model files into `vllm_omni/diffusion/models/<name>/` and ada
 
 Custom models have two common patterns for weight loading:
 
-**Pattern 1: Bypass standard loader** (DreamID-Omni style)
+**Pattern 1: Bypass standard loader** (eager custom init)
 
 When the original model has complex custom init functions that load weights in `__init__`:
 

.claude/skills/add-diffusion-model/references/custom-model-patterns.md

@@ -16,20 +16,19 @@ vllm_omni/diffusion/models/wan2_2/
 
 The transformer is loaded separately via `weights_sources` + `load_weights()`. Non-transformer components (VAE, text encoder) are loaded in `__init__` via `from_pretrained()`.
 
-### Custom model with external deps (e.g., DreamID-Omni)
+### Custom model with external deps
 
 ```
-vllm_omni/diffusion/models/dreamid_omni/
+vllm_omni/diffusion/models/<name>/
 ├── __init__.py                    # Exports pipeline only
-├── pipeline_dreamid_omni.py       # Pipeline: loads ALL weights in __init__ via custom helpers
-├── fusion.py                      # Custom fusion architecture (video + audio cross-attention)
-└── wan2_2.py                      # Re-implemented Wan backbone with split API
+├── pipeline_<name>.py             # Pipeline: loads ALL weights in __init__ via custom helpers
+└── <backbone>.py                  # Custom transformer / fusion architecture
 
-examples/offline_inference/x_to_video_audio/
-└── download_dreamid_omni.py       # Downloads weights from 3 HF repos + clones code repo
+examples/offline_inference/<task>/
+└── download_<name>.py             # Downloads weights and/or clones code repo
 ```
 
-All weights loaded eagerly in `__init__`. `load_weights()` is a no-op. External dependency (`dreamid_omni` package) imported with try/except.
+All weights loaded eagerly in `__init__`. `load_weights()` is a no-op. External dependency imported with try/except.
 
 ### Custom model with ported code (e.g., BAGEL)
 
@@ -76,7 +75,7 @@ self.weights_sources = [
 ]
 ```
 
-### Pattern 3: Fully custom loading (DreamID-Omni)
+### Pattern 3: Fully custom loading
 
 ```
 init → load ALL weights eagerly via custom helpers → load_weights() = no-op
@@ -108,8 +107,8 @@ Standard diffusers `model_index.json`:
 Custom model `model_index.json` (minimal):
 ```json
 {
-    "_class_name": "DreamIDOmniPipeline",
-    "fusion": "DreamID-Omni/dreamid_omni.safetensors"
+    "_class_name": "YourCustomPipeline",
+    "fusion": "weights/model.safetensors"
 }
 ```
 
@@ -117,7 +116,7 @@ The only **required** field is `_class_name` — it must match a key in `_DIFFUS
 
 ## External Dependency Management
 
-### Git clone + .pth injection (DreamID-Omni pattern)
+### Git clone + .pth injection
 
 ```python
 def download_dependency():
@@ -163,7 +162,7 @@ The engine checks `isinstance(pipeline, SupportImageInput)` at startup to config
 
 Diffusers models use `config.json` in each subfolder. Custom models often use:
 
-**Module-level config dicts** (DreamID-Omni):
+**Module-level config dicts**:
 ```python
 VIDEO_CONFIG = {
     "patch_size": [1, 2, 2], "model_type": "ti2v",
@@ -181,7 +180,7 @@ vae_cfg = bagel_cfg.get("vae_config", {})
 
 ## Custom Architecture Patterns
 
-### Split forward API (DreamID-Omni)
+### Split forward API
 
 When a fusion model needs to interleave blocks from two backbones:
 

.claude/skills/add-tts-model/SKILL.md

@@ -345,8 +345,8 @@ Heavier scenarios in the same file use **`advanced_model` only** (streaming, ext
 @pytest.mark.tts
 @hardware_test(res={"cuda": "L4"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_voice_clone_en_non_streaming_001(omni_server, openai_client) -> None:
-    openai_client.send_audio_speech_request({...})
+def test_voice_clone_en_non_streaming_001(omni_server, online_client) -> None:
+    online_client.send_audio_speech_request({...})
 ```
 
 **L4 consolidation:** Prefer parametrized `OmniServerParams` rows (`default`, `async_chunk`, feature flags) in one expansion module rather than many merge-only files ([#1832](https://github.com/vllm-project/vllm-omni/issues/1832)).
@@ -363,8 +363,8 @@ def test_voice_clone_en_non_streaming_001(omni_server, openai_client) -> None:
 
 | Path | Fixture | Call |
 |------|---------|------|
-| Online `/v1/*` | `openai_client` | `openai_client.send_*_request(request_config)` |
-| Offline inference | `omni_runner_handler` | `omni_runner_handler.send_*_request(request_config)` |
+| Online `/v1/*` | `online_client` | `online_client.send_*_request(request_config)` |
+| Offline inference | `offline_client` | `offline_client.send_*_request(request_config)` |
 
 1. **Grep `runtime.py` first** — reuse `send_omni_request`, `send_diffusion_request`, `send_audio_speech_request` (online + offline Qwen-style TTS), `send_single_stage_tts_request` (Coqui XTTS / MOSS-TTS-Nano offline), etc.
 2. **No matching helper** → add `send_<feature>_request` (or `send_<route>_http_request` for negative/dfx) in **`runtime.py`** with general `assert_*` bundled inside, **then** call it from the test.
@@ -396,7 +396,7 @@ See [vllm-omni-test skill](../vllm-omni-test/SKILL.md) § **Runtime send helpers
   `/health` return non-200 until you're actually ready.
 - **Mark tests per the CI Levels table** — baseline smoke: `core_model` + `advanced_model`; heavier cases: `advanced_model` only; L4 expansion: `full_model`
 - **No per-model helper modules** — do not create `tests/helpers/{slug}.py`; keep constants and `request_config` payloads in the test file
-- **Online and offline e2e go through `runtime.py`** — `openai_client.send_audio_speech_request` (online); `omni_runner_handler.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests
+- **Online and offline e2e go through `runtime.py`** — `online_client.send_audio_speech_request` (online); `offline_client.send_audio_speech_request` (Qwen-style offline) or `send_single_stage_tts_request` (single-stage offline). Add a new `send_*_request` in `runtime.py` when none fits; do not embed `omni.generate` or HTTP in tests
 
 ## Phase 4: Async Chunk (Streaming)
 

.claude/skills/vllm-omni-test/SKILL.md

@@ -94,8 +94,8 @@ Existing references: `tests/e2e/offline_inference/test_qwen2_5_omni.py` (L2-styl
 
 | Scenario | Typical location | Fixtures / runner pattern | Baseline markers & level |
 |----------|------------------|---------------------------|---------------------------|
-| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `omni_runner_handler`. **Function (isolation only):** `omni_runner_function` + `omni_runner_handler_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
-| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `openai_client`. **Function (isolation only):** `omni_server_function` + `openai_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
+| **Offline inference e2e** | `tests/e2e/offline_inference/` | **Module (default):** `omni_runner` + `offline_client`. **Function (isolation only):** `omni_runner_function` + `offline_client_function`. Diffusion/TTS may use `Omni(...).generate` directly | L2: `core_model` + **one of** `omni` / `tts` / `diffusion`; `@hardware_test(...)` when GPU/NPU is required |
+| **Online serving e2e** | `tests/e2e/online_serving/` | **Module (default):** `omni_server` + `online_client`. **Function (isolation only):** `omni_server_function` + `online_client_function`. Clients: `send_omni_request` (omni), `send_audio_speech_request` (tts), `send_diffusion_request` / `send_video_diffusion_request` / `send_images_generations_request` (diffusion) | Baseline smoke: **`core_model` + `advanced_model`**; heavier paths: `advanced_model` only; L4 expansion: `full_model` |
 | **Documentation / runnable examples** | `tests/examples/offline_inference/`, `tests/examples/online_serving/` | **Offline docs (preferred):** extract Python/Bash blocks from the doc README (e.g. `ReadmeSnippet.extract_readme_snippets`), `pytest.mark.parametrize` each snippet, run via `example_runner.run` with a stable `output_subfolder`. **Online docs:** copy client/request scripts into dedicated tests and keep them in sync with the doc page. | Usually **L4**: `advanced_model`, often `example` plus hardware marks matching the nightly docs-example job (see `.buildkite/cuda/test-nightly.yml`). Full conventions: [docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md) (introduced in [PR #1910](https://github.com/vllm-project/vllm-omni/pull/1910): naming, output directory layout, skip rules, avoid trimming `num_inference_steps` without a strong CI reason). |
 | **Performance / benchmark** | `tests/dfx/perf/tests/*.json` + `run_*_benchmark.py` | JSON or script-driven server + load config; assert explicit metrics / baselines | L4 Perf: JSON `mark` with `full_model` + `omni`/`tts`/`diffusion`; wire `test-nightly.yml` Perf steps |
 | **Invalid parameter / negative HTTP validation** | `tests/dfx/reliability/invalid_param_test/` | Live `omni_server` + low-level `send_*_http_request` with `err_code` / `err_message` | `pytest.mark.slow` + `omni` / `tts` / `diffusion` + `@hardware_marks` (`H100` or `L4`); CI in **`test-weekly.yml`** (not ready/merge/nightly) |
@@ -110,9 +110,9 @@ After choosing offline/online/docs/perf, classify the **product under test** and
 |-----------|------------------------------|----------------------------|----------------------------------------|
 | **What it is** | Multimodal LLM pipeline (thinker/talker/stages; text + vision + audio I/O) | Speech synthesis / voice models | Generative diffusion (noise → image, audio, text, or video) |
 | **Examples** | Qwen2.5-Omni, Qwen3-Omni | Qwen3-TTS, VoxCPM2, Higgs-Audio, Step-Audio2 | Qwen-Image, BAGEL, Wan2.2, HunyuanVideo |
-| **Offline runner** | `omni_runner` + `omni_runner_handler`, `generate_multimodal` | `Omni(...).generate` with deploy YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
-| **Online client** | `openai_client.send_omni_request` (chat completions, modalities) | `openai_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
-| **Typical assertions** | Stage outputs, text/audio keywords via `OmniRunnerHandler` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
+| **Offline runner** | `omni_runner` + `offline_client`, `generate_multimodal` | `Omni(...).generate` with deploy YAML + TTS params | `Omni(...).generate` + `OmniDiffusionSamplingParams` |
+| **Online client** | `online_client.send_omni_request` (chat completions, modalities) | `online_client.send_audio_speech_request` (`/v1/audio/speech`) | `send_diffusion_request` (chat/T2I), `send_video_diffusion_request` (`/v1/videos`, X2V), or `send_images_generations_http_request` / `send_images_edits_http_request` (DALL-E routes) |
+| **Typical assertions** | Stage outputs, text/audio keywords via `OfflineOmniClient` / response handler | WAV bytes, stream chunks, speech endpoint contract | Image/video dimensions, `final_output_type`, `assert_diffusion_response` |
 | **Deploy YAML** | Per-model deploy configs (`ci/qwen3_omni_moe.yaml`, …) | `qwen3_tts.yaml`, `voxcpm2.yaml`, … | Often default serve; parallel/offload YAML for heavy DiT |
 | **Nightly group** (`test-nightly.yml`) | **Omni Model Test** — `-m "full_model and omni"` | **TTS Model Test** — `-m "full_model and tts"` | **Diffusion X2I(&A&T)** *or* **Diffusion X2V** (see Step 3 table; same `diffusion` marker, different YAML group / file shard) |
 | **L4 pressure** | Expansion per modality/model as needed | Expansion + accuracy/perf in TTS group | X2I: merge feature combos per [#1832](https://github.com/vllm-project/vllm-omni/issues/1832); X2V: separate nightly group |
@@ -175,9 +175,9 @@ See **L1 unit test constraints (mocking)** below for the full do/don't list.
 @pytest.mark.omni
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_<scenario>(omni_runner, omni_runner_handler) -> None:
+def test_<scenario>(omni_runner, offline_client) -> None:
     request_config = {"prompts": ..., "modalities": [...]}  # optional: images, videos, audios
-    omni_runner_handler.send_omni_request(request_config)
+    offline_client.send_omni_request(request_config)
 ```
 
 *Offline generative e2e — **Diffusion** (representative):*
@@ -187,8 +187,8 @@ def test_<scenario>(omni_runner, omni_runner_handler) -> None:
 @pytest.mark.diffusion
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
-def test_text_to_image_001(omni_runner_handler) -> None:
-    omni_runner_handler.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
+def test_text_to_image_001(offline_client) -> None:
+    offline_client.send_diffusion_request({"prompt": "...", "extra_body": {"num_inference_steps": 4, ...}})
 ```
 
 *Offline TTS e2e — **Qwen3-TTS** (two-stage; representative):*
@@ -198,8 +198,8 @@ def test_text_to_image_001(omni_runner_handler) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_runner, omni_runner_handler) -> None:
-    omni_runner_handler.send_audio_speech_request({
+def test_text_to_audio_001(omni_runner, offline_client) -> None:
+    offline_client.send_audio_speech_request({
         "input": "...",
         "task_type": "Base",
         "ref_audio": REF_AUDIO_URL,
@@ -214,8 +214,8 @@ def test_text_to_audio_001(omni_runner, omni_runner_handler) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_runner", tts_server_params, indirect=True)
-def test_voice_clone_001(omni_runner_handler) -> None:
-    omni_runner_handler.send_single_stage_tts_request({
+def test_voice_clone_001(offline_client) -> None:
+    offline_client.send_single_stage_tts_request({
         "input": "...",
         "language": "en",
         "prompt_audio_path": REF_AUDIO_PATH,
@@ -231,9 +231,9 @@ def test_voice_clone_001(omni_runner_handler) -> None:
 @pytest.mark.omni
 @hardware_test(...)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
-def test_text_to_text_001(omni_server, openai_client) -> None:
+def test_text_to_text_001(omni_server, online_client) -> None:
     request_config = {"model": omni_server.model, "messages": ..., "modalities": ["text"]}
-    openai_client.send_omni_request(request_config)
+    online_client.send_omni_request(request_config)
 ```
 
 *Online serving e2e — **TTS** (representative):*
@@ -244,9 +244,9 @@ def test_text_to_text_001(omni_server, openai_client) -> None:
 @pytest.mark.tts
 @hardware_test(...)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
-def test_text_to_audio_001(omni_server, openai_client) -> None:
+def test_text_to_audio_001(omni_server, online_client) -> None:
     request_config = {"model": omni_server.model, "input": "...", "response_format": "wav", ...}
-    openai_client.send_audio_speech_request(request_config)
+    online_client.send_audio_speech_request(request_config)
 ```
 
 *Online serving e2e — **Diffusion** X2I (representative):*
@@ -255,8 +255,8 @@ def test_text_to_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.core_model
 @pytest.mark.diffusion
 @pytest.mark.parametrize("omni_server", _get_default_case(MODEL), indirect=True)
-def test_text_to_image_001(omni_server, openai_client) -> None:
-    openai_client.send_diffusion_request({...})  # chat completions + extra_body
+def test_text_to_image_001(omni_server, online_client) -> None:
+    online_client.send_diffusion_request({...})  # chat completions + extra_body
 ```
 
 *Online serving e2e — **Diffusion** X2V (representative):*
@@ -264,8 +264,8 @@ def test_text_to_image_001(omni_server, openai_client) -> None:
 ```python
 @pytest.mark.core_model
 @pytest.mark.diffusion
-def test_text_to_video_001(omni_server, openai_client) -> None:
-    openai_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
+def test_text_to_video_001(omni_server, online_client) -> None:
+    online_client.send_video_diffusion_request({"model": ..., "form_data": {...}})  # /v1/videos
 ```
 
 *Documentation example tests:* follow the **Preferred Test Strategy** in [l4_doc_example_tests.inc.md](../../../docs/contributing/ci/test_examples/l4_doc_example_tests.inc.md): dynamic extraction for offline READMEs; explicit copied client code for online pages until extraction is justified; use the documented **naming**, **output directory** (page folder + case id), and **skipping** rules (e.g. Gradio-only scripts).
@@ -324,21 +324,21 @@ vLLM-Omni e2e tests start a real **OmniServer** (online) or **OmniRunner** (offl
 
 | Scope | Online fixtures | Offline fixtures | When to use |
 |-------|-----------------|------------------|-------------|
-| **Module** (default) | `omni_server` → `openai_client` | `omni_runner` → `omni_runner_handler` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
-| **Function** | `omni_server_function` → `openai_client_function` | `omni_runner_function` → `omni_runner_handler_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |
+| **Module** (default) | `omni_server` → `online_client` | `omni_runner` → `offline_client` | **Default for L2/L3/L4 expansion** — amortize model/server init across `test_*` in the same module. Same `OmniServerParams` / runner config can be reused by multiple tests. |
+| **Function** | `omni_server_function` → `online_client_function` | `omni_runner_function` → `offline_client_function` | **Only when required** — each `test_*` must get a **clean** server/runner (no shared engine/GPU state). Typical: `tests/dfx/reliability/`, sleep/wakeup, crash/restart, tests that mutate global server state. |
 
 **Rules:**
 
 1. **Default to module scope** (`omni_server` / `omni_runner`) unless the scenario **explicitly** needs a fresh instance per test function.
-2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `openai_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `openai_client_function`. Do not mix module fixture with function client (or vice versa).
+2. **Indirect parametrize name must match the fixture name**: `@pytest.mark.parametrize("omni_server", ...)` with `omni_server` + `online_client`; `@pytest.mark.parametrize("omni_server_function", ...)` with `omni_server_function` + `online_client_function`. Do not mix module fixture with function client (or vice versa).
 3. Different `OmniServerParams` per `test_*` (e.g. default vs `--enable-cpu-offload`) is still OK with **module** `omni_server` — pytest parametrizes per test node; only switch to `_function` when isolation between tests matters, not merely because `server_args` differ.
 4. One `test_*` with **many** server configs (expansion matrix) → single function + `@pytest.mark.parametrize("omni_server", [...], indirect=True)` + module `omni_server` (see in-tree `test_qwen_image_expansion.py`).
 
 ```python
 # Default — module-scoped server (L4 expansion)
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=H100)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
-    openai_client.send_images_generations_request({...})
+def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
+    online_client.send_images_generations_request({...})
 
 # Function-scoped — reliability / per-test clean state only
 @pytest.mark.parametrize(
@@ -346,9 +346,9 @@ def test_foo_images_generations_default_1024(omni_server, openai_client) -> None
     [pytest.param(OmniServerParams(model=MODEL), marks=H100)],
     indirect=True,
 )
-def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) -> None:
-    openai_client_function.send_omni_sleep_http_request({...})
-    openai_client_function.send_omni_wakeup_http_request({...})
+def test_foo_sleep_wakeup_cycle(omni_server_function, online_client_function) -> None:
+    online_client_function.send_omni_sleep_http_request({...})
+    online_client_function.send_omni_wakeup_http_request({...})
 ```
 
 #### Runtime send helpers — online **and** offline (`tests/helpers/runtime.py`)
@@ -357,19 +357,19 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) ->
 
 | Principle | Action |
 |-----------|--------|
-| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OpenAIClientHandler` (online) or `OmniRunnerHandler` (offline) that matches the endpoint / pipeline shape. |
+| **Reuse first** | Grep `runtime.py` for an existing **`send_*_request`** on `OnlineOmniClient` (online) or `OfflineOmniClient` (offline) that matches the endpoint / pipeline shape. |
 | **Extend when close** | If an existing `send_*_request` almost fits (e.g. missing one optional field), extend it in `runtime.py` — do not fork logic in the test file. |
 | **Add when missing** | No suitable helper → add **`send_<feature>_request`** (high-level: call + general `assert_*`) or **`send_<route>_<verb>_http_request`** (low-level HTTP for negative/dfx) in **`runtime.py` first**, then call it from tests. |
 | **Test module owns payload only** | In `test_*.py`: `MODEL`, deploy path, vendored media, `get_prompt()`, and inline **`request_config` dicts** only. **No** `omni.generate(...)`, raw `requests.post`, OpenAI SDK calls, or `_collect_audio()` / `_process_output()` in e2e tests. |
 
-**Online** (`openai_client` from `omni_server`): `OpenAIClientHandler.send_*_request`.
+**Online** (`online_client` from `omni_server`): `OnlineOmniClient.send_*_request`.
 
-**Offline** (`omni_runner_handler` from `omni_runner`): `OmniRunnerHandler.send_*_request`.
+**Offline** (`offline_client` from `omni_runner`): `OfflineOmniClient.send_*_request`.
 
 | Do | Don't |
 |----|-------|
-| Online: `openai_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
-| Offline: `omni_runner_handler.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
+| Online: `online_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_video_diffusion_request`, `send_images_generations_request`, … | `requests.post(f"{base_url}/v1/...", json=…)` or `client.chat.completions.create(...)` inside a test |
+| Offline: `offline_client.send_omni_request`, `send_diffusion_request`, `send_audio_speech_request`, `send_single_stage_tts_request`, `send_single_stage_tts_batch_request`, … | `omni_runner.omni.generate(...)` + hand-rolled tensor/WAV extraction in `test_*.py` |
 | Add missing `send_*` to **`runtime.py`** first; bundle general `assert_*` inside the send helper | A one-off `def _post_*` / `def _collect_audio` at the bottom of a test module |
 | Mirror naming/style of neighboring `send_*` (docstring, `request_config` dict, optional `run_level`, `err_code` / `err_message` for negative cases) | Different parameter shapes per test file for the same endpoint |
 
@@ -392,7 +392,7 @@ def test_foo_sleep_wakeup_cycle(omni_server_function, openai_client_function) ->
 
 Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not** special validation — the general `assert_*` should read those from `request_config` inside `send_*`.
 
-**`send_*` ↔ `assert_*` pairing (online `OpenAIClientHandler`):**
+**`send_*` ↔ `assert_*` pairing (online `OnlineOmniClient`):**
 
 | High-level `send_*` (prefer in L2+ e2e) | Assert already invoked inside `runtime.py` | Low-level HTTP-only sibling |
 |----------------------------------------|--------------------------------------------|-----------------------------|
@@ -405,7 +405,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 
 **Rule:** Tests call high-level `send_*_request` for the general contract. **Never** call the same bundled `assert_*` again. Call an extra `assert_*` in the test **only** for special, case-specific validation.
 
-**Common `OpenAIClientHandler` entry points** (non-exhaustive — grep `runtime.py` before adding):
+**Common `OnlineOmniClient` entry points** (non-exhaustive — grep `runtime.py` before adding):
 
 | Area | High-level (SDK + assert) | Low-level HTTP (`*_http_request`) |
 |------|---------------------------|-----------------------------------|
@@ -416,7 +416,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | TTS (online) | `send_audio_speech_request` | `send_audio_speech_http_request`, `send_audio_generate_http_request`, … |
 | Ops / meta | — | `send_health_http_request`, `send_models_http_request`, `send_omni_sleep_http_request`, … |
 
-**Common `OmniRunnerHandler` entry points** (offline — grep `runtime.py` before adding):
+**Common `OfflineOmniClient` entry points** (offline — grep `runtime.py` before adding):
 
 | Area | High-level (`send_*_request` + assert) | Notes |
 |------|----------------------------------------|-------|
@@ -426,7 +426,7 @@ Changing `request_config` fields (`size`, `n`, `seed`, server flags) is **not**
 | Single-stage TTS (Coqui XTTS, MOSS-TTS-Nano, …) | `send_single_stage_tts_request`, `send_single_stage_tts_batch_request` | `prompt` + `additional_information` + `omni.generate` — **do not** duplicate in tests |
 | New model family | **Add** `send_<family>_request` here first | Then call from `tests/e2e/offline_inference/test_*.py` |
 
-L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OpenAIClientHandler` / `OmniRunnerHandler`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.
+L1 tests under `tests/entrypoints/` may use **FastAPI `TestClient`** or direct handler calls with mocks; they do **not** need `OnlineOmniClient` / `OfflineOmniClient`. **L2+ online and offline e2e** must use `runtime.py` `send_*` helpers.
 
 #### Invalid parameter validation (`tests/dfx/reliability/invalid_param_test/`)
 
@@ -452,7 +452,7 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 | **Module markers** | `pytestmark = [pytest.mark.slow, pytest.mark.<omni or tts or diffusion>]` |
 | **Server fixture** | `_PARAMS` / `_QWEN3_TTS_SPEECH`-style list of `pytest.param(OmniServerParams(...), id="...", marks=hardware_marks(...))`; `@pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)` |
 | **Hardware** | `hardware_marks(res={"cuda": "H100"})` for heavy diffusion/omni/video; `hardware_marks(res={"cuda": "L4"})` for smaller TTS models (must match weekly `-m "slow and L4"` step) |
-| **HTTP client** | **Low-level** `openai_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
+| **HTTP client** | **Low-level** `online_client.send_*_http_request({..., "err_code": 400, "err_message": (...)} )` — **not** `send_*_request` (success path) |
 | **Case shape** | Prefer **one `test_*` per route family** + `@pytest.mark.parametrize("body_spec, err_message", [...])` with stable `id=` per case; or dedicated `test_<route>_malformed_json` when not parametrized |
 | **Body helpers** | `_minimal_<endpoint>_json(omni_server)` / `_minimal_*_form_data()`; merge overrides with `body.update(body_spec)` |
 | **Known gaps** | `pytest.mark.skip(reason="…#3649")` as `_SKIP_ISSUE_3649` when server validation is not yet strict (mirror neighboring cases) |
@@ -485,13 +485,13 @@ When the user asks for **invalid parameter validation**, **invalid request bodie
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_images_generations_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     body_spec: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     body = _minimal_images_gen_json(omni_server)
     body.update(body_spec)
-    openai_client.send_images_generations_http_request(
+    online_client.send_images_generations_http_request(
         {"json": body, "timeout": 300, "err_code": 400, "err_message": err_message}
     )
 ```
@@ -542,18 +542,18 @@ def send_images_generations_request(self, request_config: dict[str, Any], ...) -
 
 # tests/e2e/.../test_foo_expansion.py — one case per test_* (no case_id branching)
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_default_1024(omni_server, openai_client) -> None:
-    openai_client.send_images_generations_request({"json": body, "timeout": 300})
+def test_foo_images_generations_default_1024(omni_server, online_client) -> None:
+    online_client.send_images_generations_request({"json": body, "timeout": 300})
 
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_images_generations_sizes_256_512_1024(omni_server, openai_client) -> None:
+def test_foo_images_generations_sizes_256_512_1024(omni_server, online_client) -> None:
     for size in ("256x256", "512x512", "1024x1024"):
-        openai_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})
+        online_client.send_images_generations_request({"json": {**body, "size": size}, "timeout": 300})
 
 # chat diffusion — separate test; send_diffusion_request bundles assert_diffusion_response
 @pytest.mark.parametrize("omni_server", [pytest.param(OmniServerParams(model=MODEL), marks=SINGLE_L4)], indirect=True)
-def test_foo_chat_completions_t2i_fallback(omni_server, openai_client) -> None:
-    openai_client.send_diffusion_request({...})
+def test_foo_chat_completions_t2i_fallback(omni_server, online_client) -> None:
+    online_client.send_diffusion_request({...})
 ```
 
 L1 tests under `tests/entrypoints/` may keep **minimal** asserts next to the handler under test when validating a single branch; still prefer `assertions.py` when the same JSON/media check appears in more than one file or level (e.g. L1 protocol test + L4 e2e).

.claude/skills/vllm-omni-test/references/test-routing.md

@@ -279,7 +279,7 @@ When adding or modifying tests, do not stop at “where the file lives” — al
 4. **API client + assert placement**: **General** validation → implement in `assertions.py`, call **inside** `send_*_request` in `runtime.py`; tests call **`send_*_request` only**. **Special** case validation → `assert_*` in `assertions.py`, called **in the test** after `send_*_request`. Low-level `send_*_http_request` is for negative/dfx tests (`err_code`), not ordinary L2+ success e2e.
 5. **Shared assertions**: logic in **`tests/helpers/assertions.py`**; general checks inside `send_*_request`; special checks only in the test. No `_assert_*` in `test_*.py`.
 6. **One case → one `test_*`**: function name reflects what is validated; no `if case_id == ...` mega-test merging multiple scenarios.
-7. **Fixture scope**: default **`omni_server` + `openai_client`** / **`omni_runner` + `omni_runner_handler`** (module). Use **`omni_server_function` + `openai_client_function`** / **`omni_runner_function` + `omni_runner_handler_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
+7. **Fixture scope**: default **`omni_server` + `online_client`** / **`omni_runner` + `offline_client`** (module). Use **`omni_server_function` + `online_client_function`** / **`omni_runner_function` + `offline_client_function`** only when each `test_*` must spawn a fresh instance. Parametrize name must match fixture (`omni_server` vs `omni_server_function`).
 8. **Type marker**: `omni`, `tts`, or `diffusion` on every model e2e module.
 9. **Diffusion L4 Function**: wire `*_expansion.py` into **X2I(&A&T)** or **X2V** **Function Test** in **`test-nightly.yml` only** — do not add `test-merge.yml` unless the user also requested L3.
 10. **Diffusion L4 Perf** (only when requested): add `tests/dfx/perf/tests/test_<slug>_vllm_omni.json` + **Perf Test · &lt;Model&gt;** step (artifact upload); not part of “L4 functional cases” by default.

benchmarks/tts/benchmark_cosyvoice3_trt_streams.py

@@ -0,0 +1,151 @@
+"""Benchmark CosyVoice3-style CUDA stream handoff strategies."""
+
+import argparse
+import json
+import statistics
+import time
+
+import torch
+
+from vllm_omni.platforms import current_omni_platform
+
+
+def run_case(
+    *,
+    device: torch.device,
+    steps: int,
+    warmup_steps: int,
+    producer_cycles: int,
+    estimator_cycles: int,
+    use_stream_dependencies: bool,
+) -> dict[str, float | int | str]:
+    caller_stream = torch.cuda.Stream(device=device)
+    estimator_stream = torch.cuda.Stream(device=device)
+    input_tensor = torch.zeros(1024, device=device)
+    output_tensor = torch.empty_like(input_tensor)
+    submission_ms = []
+
+    def step(measure: bool) -> None:
+        with torch.cuda.stream(caller_stream):
+            torch.cuda._sleep(producer_cycles)
+            input_tensor.add_(1)
+
+        start = time.perf_counter_ns()
+        if use_stream_dependencies:
+            estimator_stream.wait_stream(caller_stream)
+        else:
+            caller_stream.synchronize()
+
+        with torch.cuda.stream(estimator_stream):
+            torch.cuda._sleep(estimator_cycles)
+            output_tensor.copy_(input_tensor)
+
+        if use_stream_dependencies:
+            caller_stream.wait_stream(estimator_stream)
+        else:
+            estimator_stream.synchronize()
+
+        if measure:
+            submission_ms.append((time.perf_counter_ns() - start) / 1e6)
+
+    for _ in range(warmup_steps):
+        step(measure=False)
+    caller_stream.synchronize()
+    estimator_stream.synchronize()
+    input_tensor.zero_()
+    torch.accelerator.reset_peak_memory_stats(device)
+    initial_allocated = torch.accelerator.memory_allocated(device)
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+    wall_start = time.perf_counter()
+    with torch.cuda.stream(caller_stream):
+        start_event.record()
+    for _ in range(steps):
+        step(measure=True)
+    with torch.cuda.stream(caller_stream):
+        end_event.record()
+    end_event.synchronize()
+    wall_ms = (time.perf_counter() - wall_start) * 1e3
+    peak_delta_bytes = torch.accelerator.max_memory_allocated(device) - initial_allocated
+
+    expected = torch.full_like(output_tensor, steps)
+    torch.testing.assert_close(output_tensor, expected, rtol=0, atol=0)
+    return {
+        "mode": "wait_stream" if use_stream_dependencies else "synchronize",
+        "steps": steps,
+        "submission_median_ms": statistics.median(submission_ms),
+        "submission_p95_ms": statistics.quantiles(submission_ms, n=20)[18],
+        "submission_mean_ms": statistics.mean(submission_ms),
+        "gpu_elapsed_ms": start_event.elapsed_time(end_event),
+        "wall_elapsed_ms": wall_ms,
+        "peak_allocation_delta_bytes": peak_delta_bytes,
+        "final_value": int(output_tensor[0].item()),
+    }
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--device", default="cuda:0")
+    parser.add_argument("--steps", type=int, default=100)
+    parser.add_argument("--warmup-steps", type=int, default=10)
+    parser.add_argument("--repeats", type=int, default=7)
+    parser.add_argument("--producer-cycles", type=int, default=1_000_000)
+    parser.add_argument("--estimator-cycles", type=int, default=2_000_000)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    device = torch.device(args.device)
+    current_omni_platform.set_device(device)
+    torch.manual_seed(0)
+
+    old_results = []
+    new_results = []
+    for repeat in range(args.repeats):
+        common_args = {
+            "device": device,
+            "steps": args.steps,
+            "warmup_steps": args.warmup_steps,
+            "producer_cycles": args.producer_cycles,
+            "estimator_cycles": args.estimator_cycles,
+        }
+        if repeat % 2 == 0:
+            old_results.append(run_case(**common_args, use_stream_dependencies=False))
+            new_results.append(run_case(**common_args, use_stream_dependencies=True))
+        else:
+            new_results.append(run_case(**common_args, use_stream_dependencies=True))
+            old_results.append(run_case(**common_args, use_stream_dependencies=False))
+
+    old_submission_ms = statistics.median(result["submission_median_ms"] for result in old_results)
+    new_submission_ms = statistics.median(result["submission_median_ms"] for result in new_results)
+    old_gpu_ms = statistics.median(result["gpu_elapsed_ms"] for result in old_results)
+    new_gpu_ms = statistics.median(result["gpu_elapsed_ms"] for result in new_results)
+    print(
+        json.dumps(
+            {
+                "device": torch.cuda.get_device_name(device),
+                "torch": torch.__version__,
+                "cuda": torch.version.cuda,
+                "repeats": args.repeats,
+                "warmup_steps": args.warmup_steps,
+                "producer_cycles": args.producer_cycles,
+                "estimator_cycles": args.estimator_cycles,
+                "old_submission_median_ms": old_submission_ms,
+                "new_submission_median_ms": new_submission_ms,
+                "submission_median_speedup": old_submission_ms / new_submission_ms,
+                "old_gpu_elapsed_median_ms": old_gpu_ms,
+                "new_gpu_elapsed_median_ms": new_gpu_ms,
+                "gpu_elapsed_change_percent": (new_gpu_ms - old_gpu_ms) / old_gpu_ms * 100,
+                "parity": all(result["final_value"] == args.steps for result in old_results + new_results),
+                "old_runs": old_results,
+                "new_runs": new_results,
+            },
+            indent=2,
+        )
+    )
+
+
+if __name__ == "__main__":
+    main()

collect_env.py

@@ -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
 
 # ruff: noqa
 # code borrowed from https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py
@@ -494,6 +494,16 @@ def is_xnnpack_available():
 
 
 def get_env_vars():
+    try:
+        from vllm_omni.config.environment_variable_inventory import ENVIRONMENT_VARIABLE_INVENTORY
+    except Exception:
+        # Keep environment collection usable when an incomplete/broken Omni
+        # installation is the reason the user is running this diagnostic.
+        # Importing an Omni submodule executes package initialization, which
+        # may fail with RuntimeError, AssertionError, or OSError as well as
+        # ImportError when vLLM or the accelerator stack is broken.
+        ENVIRONMENT_VARIABLE_INVENTORY = {}
+
     env_vars = ""
     secret_terms = ("secret", "token", "api", "access", "password")
     report_prefix = (
@@ -507,12 +517,17 @@ def get_env_vars():
         "MKL_",
         "NVIDIA",
     )
-    for k, v in os.environ.items():
-        if any(term in k.lower() for term in secret_terms):
+    for k, v in sorted(os.environ.items()):
+        omni_classification = ENVIRONMENT_VARIABLE_INVENTORY.get(k)
+        if (omni_classification is not None and omni_classification.redact_value) or any(
+            term in k.lower() for term in secret_terms
+        ):
             continue
-        if k in environment_variables:
-            env_vars = env_vars + "{}={}".format(k, v) + "\n"
-        if k.startswith(report_prefix):
+        if (
+            k in environment_variables
+            or (omni_classification is not None and omni_classification.is_public_omni)
+            or k.startswith(report_prefix)
+        ):
             env_vars = env_vars + "{}={}".format(k, v) + "\n"
 
     return env_vars

docs/.nav.yml

@@ -24,7 +24,6 @@ nav:
       - Text-To-Speech: user_guide/examples/offline_inference/text_to_speech.md
       - Text-To-Video: user_guide/examples/offline_inference/text_to_video.md
       - X-To-Text: user_guide/examples/offline_inference/x_to_text.md
-      - X-To-Video-Audio: user_guide/examples/offline_inference/x_to_video_audio.md
     - Online Serving:
       - Image-To-Image: user_guide/examples/online_serving/image_to_image.md
       - Image-To-Video: user_guide/examples/online_serving/image_to_video.md

docs/cli/README.md

@@ -2,6 +2,11 @@
 
 The CLI for vLLM-Omni inherits from vllm with some additional arguments.
 
+Environment variables are process configuration rather than CLI-only options.
+See the [Environment Variables](../configuration/environment_variables.md)
+reference for ownership, precedence, evaluation time, and per-stage `env`
+scoping.
+
 ## serve
 
 Starts the vLLM-Omni OpenAI Compatible API server.
@@ -39,4 +44,5 @@ vllm bench serve --omni \
     --num-prompts  5
 ```
 
-See [vllm bench serve](./bench/serve.md) for the full reference of all available arguments.
+See [vllm bench serve](./bench/serve.md) for serving benchmark arguments and
+dataset-specific examples, including OmniInteract native-duplex sessions.

docs/cli/bench/serve.md

@@ -2,7 +2,7 @@
 The vllm bench command launches the vLLM-Omni benchmark to evaluate the performance of multimodal models.
 
 ## Notes
-vLLM-Omni registers the `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, and `daily-omni` serving benchmark backends.
+vLLM-Omni registers the `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, `daily-omni`, and `openai-realtime-duplex` serving benchmark backends. It also adds the `omniinteract` dataset.
 
 ## Basic Parameter Description
 You can use `vllm bench serve --omni --help=all` to get descriptions of all parameters. The commonly used parameters are described below:
@@ -10,7 +10,7 @@ You can use `vllm bench serve --omni --help=all` to get descriptions of all para
   Enable Omni (multimodal) mode, supporting multimodal inputs and outputs such as images, videos, and audio.
 
 - `--backend`
-  Specify the backend adapter. vLLM-Omni adds `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, and `daily-omni` to the upstream vLLM backend choices.
+  Specify the backend adapter. vLLM-Omni adds `openai-chat-omni`, `openai-audio-speech`, `openai-image-edits-omni`, `daily-omni`, and `openai-realtime-duplex` to the upstream vLLM backend choices.
 
 - `--model`
   The model identifier to load, filled according to the models supported by vLLM-Omni.
@@ -19,7 +19,7 @@ You can use `vllm bench serve --omni --help=all` to get descriptions of all para
   The API endpoint exposed externally, to which clients send their requests.
 
 - `--dataset-name`
-  The name of the dataset used; random-mm indicates generating random multimodal inputs (images, videos, audio).
+  The name of the dataset used; random-mm indicates generating random multimodal inputs (images, videos, audio), while `omniinteract` replays official OmniInteract videos.
 
 - `--num-prompts`
   The total number of requests to send, an integer.
@@ -269,6 +269,44 @@ We use audio generation time / audio duration to calculate RTF.
 
 </details>
 
+### OmniInteract Realtime Benchmark
+
+OmniInteract runs each video as one native-duplex WebSocket sample in the standard serving-benchmark lifecycle:
+
+```bash
+vllm bench serve --omni \
+  --backend openai-realtime-duplex \
+  --dataset-name omniinteract \
+  --dataset-path /path/to/OmniInteract \
+  --model openbmb/MiniCPM-o-4_5 \
+  --base-url http://127.0.0.1:8000 \
+  --endpoint /v1/realtime \
+  --omniinteract-ref-audio /path/to/reference.wav \
+  --omniinteract-output-dir ./omniinteract-artifacts \
+  --num-prompts 3 \
+  --num-warmups 0
+```
+
+`--dataset-path` accepts an extracted directory, `data.tar[.gz]`, or a Hugging Face dataset ID; omitting it uses
+`lucky-lance/OmniInteract`. `--num-prompts` is the total across subsets and defaults to 3 for OmniInteract; explicit `0`
+selects all and oversize values use all available cases. Reference audio is required, and OmniInteract uses the
+`/v1/realtime` endpoint.
+
+Audio is replayed as 16 kHz PCM16 in 200 ms chunks and video at 1 FPS with real-time pacing. All selected media is decoded
+before timing and remains in client memory for the run, so `--max-concurrency` does not limit media preparation memory; use
+explicit `--num-prompts 0` only when the client has enough RAM for the full dataset. Media commands are bounded by
+`--omniinteract-media-timeout-s`, and concurrency defaults to 1. Standard request-rate, warmup, result-saving, and summary
+options apply. Use `--omniinteract-require-response` only for functional E2E cases; LISTEN is a valid benchmark result.
+
+Each completed case writes `output.wav`, `wav_transcript.json`, `events.json`, `result.json`, and a final `.done` marker under
+`--omniinteract-output-dir`. The root also contains `batch_summary.json` and `official_eval_manifest.jsonl`; failed cases write
+`.failed.json`. Runs sharing one output root are serialized. Completion validates transport, response lifecycle, and artifacts,
+not answer accuracy. Transcript timestamps are serialized playback-queue times. Clipped or cancelled outputs are ineligible and
+omitted from the official manifest; `audio_clipped_bytes` records output beyond the rounded video horizon.
+
+TTFT, TTFP, and RTF start at client receipt of `response.created`. TPOT/ITL use engine stage-0 timing; ITL is emitted only when
+every token interval is present.
+
 ### Multi-Modal Benchmark
 
 <details class="admonition abstract" markdown="1">

docs/configuration/README.md

@@ -7,6 +7,9 @@ For options within a vLLM Engine, please refer to the [vLLM 0.26 configuration g
 Each model defines fixed topology in a registered `PipelineConfig` and runtime
 overrides in a deploy YAML.
 
+For process-level settings shared by the CLI, deploy YAML, workers, and model
+integrations, see [Environment Variables](./environment_variables.md).
+
 For a specific example, see the [Qwen2.5-Omni deploy config](gh-file:vllm_omni/deploy/qwen2_5_omni.yaml). The matching frozen pipeline topology lives at [vllm_omni/model_executor/models/qwen2_5_omni/pipeline.py](gh-file:vllm_omni/model_executor/models/qwen2_5_omni/pipeline.py).
 
 For an introduction, see [Pipeline and deploy configurations](./stage_configs.md).

docs/configuration/environment_variables.md

@@ -0,0 +1,237 @@
+# Environment Variables
+
+Environment variables are process-level configuration shared by the CLI,
+deploy YAML, Python APIs, worker processes, and model integrations. They are
+therefore documented under **User Guide → Configuration**, rather than under
+CLI Reference.
+
+This page describes the public vLLM-Omni environment-variable contract on the
+`latest` documentation branch. Use the documentation for your installed
+vLLM-Omni release when behavior differs.
+
+!!! important "Public contract"
+    Only names in [Public vLLM-Omni variables](#public-vllm-omni-variables) are
+    public Omni configuration. A name found in source code is not automatically
+    supported: model-specific, benchmark, platform, and internal variables have
+    separate ownership and lifecycle rules.
+
+## Naming and implementation boundary
+
+New public variables owned by vLLM-Omni must use the `VLLM_OMNI_` prefix.
+Existing public names with older prefixes remain supported for compatibility,
+but they do not establish naming precedent for new settings. Inherited vLLM
+and third-party names retain the prefix chosen by their owning project.
+
+The source [environment-variable inventory](gh-file:vllm_omni/config/environment_variable_inventory.py)
+records ownership, migration disposition, and redaction metadata. Unlike
+`vllm.envs`, it is not an executable value resolver: existing consumers still
+own parsing, defaults, precedence, and evaluation time. The public tables below
+describe those live consumer contracts. Stable model or stage behavior should
+move to typed configuration, and request-varying behavior should move to a
+request schema instead of adding another environment variable.
+
+## Precedence and evaluation time
+
+There is no repository-wide rule that makes environment variables override (or
+lose to) every CLI, YAML, or request option. The tables below state precedence
+for each public setting. In general:
+
+- request fields should control behavior that can vary per request;
+- typed deploy or stage configuration should control stable model and stage
+  behavior;
+- environment variables are appropriate for process launch, platform
+  integration, emergency diagnostics, and compatibility fallbacks.
+
+An environment variable can be read while a module is imported, when server or
+model configuration is constructed, when a worker is spawned, or whenever a
+property is accessed. Set import- and startup-time variables before invoking
+`vllm serve` or importing vLLM-Omni. Changing the parent shell after workers
+start does not update existing worker environments.
+
+The lifecycle labels used below are:
+
+- **Stable** — reviewed public configuration; changes require compatibility
+  consideration.
+- **Experimental** — public for evaluation, but its contract may still evolve.
+- **Diagnostic** — supported for investigation or temporary mitigation, not as
+  a normal tuning interface.
+- **Deprecated** — compatibility alias; migrate to the named replacement.
+
+## Public vLLM-Omni variables
+
+### Diffusion backend and cache selection
+
+| Name | Type and default | Applies to and read time | Precedence and invalid values | Lifecycle |
+| --- | --- | --- | --- | --- |
+| `DIFFUSION_ATTENTION_BACKEND` | Backend name or `auto`; default `auto` (platform selection) | Diffusion stages; read when `OmniDiffusionConfig` is constructed | `diffusion_attention_config.default` wins. An unknown backend fails during backend resolution. | Stable fallback |
+| `DIFFUSION_CACHE_BACKEND` | `none`, `cache_dit`, `tea_cache`, `mag_cache`, `step_cache`, `stepcache`, or `step_cache_dit`; default `none` | Diffusion runner startup | Explicit `cache_backend` in config wins. Otherwise this name wins over the deprecated alias. Unsupported values raise `ValueError` during runner setup. | Stable fallback |
+| `DIFFUSION_CACHE_ADAPTER` | Same values as `DIFFUSION_CACHE_BACKEND`; default `none` | Diffusion runner startup | Used only when neither explicit `cache_backend` nor `DIFFUSION_CACHE_BACKEND` is set. Unsupported values raise `ValueError`. | Deprecated; use `DIFFUSION_CACHE_BACKEND` |
+| `OMNI_DIFFUSION_PROMPT_EMBED_CACHE` | Boolean: `1`, `true`, `yes`, `on`, `0`, `false`, `no`, or `off`; default disabled | Each diffusion runner; resolved during model setup | A recognized environment value overrides the explicit enable setting. An unrecognized value is ignored. | Experimental |
+| `OMNI_DIFFUSION_PROMPT_EMBED_CACHE_SIZE` | Positive integer; default `32` entries | Each diffusion runner; resolved during model setup | A valid environment value overrides the explicit cache size. A non-integer logs a warning; a non-positive value is ignored. | Experimental |
+| `OMNI_DIFFUSION_SESSION_STATE_MANAGER` | Boolean with the same accepted spellings as the prompt cache; default disabled | Experimental diffusion session manager; model setup | A recognized environment value overrides the explicit enable setting. An unrecognized value is ignored. | Experimental |
+| `OMNI_DIFFUSION_SESSION_STATE_MANAGER_MAX_SESSIONS` | Positive integer; default `64` | Experimental diffusion session manager; model setup | A valid environment value overrides the explicit maximum. A non-integer logs a warning; a non-positive value is ignored. | Experimental |
+
+Backend names for `DIFFUSION_ATTENTION_BACKEND` are the members of
+`DiffusionAttentionBackendEnum`, such as `FLASH_ATTN`, `TORCH_SDPA`,
+`SAGE_ATTN`, `FLASHINFER_ATTN`, and `TRTLLM_ATTN`. Platform support still
+depends on the installed kernels and model path.
+
+### Serving and runtime
+
+| Name | Type and default | Applies to and read time | Precedence and invalid values | Lifecycle |
+| --- | --- | --- | --- | --- |
+| `SPEAKER_SAMPLES_DIR` | Filesystem path; default `~/.cache/vllm-omni/speakers` | Speech server; read when speaker storage initializes | Environment-only setting. The directory is created; filesystem errors propagate. | Stable |
+| `SPEAKER_MAX_UPLOADED` | Integer; default `1000` | Speech server; read when speaker storage initializes | Environment-only setting. A non-integer logs a warning and uses `1000`; range is not otherwise validated. | Stable |
+| `VLLM_OMNI_ASYNC_OUTPUT_TIMEOUT` | Float seconds; default `600` | Diffusion engine async-output wait in `step_streaming`; resolved per call on the request path, not at import | Environment-only setting. A non-float or `<=0` value warns once and uses the default. | Experimental |
+| `VLLM_OMNI_INPUT_WAIT_TIMEOUT_S` | Float seconds; default `600`; `<=0` disables | Full-payload input coordinator, not async-chunk transfer; read when the scheduler module imports in each worker | Environment-only setting. A non-float logs a warning and uses `600`. | Stable operational control |
+| `VLLM_OMNI_ORCH_MONITOR_PATH` | Filesystem path; default `<current-working-directory>/vllm_omni_orch_monitor_<timestamp>.json` | Orchestrator monitor enabled by `--enable-orch-monitor`; read when the monitor is created | Environment-only path override. Parent directories are created; write errors are logged. | Diagnostic |
+| `VLLM_OMNI_VIDEO_SYNC_TIMEOUT` | Float seconds; default `600` | Synchronous Videos API; read when the API server module imports | Environment-only setting. A non-float raises `ValueError` during import. | Experimental |
+| `VLLM_VIDEO_ASYNC_CHUNK` | `on` or `off`; default `on` | Streaming video output; read on attribute access | Environment-only setting. Values are trimmed and case-normalized; an invalid value warns once and uses `on`. | Experimental |
+| `VLLM_VIDEO_AUDIO_DELTA_MODE` | `fast` or `slow`; default `fast` | Streaming video audio deltas; read on attribute access | Environment-only setting. Values are trimmed and case-normalized; an invalid value warns once and uses `fast`. | Experimental |
+
+### Server storage
+
+Storage names use Pydantic's nested-settings delimiter (`__`). They are read
+when `ServerSettings` is initialized, normally during server import/startup.
+Invalid integer values raise a Pydantic validation error.
+
+| Name | Type and default | Meaning | Lifecycle |
+| --- | --- | --- | --- |
+| `VLLM_OMNI_SERVER_STORAGE__PATH` | Path; default `/tmp/storage` | Directory for completed server files. | Stable |
+| `VLLM_OMNI_SERVER_STORAGE__FILE_CONCURRENCY` | Integer; default `4` | Maximum concurrent file operations. No additional range validation is currently applied. | Stable |
+| `VLLM_OMNI_SERVER_STORAGE__FILE_TTL` | Integer seconds or unset; default unset | Optional lifetime for locally stored files. | Stable |
+| `VLLM_OMNI_SERVER_STORAGE__TTL_SWEEP_INTERVAL` | Integer seconds or unset | Sweep frequency. If file TTL is set and this is unset, the effective default is `300`. | Stable |
+| `VLLM_OMNI_STORAGE_PATH` | Path | Deprecated alias for `VLLM_OMNI_SERVER_STORAGE__PATH`. The new name wins when both are set. | Deprecated |
+| `VLLM_OMNI_STORAGE_MAX_CONCURRENCY` | Integer | Deprecated alias for `VLLM_OMNI_SERVER_STORAGE__FILE_CONCURRENCY`. The new name wins when both are set. | Deprecated |
+
+The deprecated aliases emit `DeprecationWarning`. They do not affect file TTL
+settings.
+
+### Quantization diagnostics and performance
+
+| Name | Type and default | Applies to and read time | Precedence and invalid values | Lifecycle |
+| --- | --- | --- | --- | --- |
+| `VLLM_OMNI_SKIP_NVFP4_NAN_CLAMP` | Boolean truthy spellings: `1`, `true`, `yes`, `on`; default false | ModelOpt NVFP4 compatibility patch; read when `vllm_omni.patch` imports | Environment-only escape hatch. Any other value means false. Set only to diagnose the upstream NaN-scale issue. | Diagnostic and temporary |
+| `VLLM_OMNI_USE_QUACK_FP8` | Boolean truthy spellings: `1`, `true`, `yes`, `on`; unset means hardware auto-detection | FP8 scaled matrix multiplication; evaluated when quack capability is selected | A set value overrides auto-detection. Any non-truthy value forces quack off. If quack cannot load, vLLM-Omni warns and falls back to FlashInfer. | Experimental performance control |
+
+`QUACK_CACHE_DIR` is owned by the external quack library and is not an
+Omni-owned variable, even though vLLM-Omni supplies a persistent default when
+it is unset.
+
+## Per-stage environment
+
+Deploy configurations can set arbitrary environment keys for one stage:
+
+```yaml
+stages:
+  - stage_id: 0
+    env:
+      HCCL_IF_BASE_PORT: 23000
+      MY_MODEL_CACHE: /mnt/model-cache
+```
+
+The legacy nested form under `runtime.env` is also accepted. vLLM-Omni converts
+keys and values to strings, applies them while the stage engine or child process
+is launched, logs keys but not values, and restores the parent process's prior
+environment afterward. A child process inherits the scoped values. This keeps
+sibling stages from permanently changing one another's parent environment.
+
+This scoping has two consequences:
+
+1. A setting read inside stage/model initialization can see the stage value.
+2. A setting already cached by an import that occurred before stage launch
+   cannot be changed retroactively with stage `env`.
+
+Stage `env` is an escape hatch for process and third-party settings, not a
+replacement for typed stage configuration. Do not put request-varying behavior
+there. Secret values may reach child processes; logs and diagnostics must emit
+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)
+for their definitions. This includes vLLM launch, cache, logging, plugin, ROCm,
+XPU, ModelScope, and FlashInfer workspace settings.
+
+An inherited name does not become Omni-owned because Omni reads or forwards it.
+Where behavior differs, the Omni-specific page should document only that
+difference. For example, device-visibility values can be further narrowed by a
+stage's `devices` setting.
+
+## Platform, external, and secret variables
+
+Hardware runtimes, distributed launchers, and third-party libraries retain
+ownership of their variables. Common groups include:
+
+- device visibility (`ASCEND_RT_VISIBLE_DEVICES`, `HIP_VISIBLE_DEVICES`, and
+  `MUSA_VISIBLE_DEVICES`);
+- distributed launch (`MASTER_ADDR`, `MASTER_PORT`, `RANK`, and `WORLD_SIZE`);
+- Hugging Face cache and authentication (`HF_HOME`, `HF_TOKEN`, and related
+  names);
+- accelerator libraries such as NCCL, FlashInfer, MindIE, MORI, and quack.
+
+Consult the owning project's documentation for accepted values and support.
+`HF_TOKEN`, `HUGGINGFACE_HUB_TOKEN`, and `OPENAI_API_KEY` are explicitly marked
+for redaction in the reviewed inventory. Never print their values in environment
+collection, examples, bug reports, or logs.
+
+## Model-specific variables
+
+The audit found 56 variables read by a single model or pipeline family. They are
+not listed as public usage options here because doing so would turn implementation
+escape hatches into an accidental compatibility contract.
+
+Every audited model-specific name has a migration disposition in the
+[environment-variable inventory](gh-file:vllm_omni/config/environment_variable_inventory.py):
+
+| Disposition | Count | Required outcome |
+| --- | ---: | --- |
+| Promote | 33 | Move a stable setting into typed stage or model configuration. |
+| Request scope | 6 | Move request-varying behavior into a declared request-option schema. |
+| External | 0 | Retain only when a supported third-party library owns the contract. |
+| Internalize | 11 | Keep a debug or diagnostic switch out of public documentation and configuration. |
+| Deprecate/remove | 6 | Remove a compatibility escape hatch that has no continuing contract. |
+
+The disposition is a migration target, not a statement that the existing
+environment switch is stable. Promote or request-scope work should land in
+model-owner-reviewed follow-up changes with validation and model-page updates.
+The family-sized migration packages and their ownership gate are tracked in
+[#6232](https://github.com/vllm-project/vllm-omni/issues/6232).
+
+## Environment collection and troubleshooting
+
+Run the repository collector when preparing a bug report:
+
+```bash
+python collect_env.py
+```
+
+The environment section includes the safe public Omni allowlist, registered
+vLLM variables, and selected platform prefixes. It omits names marked for
+redaction and names containing common secret terms. Review the output before
+sharing it. A key supplied through stage `env` is not collected merely because
+it appears in deploy configuration; the same allowlist and prefix rules apply.
+
+If a setting appears ineffective, first check its read time in the tables
+above. Restart processes after changing import- or startup-time values, verify
+that the variable reaches the intended worker, and check whether an explicit
+typed configuration field takes precedence.
+
+## Benchmark and internal variables
+
+The reviewed inventory marks 20 benchmark variables as transitional while
+benchmark configuration moves to CLI/request schemas in
+[issue #5376](https://github.com/vllm-project/vllm-omni/issues/5376). They are not
+part of the serving/runtime contract.
+
+Internal process-coordination and debug variables are deliberately excluded
+from this page. The CI scanner resolves string literals and module-level string
+constants used by direct `os` access, local wrapper helpers, aliased `os`
+imports, and environment membership checks. Server-storage names generated by
+Pydantic settings are checked separately against their model fields.
+
+Names assembled from cross-module metadata and arbitrary keys supplied through
+stage `env` cannot be inferred reliably. Contributors must classify those names
+explicitly. CI fails when a statically resolved name has no classification or
+when a reviewed public name is absent from this page.

docs/configuration/stage_configs.md

@@ -14,7 +14,7 @@ describe fixed model topology and are not accepted in deploy YAMLs.
 Common `PipelineConfig` fields include:
 
 | Field | Description |
-|-------|-------------|
+| ------- | ------------- |
 | `model_type` | Pipeline identifier used during model and config resolution. |
 | `default_deploy_config_name` | Bundled deploy YAML loaded when the user does not pass `deploy_config`. |
 | `model_arch` | Default Hugging Face architecture for the pipeline. |
@@ -26,7 +26,7 @@ Common `PipelineConfig` fields include:
 Common `StagePipelineConfig` fields include:
 
 | Field | Description |
-|-------|-------------|
+| ------- | ------------- |
 | `stage_id` | Stable stage identifier. |
 | `model_stage` | Logical stage name used by runtime and strategy resolution. |
 | `execution_type` | `LLM_AR`, `LLM_GENERATION`, or `DIFFUSION`. |
@@ -49,10 +49,10 @@ deployment overrides.
 The new deploy schema lives under `vllm_omni/deploy/` and is paired with a frozen `PipelineConfig` registered by the model's `pipeline.py`. Each deploy YAML has these top-level fields:
 
 | Field | Type | Required | Default | Description |
-|-------|------|----------|---------|-------------|
+| ------- | ------ | ---------- | --------- | ------------- |
 | `base_config` | str (path) | optional | — | Overlay parent (relative or absolute). `stages:` / `platforms:` deep-merged by stage_id; other scalars overlay-wins. Intended for user-authored overlays; prod yamls stay flat. |
 | `async_chunk` | bool | optional | `true` | Enable chunked streaming between stages. Pin to `false` if the pipeline runs end-to-end. |
-| `session_mode` | str | optional | `"turn"` | Session behavior. Use `"duplex"` only with a pipeline that enables duplex control. |
+| `session_mode` | str | optional | `"turn"` | Session behavior. MiniCPM-o 4.5 deploy YAMLs set `"duplex"` so `/v1/realtime` and chat share the same profile. |
 | `active_stream_window` | int | optional | `0` | Number of active downstream stream slots; `0` preserves all-stream cycling. |
 | `duplex_session` | dict | optional | runtime defaults | Full-duplex session lifecycle, buffering, replay, and capacity limits. |
 | `connectors` | dict | optional | `null` | Named connector specs (`{name, extra}`). Referenced by each stage's `input_connectors` / `output_connectors`. See [Connector schema](#connector-schema). |
@@ -84,7 +84,7 @@ subprocess on one GPU. `ray` / `external_launcher` are not fully supported yet.
 Each entry under `stages:` accepts any `StageDeployConfig` field directly (no nested `engine_args:`). Only fields whose value legitimately varies across stages live here; pipeline-wide settings (trust_remote_code, distributed_executor_backend, dtype, quantization, prefix/chunked prefill, DP/PP sizes) are declared at the top level and applied to every stage. Unknown keys fall through to `engine_extras:` and are forwarded to the engine. Frequently used fields are listed below; the source-of-truth schema is `StageDeployConfig` in `vllm_omni/config/stage_config.py`.
 
 | Field | Type | Required | Default | Description |
-|-------|------|----------|---------|-------------|
+| ------- | ------ | ---------- | --------- | ------------- |
 | `stage_id` | int | required | — | Stage identity; matched against `PipelineConfig.stages[*].stage_id`. |
 | `max_num_seqs` | int \| null | optional | `null` | Max concurrent sequences per stage. |
 | `gpu_memory_utilization` | float \| null | optional | `null` | Per-stage total memory target; used for automatic KV-cache sizing. |
@@ -113,9 +113,9 @@ connectors:
       # Additional connector-specific options
 ```
 
-| Connector class | Use case | `extra` keys |
-|-----------------|----------|--------------|
-| `SharedMemoryConnector` | Same-host KV transfer between stages (default for bundled YAMLs). | None. All payloads use shared memory. |
+| Connector class          | Use case                                                              | `extra` keys                                                                                                      |
+|--------------------------|-----------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------|
+| `SharedMemoryConnector`  | Same-host KV transfer between stages (default for bundled YAMLs).     | None. All payloads use shared memory.                                                                             |
 | `MooncakeStoreConnector` | Cross-host KV transfer over TCP. Required for multi-node deployments. | `host`, `metadata_server`, `master`, `segment` (int bytes), `localbuf` (int bytes), `proto` (`"tcp"` / `"rdma"`). |
 
 A stage references a connector by name in its `input_connectors` / `output_connectors`:
@@ -135,7 +135,7 @@ stages:
 ### CLI flags
 
 | Flag | Description |
-|------|-------------|
+| ------ | ------------- |
 | `--deploy-config PATH` | Load a deploy YAML. **Optional** — when omitted, the bundled `vllm_omni/deploy/<model_type>.yaml` is auto-loaded by the model registry. |
 | `--stage-overrides JSON` | Per-stage JSON overrides, e.g. `'{"0":{"gpu_memory_utilization":0.5}}'`. Per-stage values always win over global flags. |
 | `--async-chunk` / `--no-async-chunk` | Flip the deploy YAML's `async_chunk:` bool. Unset (default) leaves the YAML value in force. |
@@ -249,7 +249,7 @@ CUDA_VISIBLE_DEVICES=0 vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --omni \
 Effective config per stage after the merge:
 
 | Stage | Field | Final value | Source |
-|-------|-------|-------------|--------|
+| ------- | ------- | ------------- | -------- |
 | 0 | `gpu_memory_utilization` | `0.9` | base YAML (overlay didn't touch stage 0) |
 | 0 | `max_num_seqs` | `8` | per-stage CLI (`--stage-overrides`) — wins over base `32` |
 | 0 | `max_model_len` | `16384` | global CLI |
@@ -273,12 +273,14 @@ when loading a deploy configuration.
 Examples:
 
 For offline inference (assuming the necessary dependencies have been imported):
+
 ```python
 model_name = "Qwen/Qwen2.5-Omni-7B"
 omni = Omni(model=model_name, deploy_config="/path/to/deploy_config.yaml")
 ```
 
 For online serving:
+
 ```bash
 vllm serve Qwen/Qwen2.5-Omni-7B --omni --port 8091 --deploy-config /path/to/deploy_config.yaml
 ```

docs/contributing/ci/test_examples/l4_functionality_tests.inc.md

@@ -1,8 +1,8 @@
-**Scope**
+#### Scope
 
 For diffusion models, the L4 functionality test covers all or common *diffusion features* that are supported by this model, including several [parallelism acceleration methods](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/parallelism_acceleration/), [CPU offloading](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/cpu_offload/), [TeaCache](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/teacache/) and [Cache-DiT](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/diffusion/cache_dit_acceleration/) cache backends, [quantization methods](https://docs.vllm.ai/projects/vllm-omni/en/latest/user_guide/quantization/overview/).
 
-**Test Case Design**
+#### Test Case Design
 
 For a *high priority* model (currently listed in [issue #1832](https://github.com/vllm-project/vllm-omni/issues/1832)), we use several test cases, each with multiple features turned on, so that each supported feature is tested in at least one test case. This is to relieve the GPU workload on the CI machine. The suggested test case combination is as follows:
 
@@ -34,12 +34,12 @@ For a *normal priority* model, further reduce the number of test cases.
 
 Currently all the features are available in online serving mode. Hence, only need to add `tests/e2e/online_serving/test_{model}_expansion.py`.
 
-**Code Style**
+#### Code Style
 
-- Validation: test that the multimodal output files of your model have the correct shapes. `OpenAIClientHandler.send_diffusion_request` should have taken care of this.
+- Validation: test that the multimodal output files of your model have the correct shapes. `OnlineOmniClient.send_diffusion_request` should have taken care of this.
 - Test marks: always add `full_model` and `diffusion` for L4 nightly `test_*_expansion.py` cases. Add GPU-related marks if needed. Ref: [Markers for Tests](https://docs.vllm.ai/projects/vllm-omni/en/latest/contributing/ci/test_writing_guide/#markers-for-tests).
 - To maximize code reuse, you may refer to
-    - `tests/conftest.py` for `omni_server` (running server in subprocess) and `openai_client` fixtures (sending requests and validating output), `generate_synthetic_image` and `assert_XXX_valid` helper.
+    - `tests/conftest.py` for `omni_server` (running server in subprocess) and `online_client` fixtures (sending requests and validating output), `generate_synthetic_image` and `assert_XXX_valid` helper.
     - `tests/helpers/mark.py` for `@hardware_test(...)` and `hardware_marks`.
     - [Parametrizing tests (pytest doc)](https://docs.pytest.org/en/stable/example/parametrize.html) to reuse test function implementation for different cases.
 - Doc: add a concise docstring for each test function.

docs/contributing/ci/test_examples/l4_performance_tests.inc.md

@@ -4,8 +4,8 @@ When you want to add L4-level ***performance test*** cases, add entries to JSON
 
 | Model type | Runner | Example JSON files |
 | ---------- | ------ | ------------------ |
-| Omni (nightly) | `run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json` (`full_model` in `mark`) |
-| Omni (weekly) | `run_benchmark.py` | `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` (`slow` in `mark`; **Perf Test** in `test-weekly.yml`) |
+| Omni (nightly) | `run_benchmark.py` | `test_qwen3_omni_no_async_chunk.json`, `test_qwen3_omni_async_chunk.json` (`full_model` without `slow` in `mark`) |
+| Omni (weekly) | `run_benchmark.py` | `test_qwen3_omni_async_chunk.json` (CUDA only), `test_qwen3_omni_vllm_text.json`, `test_qwen3_omni_multi_replicas.json` (`slow` in `mark`; **Perf Test** in `test-weekly.yml`) |
 | TTS | `run_benchmark.py` | `test_tts.json`, `test_voxcpm2.json`, `test_higgs_audio_v3.json` |
 | Diffusion | `run_diffusion_benchmark.py` | `test_qwen_image_vllm_omni.json`, `test_bagel_vllm_omni.json`, `test_wan22_i2v_vllm_omni.json`, `test_cosmos3_vllm_omni.json`, … |
 
@@ -20,7 +20,7 @@ Diffusion cases are detected when the JSON has `server_type` (typically `"vllm-o
 
 #### Running perf cases
 
-Pass **`--test-config-file`** to run one JSON file as-is, or omit it for the bulk scan above and filter with pytest **`-m`** on each case's JSON `mark` (for example `pytest … run_benchmark.py -m "full_model and tts and H100"` or `-m "slow and omni and H100"` for weekly Omni configs).
+Pass **`--test-config-file`** to load one JSON file, or omit it for the bulk scan above. In either mode, pytest **`-m`** filters each case using its JSON `mark`. A single file can contain nightly, weekly, and platform-specific case objects. For example, the nightly object in `test_qwen3_omni_async_chunk.json` declares both H100 and A3, while its `slow` weekly object is CUDA-only; there is no NPU weekly counterpart. Use `-m "H100 and full_model and not slow"`, `-m "H100 and full_model and slow"`, or `-m "npu"` to select the intended schedule and platform.
 
 ```JSON
 {
@@ -75,7 +75,7 @@ Omit `mark` only for configs not meant to be filtered by `-m`. `server_type` is
 
 Optional top-level field on each perf JSON **case object** (one per `test_name`). `run_benchmark.py` and `run_diffusion_benchmark.py` read it via `tests.dfx.conftest.resolve_pytest_marks` and attach the marks to the corresponding `pytest.param`, so you can filter locally with `-m`.
 
-When `mark` is present, it must be an **array** with exactly one ``hardware_marks`` object (same shape as `@hardware_test` / `hardware_marks()` in `tests/helpers/mark.py`), followed by registered pytest marker name strings such as `full_model`, `omni`, `tts`, `diffusion`, or `local_model`.
+When `mark` is present, it must be an **array** with exactly one ``hardware_marks`` object (same shape as `@hardware_test` / `hardware_marks()` in `tests/helpers/mark.py`), followed by registered pytest marker name strings such as `full_model`, `slow`, `omni`, `tts`, `diffusion`, or `local_model`.
 
 Supported form:
 
@@ -142,11 +142,18 @@ Examples:
 pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py -m "full_model and H100 and diffusion"
 pytest -s -v tests/dfx/perf/scripts/run_benchmark.py -m "full_model and omni and H100"
 
-# Single file (same as nightly CI Perf steps)
+# Single file (same selectors as the CI Perf steps)
 pytest -s -v tests/dfx/perf/scripts/run_diffusion_benchmark.py \
   --test-config-file tests/dfx/perf/tests/test_bagel_vllm_omni.json
 pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
-  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json
+  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json \
+  -m "H100 and full_model and not slow"
+pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
+  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json \
+  -m "H100 and full_model and slow"
+pytest -s -v tests/dfx/perf/scripts/run_benchmark.py \
+  --test-config-file tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json \
+  -m "npu"
 ```
 
 <!-- markdownlint-disable-next-line MD051 -->

docs/design/feature/attention_backend_selection.md

@@ -61,7 +61,8 @@ unstructured keyword dictionaries:
 - `quant` is consumed by FlashInfer and TRTLLM with backend-specific value
   validation;
 - `skip_softmax` is consumed by TRTLLM; and
-- `block_sparse` is consumed by block-sparse backends such as RainFusion.
+- `block_sparse` is consumed by block-sparse backends such as RainFusion; and
+- `fastvideo_vsa_topk` is consumed by FastVideo VSA.
 
 A backend reads only the fields it owns and rejects incompatible values. New
 options should be added to a shared typed spec only when more than one backend
@@ -87,3 +88,19 @@ An implementation change is complete only when it:
    corresponding user guide; and
 6. updates the overview matrix without moving algorithm details back into the
    landing page.
+
+## FastVideo VSA model metadata
+
+FastVideo VSA support is scoped to `FastVideo/FastWan2.2-TI2V-5B-Diffusers`, whose text-to-video and image-to-video modes both use `Wan22Pipeline`. It operates on a flattened DiT sequence but partitions tokens in
+the original latent video grid. Wan integrations therefore attach the
+post-patch `(T, H, W)` shape as `vsa_dit_seq_shape` attention metadata. The
+backend validates that `T * H * W` equals the sequence length, derives the
+runtime block count, and selects the configured top-k key/value blocks for
+each query block.
+
+Checkpoint-specific compensation remains model-owned. Wan layers pass an
+optional learned `gate_compress` projection through attention metadata when
+the checkpoint contains those weights. Checkpoints without the projection do
+not allocate or execute it. When top-k selects every block, native checkpoints
+route to SDPA, while FastVideo DMD checkpoints preserve the VSA all-block path
+to retain checkpoint semantics.

docs/design/feature/cfg_parallel.md

@@ -63,7 +63,7 @@ vLLM-omni provides `CFGParallelMixin` that encapsulates all CFG parallel logic.
 
 ### N-Branch CFG (3+ branches)
 
-Some models require more than 2 CFG branches. For example, Bagel and OmniGen2 use 3 branches, DreamID Omni uses 4 branches.
+Some models require more than 2 CFG branches. For example, Bagel and OmniGen2 use 3 branches.
 
 `predict_noise_with_multi_branch_cfg()` handles these by automatically dispatching N branches across M GPUs using round-robin (rule: branch `i` → rank `i % M`):
 

docs/design/feature/host_weight_runtime.md

@@ -14,12 +14,12 @@ specified in the [Host Weight Runtime module design](../module/host_weight_runti
 
 ## Status
 
-The first implementation provides contracts and a CPU local-filesystem store.
-The initial diffusion consumer contract additionally defines typed,
+The implementation provides contracts and a CPU local-filesystem store. The
+diffusion consumer additionally defines typed,
 representation-independent final-layout identity/restoration mechanics plus a
-concrete BF16-with-preserved-FP32 policy for MiniMax H3. It remains
-library-only: no loader or DLO path selects, publishes, restores, or transports
-that artifact yet.
+concrete BF16-with-preserved-FP32 policy for MiniMax H3 and
+`black-forest-labs/FLUX.2-klein-4B`. The opt-in no-AllGather DLO integration
+selects, publishes, restores, and transfers these artifacts.
 
 V1 includes:
 
@@ -33,13 +33,12 @@ V1 includes:
 
 V1 does not include:
 
-- a public CLI, loader activation, or default-on model integration;
-- lease handoff to DLO or another transport consumer;
-- FP8, quantized, merged-adaptation, or additional model producers;
-- CUDA registration, pinned staging, H2D scheduling, or GPU kernels;
+- default-on activation or consumers outside no-AllGather DLO;
+- online FP8, quantized, merged-adaptation, or additional model producers;
+- HWR interaction with DLO AllGather;
 - a remote artifact provider or cross-node coordination;
 - automatic eviction; or
-- a change to DLO AllGather or no-AllGather behavior.
+- a change to DLO collective or execution behavior.
 
 ## Motivation and use cases
 
@@ -255,7 +254,8 @@ The contract is intentionally separate from loader activation:
 - `FinalLayoutBF16Producer` accepts only the matching identity context and a
   finalized CPU model. It is `POST_LOAD_ONLY` and `SINGLE_PROCESS` per exact TP
   coordinate. Its BF16 policy preserves model-declared FP32 parameters and
-  buffers and revalidates MiniMax H3 mixed-precision invariants.
+  buffers, revalidates MiniMax H3 mixed-precision invariants, and revalidates
+  FLUX.2-klein's two block stacks, packed QKV mapping, and BF16 base layout.
 
 Other representations reuse source identity, typed parallel identity, tensor
 ownership, and exact restoration only when their policy proves those semantics.

docs/design/feature/offloader/distributed_layerwise_offload.md

@@ -20,7 +20,7 @@ implemented but has less end-to-end coverage than the primary path.
 Legend: ✅ supported, ⚠️ compatibility path or limited validation, ❌ unsupported.
 
 | Feature | DLO + AllGather | DLO without AllGather |
-|---|---|---|
+| --- | --- | --- |
 | **DP** | ✅ Primary path; host weights are sharded across the DP group. | ✅ Each DP rank streams complete rank-local blocks. |
 | **SP** | ✅ When DP=1, DLO uses the SP group for weight sharding. | ✅ SP remains active without a DLO weight collective. |
 | **TP > 1** | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. | ⚠️ Ordinary TP-aware loader only; no direct checkpoint mmap. |
@@ -169,10 +169,242 @@ This mode means:
   copy per rank.
 - The scheduler does not require a synchronized DP request wave for DLO.
 
+### Component allocator-cache retention
+
+MiniMax-H3 DLO keeps one bounded PyTorch allocator cache across its staged
+encoder, DiT, and VAE component boundaries so allocations released by one
+component can be reused by the next. The DiT prefetch path and its two shared
+device buffers are unchanged.
+
+After a component is offloaded, cached-but-unallocated memory is retained only
+while it is at most 25% of device capacity and at least 5% of device capacity
+remains physically free. Crossing either bound releases the allocator cache.
+Missing allocator telemetry also releases it conservatively. Component or
+staging failure forces a release, and an out-of-memory allocation gets one
+retry after release. This is a component-local policy rather than a global
+`empty_cache` override, so the executor's unconditional shutdown cleanup is
+preserved.
+
+Retaining allocator blocks trades higher reserved and peak device memory for
+lower lifecycle latency. The benefit is device- and workload-dependent, and a
+smaller or externally loaded device may cross a bound and return to the
+ordinary per-component cache-release behavior. Current performance validation
+covers MiniMax-H3 no-AllGather DLO on CUDA. Other MiniMax-H3 DLO transfer
+topologies and platforms receive the same bounded policy but are not claimed
+performance targets; other DLO models remain unchanged unless they explicitly
+adopt it.
+
+### Final-layout Host Weight Runtime consumer
+
+> **Status:** PR2 landed in
+> [PR #6486](https://github.com/vllm-project/vllm-omni/pull/6486). PR3 adds
+> registered direct H2D plus generic HWR publication/source-digest
+> optimizations without changing AllGather behavior.
+
+Final-layout Host Weight Runtime (HWR) backing is opt-in and currently applies
+only to no-AllGather DLO. Enable it with
+`--host-weight-runtime-mode preferred` (or `required`) and
+`--host-weight-runtime-root <node-local-root>`.
+
+The modes express operator fallback policy, not different artifact formats:
+
+- `preferred` consumes an exact hit, but on a miss it allows canonical loading
+  followed by post-load publication. It is the normal population path.
+- `required` consumes the same exact artifact but fails startup on a miss or
+  unusable artifact. It never invokes canonical DiT fallback or post-load
+  publication and therefore cannot populate an empty store.
+
+PR2 has no separate prewarm command. Operators populate one matching producer
+cohort per node-local storage domain in `preferred` mode, then restart the same
+model revision and parallel layout with `required`. TP coordinates have distinct
+identities, while equivalent DP replicas share them.
+
+The current producer/consumer boundary is the model-declared final-layout BF16
+contract (MiniMax H3 and `black-forest-labs/FLUX.2-klein-4B`). The FLUX.2-klein
+contract covers both transformer block stacks, constructor-stable packed QKV
+mapping state, BF16 parameters, and any persistent loader-owned `beta`/`eps`
+buffers. The shared HWR machinery supports ordinary-loader final layouts for
+TP1 and TP2 rank identities plus SP layout identities. FLUX.2-klein evidence
+in this change is TP1-only; its TP2/SP layouts rely on the existing
+per-coordinate identity mechanics and remain unmeasured. Online quantization,
+HSDP, LoRA/adapted weights, and non-default load formats remain ineligible.
+HWR mode
+`disabled`, DLO-disabled, and DLO AllGather configurations stop before source
+identity or store construction and retain the existing checkpoint-mmap or
+ordinary-loader path.
+
+Eligibility is decided before constructing HWR, resolving canonical sources,
+hashing identity inputs, probing a filesystem, or emitting an HWR observer
+event:
+
+```text
+DLO disabled / HWR disabled / DLO AllGather
+  -> zero final-layout HWR interaction
+  -> preserve current checkpoint-mmap or ordinary-loader behavior
+```
+
+When final-layout HWR is explicitly enabled for no-AllGather DLO, runtime
+outcomes are authoritative:
+
+- `LOCAL_HIT`: plan and commit an exact restoration, then transfer the lease
+  transactionally to DLO.
+- `CANONICAL_FALLBACK`: bypass checkpoint mmap, canonically materialize and
+  finalize the DiT, then call `publish_after_load()` for a future startup. The
+  current startup retains its canonical tensors.
+- `FAILED`: fail startup. Preferred mode does not reinterpret nonretryable
+  identity, configuration, or compatibility failures as misses.
+- Required mode cannot bootstrap an empty store through a `POST_LOAD_ONLY`
+  producer.
+
+Checkpoint mmap remains an unchanged control path whenever final-layout HWR is
+not selected.
+
+Local source identity and publication avoid repeated full-payload passes while
+preserving the same semantic contract:
+
+- Canonical files without a trusted immutable Hub blob identity are hashed in
+  parallel. Their content IDs are cached per storage domain under the HWR root
+  and reused only when the path, inode, size, timestamps, symlink target, and
+  cache-record checksum still match. Corrupt entries are rebuilt; cache-lock or
+  cache-I/O failure falls back to hashing the canonical file directly.
+- Producers that emit payload bytes in canonical storage-key order may declare
+  `ordered=True`. The filesystem writer then computes file and tensor SHA256
+  values during the write and overlaps payload `fsync` with later producer
+  work. Producers without that guarantee retain unordered writes and use a
+  parallel readback checksum fallback.
+- Manifest and `READY` publication still wait for every payload `fsync` and
+  checksum to complete. These optimizations do not weaken artifact validation
+  or change lease ownership.
+
+#### Registered mmap transport
+
+After DLO takes an HWR lease, the no-AllGather backend may register the lease's
+complete immutable mapped ranges under the existing `pin_cpu_memory` policy.
+Registration is transport state: it does not change artifact identity, the
+store, tensor ownership, or H2D payload. On success, each tensor view copies
+directly into the existing rotating HBM block buffers and the two private host
+staging slots are not allocated.
+
+`--dlo-host-registration-limit-gib` is an optional per-worker preflight ceiling
+over page-aligned registered bytes. Zero adds no ceiling. A disabled pinned
+memory policy, unsupported platform/capability, over-budget mapping, or fully
+rolled-back registration error selects the existing two-slot staging path. A
+partial registration that cannot be rolled back aborts startup because closing
+the lease would unmap memory still owned by the platform.
+
+Direct checkpoint mmap remains unchanged and continues to use staging. It may
+require loader-owned per-block transforms, while the HWR artifact already
+contains final runtime bytes. DLO AllGather never receives an HWR final-layout
+lease and therefore never enters this registration path.
+
+#### Pre-service transaction boundary
+
+The startup transaction does not end at restore commit. A lease-backed model is
+disposable until backend setup and initial prefetch complete:
+
+```text
+UNRESOLVED
+  -> LEASE_OWNED_BY_LOADER
+  -> RESTORE_PLANNED
+  -> COMMIT_STARTED
+  -> CARRIER_OWNED
+  -> BACKEND_OWNED
+  -> BACKEND_READY
+  -> IN_SERVICE
+```
+
+Failure before `COMMIT_STARTED` closes the loader-owned lease. Preferred mode
+may canonically load using the untouched skeleton; required mode fails.
+
+Failure from `COMMIT_STARTED` through `BACKEND_READY` must:
+
+1. synchronize or otherwise quiesce partial backend and initial-prefetch work;
+2. release hook, staging, restore-plan, and model references;
+3. close the lease through its current owner;
+4. discard the restored model; and
+5. in preferred mode, construct a fresh canonical model while bypassing HWR
+   lookup, HWR publication, and checkpoint mmap for that recovery attempt.
+
+Required mode fails instead of constructing the fresh fallback. Once the model
+enters service, failures follow normal runtime handling rather than startup
+fallback.
+
+#### Finalization phases
+
+Cold canonical loading separates byte-changing work from shared runtime-state
+finalization:
+
+```text
+Cold-only byte-changing:
+  casting, reordering, packing, quantization, generated scales,
+  model-specific weight transforms, and calibration that mutates
+  parameters or persistent buffers
+
+Shared cold/warm non-byte:
+  validation, hook installation, eval state, bookkeeping,
+  and non-persistent runtime state
+```
+
+The warm path must snapshot restored tensor bytes and backing pointers before
+shared finalization and prove both are unchanged afterward.
+
+#### Lease ownership
+
+Lease transfer is single-owner and single-take:
+
+```text
+loader owns
+  -> restore plan borrows
+  -> commit succeeds
+  -> runner carrier takes once
+  -> backend takes once before asynchronous work
+  -> backend drains pending H2D work
+  -> backend releases hooks, staging, and model references
+  -> backend unregisters every mapped range
+  -> backend closes lease
+```
+
+The runner carrier is process-local, rejects serialization, and rejects a
+duplicate `take()`. If the carrier still owns the lease, runner cleanup closes
+it. Once the backend takes ownership, only backend abort or teardown may drain
+work and close it.
+
+The implementation keeps the final-layout lease through backend setup and
+initial prefetch. Successful registration copies directly from immutable HWR
+views; otherwise the backend uses the existing two bounded rank-local staging
+slots. Transactional cleanup drains device work, removes source references,
+unregisters mappings, and only then closes the lease. An unregistration failure
+retains both registration and lease for retry/process teardown. Preferred mode
+then uses the runner's fresh canonical retry; required mode propagates the
+failure.
+
+#### PR2 and PR3 promotion gates
+
+- Warm hit performs zero ordinary DiT materialization and zero producer calls.
+- Shared warm finalization changes neither restored bytes nor backing pointers.
+- Preferred `FAILED` outcomes do not silently fall back.
+- Planning failure leaves the skeleton untouched.
+- Any pre-service failure after commit begins discards the restored model;
+  preferred mode uses a fresh canonical model and required mode fails.
+- Mixed components load normally and no required tensor remains on `meta`.
+- Disabled mode and AllGather emit zero final-layout HWR interaction.
+- The lease carrier rejects duplicate take and serialization.
+- Backend setup or prefetch failure drains asynchronous work before lease close.
+- Registration is all-or-nothing; rollback/unregistration failure never closes
+  a mapping still owned by the platform.
+- Registered HWR transport bypasses host staging while preserving output and
+  H2D payload; unsupported registration retains the bounded staging path.
+- Compatible checkpoint mmap remains an unchanged benchmark and control path.
+- Prewarm uses one matching producer cohort per TP/SP topology and storage
+  domain.
+- Benchmarks compare ordinary loading, cold publication, and warm staging with
+  output parity, startup latency, aggregate PSS, page-cache state, HBM, and H2D
+  payload.
+
 ## Parallelism compatibility
 
 | Parallelism | DLO + AllGather | DLO without AllGather |
-|---|---|---|
+| --- | --- | --- |
 | **DP** | Supported primary path. DLO shards host weights across the DP group and can run DP multi-concurrency. | Supported rank-local path. Compatible TP1 replicas can share checkpoint pages on each node; fallback runtime tensors remain private. |
 | **SP** | Supported in the implementation. With DP=1, DLO uses the SP group for host-weight sharding; SP still shards sequence/activation work. | SP remains active, but DLO keeps standard-loader rank-local weights and adds no SP weight collective. |
 | **TP > 1** | Outside the Phase A shared-mmap support scope. The loader falls back before mutation, preserves TP-local layouts, and DLO may apply DP/SP host sharding to those ordinary runtime tensors. | Outside the Phase A shared-mmap support scope. The ordinary TP-aware loader produces rank-local tensors, which DLO streams without an additional weight collective; DP replicas retain private runtime storage. |
@@ -185,7 +417,8 @@ This mode means:
   the SP group becomes DLO's sharding group in AllGather mode.
 - **DP + TP/SP without AllGather:** standard model loading defines the
   rank-local tensor layout. DLO adds no cross-DP, cross-TP, or cross-SP weight
-  collective.
+  collective. When the model declares the final-layout contract, HWR keys the
+  reusable artifact by TP rank/size and SP layout.
 - **HSDP + SP:** the general parallel configuration permits HSDP over SP, but
   DLO must use `--dlo-no-use-allgather`. HSDP remains responsible for weight
   materialization and synchronization.
@@ -211,10 +444,12 @@ reconstructs those tensors with their recorded layouts. Other online methods
 must use `--dlo-no-use-allgather` or disable online quantization until their
 runtime layouts are validated.
 
-A normalized runtime mmap cache, built through the ordinary loader, is the
-proposed general mechanism for sharing transformed TP or quantized layouts.
-That cache and its publication/lifecycle protocol are intentionally outside
-this phase; see [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195).
+The Host Weight Runtime representation, publication, and no-AllGather consumer
+contracts are merged; see
+[RFC #6414](https://github.com/vllm-project/vllm-omni/issues/6414),
+[PR #6445](https://github.com/vllm-project/vllm-omni/pull/6445), and
+[PR #6486](https://github.com/vllm-project/vllm-omni/pull/6486). Registered
+direct H2D is an optional transport layer over that merged lease contract.
 
 ## Validation coverage
 
@@ -227,10 +462,16 @@ Current source-level validation includes:
 - ordinary-loader fallback for per-tensor online FP8 linears followed by DLO
   sharding of finalized weights and scales;
 - exact loader-to-backend plan transfer and ordinary-loader fallback;
+- ordered publication hashing, overlapped durability, unordered parallel
+  checksum fallback, and node-local source-digest reuse/invalidation;
 - rank-local mmap source retention, bounded two-slot staging, and adapter
   transforms without parameter-side flags;
+- read-only registration preflight, direct source-to-device copies, safe
+  fallback, partial rollback, retryable unregistration, and lease ordering;
 - resident-layer requests requiring no-AllGather;
 - DP request-wave validation for denoising-step compatibility;
+- bounded component allocator-cache retention, conservative/forced release,
+  OOM retry, and immutable encoder non-block staging;
 - sharding, double-buffer, AllGather-size, and heterogeneous-block regression
   tests.
 
@@ -243,7 +484,7 @@ VAEs at TP1. They validate the ordinary-loader fallback only, not direct mmap
 or shared-mmap host-memory savings.
 
 | Configuration | Result | Warm E2E | Peak device memory | Host PSS |
-|---|---:|---:|---:|---:|
+| --- | ---: | ---: | ---: | ---: |
 | DP4xTP1 AllGather | Passed, 4 concurrent requests | 2.87 s / 4 requests | 13.84 GiB | 211.99 GiB |
 | DP4xTP1 no-AllGather | Passed, 1 request | 15.02 s | 13.23 GiB | 187.77 GiB |
 | DP2xTP2 AllGather | Passed, 2 concurrent requests | 4.16 s / 2 requests | 12.50 GiB | 211.97 GiB |
@@ -275,7 +516,7 @@ pipeline components, so each worker should be compared with the same worker in
 the other storage mode.
 
 | Worker | Ordinary RSS | mmap RSS | Ordinary PSS | mmap PSS | PSS reduction |
-|---|---:|---:|---:|---:|---:|
+| --- | ---: | ---: | ---: | ---: | ---: |
 | DP worker 0 | 168.27 GiB | 132.76 GiB | 167.84 GiB | 101.43 GiB | 66.40 GiB |
 | DP worker 1 | 116.19 GiB | 79.97 GiB | 115.73 GiB | 48.64 GiB | 67.09 GiB |
 | **Two-worker total** | — | — | **283.56 GiB** | **150.08 GiB** | **133.48 GiB (47.1%)** |

docs/design/metrics.md

@@ -6,7 +6,7 @@ This document describes how vLLM-Omni exposes Prometheus metrics for multi-stage
 
 - Expose pipeline-level request and latency metrics that span the full multi-stage execution (orchestrator scope).
 - Preserve all upstream vLLM per-engine metrics (`vllm:*`) for stages backed by an AR LLM engine, and reshape their `engine` label into `stage` + `replica` so multi-replica deployments gain per-replica visibility automatically.
-- Expose per-modality SLO metrics that the upstream `vllm:*` families do not capture — audio TTFP / RTF / duration / frames / streaming continuity / silent-loss.
+- Expose per-modality SLO metrics that the upstream `vllm:*` families do not capture — audio TTFP / RTF / duration / frames / streaming continuity / silent-loss, plus image and diffusion workload, queueing, execution, and memory signals.
 - Expose per-replica-edge cross-stage transfer metrics so the slack between E2E latency and the sum of per-stage `gen_time` (queueing, serialization, network) becomes attributable.
 - Keep the metrics collection overhead low enough that it does not regress TTFA or throughput.
 
@@ -20,7 +20,7 @@ vLLM's `unregister_vllm_metrics()` function strips every `prometheus_client` col
 
 ### The Problem
 
-vLLM-Omni runs multiple engine instances (stages × replicas) within a single process, coordinated by an Orchestrator. The pipeline needs its own metrics — aggregate request counts, end-to-end latency across all stages, per-modality SLO signals, and cross-stage transfer attribution — that do not exist in upstream vLLM. All pipeline-level metrics use the `vllm:omni_` prefix to distinguish them from upstream per-engine metrics. At import time (see `vllm_omni/patch.py`) we replace `unregister_vllm_metrics()` with a scoped version that still strips upstream `vllm:*` collectors before each new `PrometheusStatLogger` registers (so multi-engine processes don't crash on duplicate registration), but preserves anything prefixed `vllm:omni_` / `vllm_omni`.
+vLLM-Omni runs multiple engine instances (stages × replicas) within a single process, coordinated by an Orchestrator. The pipeline needs its own metrics — aggregate request counts, end-to-end latency across all stages, per-modality SLO signals, and cross-stage transfer attribution — that do not exist in upstream vLLM. All pipeline-level metrics use the `vllm_omni:` prefix to distinguish them from upstream per-engine metrics. At import time (see `vllm_omni/patch.py`) we replace `unregister_vllm_metrics()` with a scoped version that still strips upstream `vllm:*` collectors before each new `PrometheusStatLogger` registers (so multi-engine processes don't crash on duplicate registration), but preserves anything prefixed `vllm_omni:` / `vllm_omni`.
 
 Upstream per-engine metrics retain the `vllm:` prefix but are now registered by `OmniPrometheusStatLogger`, a thin subclass of upstream's `PrometheusStatLogger` that reshapes the single `engine` label into a `stage` + `replica` pair (see "OmniPrometheusStatLogger wrap" below).
 
@@ -38,7 +38,7 @@ Upstream per-engine metrics retain the `vllm:` prefix but are now registered by
                                    |
         +--------+--------+--------+--------+--------+
         |                                            |
-   vllm:omni_*                                    vllm:*
+   vllm_omni:*                                    vllm:*
    collectors                                  collectors
         |                                            |
    +----+--------+   +-----------+   +----------+   +-----------+
@@ -58,17 +58,30 @@ Upstream per-engine metrics retain the `vllm:` prefix but are now registered by
 
 ### Data Flow
 
-There are four independent paths for metric collection.
+There are five independent paths for metric collection.
 
-**Path 1: Pipeline-level metrics (`vllm:omni_*`)**
+**Path 1: Pipeline-level metrics (`vllm_omni:*`)**
 
 `OmniPrometheusMetrics` registers the Gauge / Counter / Histogram collectors at import time. It is instantiated once per entrypoint, labeled with the model name. The entrypoint calls its methods as requests progress:
 
 - `set_running(n)` / `set_waiting(n)` — updated after each request completes. The running count comes from `OmniRequestCounter`, a simple counter incremented/decremented by the Orchestrator as it tracks requests. Waiting is derived as `total - running`.
-- `request_succeeded(e2e_seconds, finished_reason="stop")` — recorded when a request finishes at the final stage. `finished_reason` is extracted from `engine_outputs.outputs[0].finish_reason` (vLLM `CompletionOutput` convention) and increments `vllm:omni_requests_success_total{finished_reason}`.
-- `request_failed()` — recorded by the cleanup path when a request exits without natural completion. Internally maps to `finished_reason="abort"` so a single Counter family covers both natural and aborted completion.
+- `request_succeeded(e2e_seconds, finished_reason="stop")` — recorded when a request finishes at the final stage. `finished_reason` is extracted from `engine_outputs.outputs[0].finish_reason` (vLLM `CompletionOutput` convention) and increments `vllm_omni:requests_success_total{finished_reason}`.
+- `request_failed()` — recorded by the cleanup path when a request exits without natural completion. Internally maps to `finished_reason="abort"` on the existing completion Counter.
+- `inc_requests_failed(reason)` — increments the dedicated failure-attribution Counter once per request. Reasons are normalized to the bounded set `client_abort`, `client_disconnect`, `stage_error`, and `unknown`.
 
-**Path 2: Audio modality metrics (`vllm:omni_audio_*`)**
+**Path 2: Image and diffusion metrics**
+
+Image and diffusion metrics use two cooperating collectors. `OmniPrometheusMetrics` owns pipeline/stage workload families, while `OmniModalityMetrics` owns per-replica diffusion breakdowns. Finished stage messages are consumed once in `OmniBase._process_single_result`, which prevents replayed terminal messages from incrementing Counters or Histograms twice.
+
+- `stage_gen_time_s`, `stage_in_queue_s`, image workload, inference-step, and per-stage peak-memory values are emitted from `StageRequestStats` when a stage finishes. `image_pixels` and `image_count_total` are restricted to image outputs; `num_inference_steps` is restricted to diffusion stages.
+- Diffusion engine timings are accumulated per request by `OrchestratorAggregator.accumulate_diffusion_metrics`. Engine-side millisecond values are converted to seconds before `observe_modality_at_finalize(...)` dispatches the per-replica Histogram observations.
+- `request_queue_wait_s` is emitted once at final request completion from `pipeline_timings["queue_wait_ms"]`. A present value of `0` is a valid observation; a missing key produces no sample.
+- `stage_waiting_requests` is the sum of the latest scheduler snapshots for all live replicas in a stage. AR snapshots arrive through `SchedulerStats`; diffusion snapshots ride normal diffusion output metrics. A dead replica's snapshot is removed. Because this deliberately reuses existing message paths, the gauge can retain its last value while a replica produces no output.
+- `kv_wait_s` measures the AR scheduler interval from entering the KV-transfer-free wait state until extraction acknowledgement. The sample is carried to the Orchestrator through the existing engine output and labeled by connector type. Abort/error/removal paths discard incomplete timers rather than emitting partial waits.
+
+Zero and missing values have different meanings. Valid zero-duration queue observations are recorded, while unavailable optional measurements are omitted instead of being exported as synthetic zeroes. `peak_memory_mb`, `diffusion_forward_s`, `diffusion_kv_load_s`, `vae_decode_s`, and `kv_wait_s` therefore appear only when their data source is active.
+
+**Path 3: Audio modality metrics (`vllm_omni:audio_*`)**
 
 `OmniModalityMetrics` registers seven audio families with `{model_name, stage, replica}` (plus an extra `threshold_ms` / `reason` label on the two extra-cardinality Counters). Three observation sites:
 
@@ -76,7 +89,7 @@ There are four independent paths for metric collection.
 - `observe_audio_first_packet(...)` — called from the OpenAI SSE audio branch in `serving_chat.py` on the first audio packet for a request. The once-per-request guard is held by `ClientRequestState.first_audio_ts`. The `request_arrival_ts` anchor is stored in `ClientRequestState` by `async_omni.generate()`, computed at request entry.
 - `observe_audio_streaming_finalize(...)` — called from `serving_chat.py` after the streaming chunk loop exhausts. It runs the per-chunk player simulation from `vllm_omni/benchmarks/audio_continuity.py` to compute the worst-case underrun and emits `audio_underrun_s` plus (when the request stayed below the threshold) `audio_continuity_ok_total{threshold_ms}`. Per-chunk PCM byte counts and arrival timestamps are recorded by the same audio branch that updates `first_audio_ts`.
 
-**Path 3: Cross-stage transfer metrics (`vllm:omni_transfer_*`)**
+**Path 4: Cross-stage transfer metrics (`vllm_omni:transfer_*`)**
 
 `OmniTransferMetrics` registers four Histogram families with `{model_name, from_stage, from_replica, to_stage, to_replica}` labels. Each observation corresponds to one physical transfer hop (one chunk between adjacent stages), not the per-request accumulated total — so the histograms track per-transfer distribution.
 
@@ -87,7 +100,7 @@ The hook lives in `OrchestratorAggregator.record_transfer_tx` and `record_transf
 
 Defensive fail-safe: if `transfer_emitter` or `replica_resolver` is missing, or the resolver returns `None` for either side, the emit is skipped silently (the underlying `TransferEdgeStats` accumulation is unaffected).
 
-**Path 4: Per-engine metrics (`vllm:*`, stage/replica wrap)**
+**Path 5: Per-engine metrics (`vllm:*`, stage/replica wrap)**
 
 The Orchestrator instantiates `OmniPrometheusStatLogger` (a thin subclass of upstream `vllm.v1.metrics.loggers.PrometheusStatLogger`) and feeds it scheduler stats and iteration stats after processing each batch of engine outputs. This populates the standard ~37 vLLM metric families (TTFT, ITL, TPOT, KV cache usage, etc.) using the same upstream code path — but with the `engine` label reshaped into `stage` + `replica` so multi-replica deployments produce distinct series per replica. See the next section for the wrap mechanics.
 
@@ -97,9 +110,9 @@ The Orchestrator runs in a background thread. The API server (OmniBase) runs in
 
 ### Metric Registration and Lifecycle
 
-All `vllm:omni_*` collectors are registered once when their owning class (`OmniPrometheusMetrics` / `OmniModalityMetrics` / `OmniTransferMetrics`) is imported. Per-`(stage, replica)` labels are bound lazily on first observation to avoid registering label sets for combinations that never produce data.
+All `vllm_omni:*` collectors are registered once when their owning class (`OmniPrometheusMetrics` / `OmniModalityMetrics` / `OmniTransferMetrics`) is imported. Per-`(stage, replica)` labels are bound lazily on first observation to avoid registering label sets for combinations that never produce data.
 
-The `prometheus_client` default registry holds all collectors. FastAPI's `/metrics` endpoint serves the default registry, so `vllm:omni_*` and the wrapped `vllm:*` metrics appear in the same scrape response alongside `http_*` and `process_*` metrics from the instrumentator and the Python client runtime.
+The `prometheus_client` default registry holds all collectors. FastAPI's `/metrics` endpoint serves the default registry, so `vllm_omni:*` and the wrapped `vllm:*` metrics appear in the same scrape response alongside `http_*` and `process_*` metrics from the instrumentator and the Python client runtime.
 
 ## OmniPrometheusStatLogger Wrap
 
@@ -130,11 +143,11 @@ A reverse map `(stage_id, replica_id) -> flat_idx` is maintained on the Orchestr
 
 ## Gating: `--log-stats`
 
-All metrics — both the 15 `vllm:omni_*` families and the ~65 upstream `vllm:*` wrap families — are gated by the user's `--log-stats` CLI flag (default off). The flag is plumbed from `OmniBase.__init__(log_stats=...)` through `AsyncOmniEngine` to the stage-spawn helpers and to the three Prometheus metric classes (`OmniPrometheusMetrics` / `OmniModalityMetrics` / `OmniTransferMetrics`), and is forwarded to `Orchestrator._init_metrics_state(...)`.
+All metrics — both the omni-specific `vllm_omni:*` families and the upstream `vllm:*` wrap families — are gated by the user's `--log-stats` CLI flag (default off). The flag is plumbed from `OmniBase.__init__(log_stats=...)` through `AsyncOmniEngine` to the stage-spawn helpers and to the three Prometheus metric classes (`OmniPrometheusMetrics` / `OmniModalityMetrics` / `OmniTransferMetrics`), and is forwarded to `Orchestrator._init_metrics_state(...)`.
 
 Behavior with `--log-stats=off` (default):
 
-- The three `Omni*Metrics` classes register their module-level Gauge / Counter / Histogram families at import time (prometheus_client requires up-front registration), but each `observe / inc / set` method early-returns. The per-label child series for `vllm:omni_*` stay materialized but never have data written to them.
+- The three `Omni*Metrics` classes register their module-level Gauge / Counter / Histogram families at import time (prometheus_client requires up-front registration), but each `observe / inc / set` method early-returns. The per-label child series for `vllm_omni:*` stay materialized but never have data written to them.
 - `OmniPrometheusStatLogger` is not constructed in `_init_metrics_state`, so the ~65 upstream `vllm:*` wrap families are not registered in the default registry at all.
 - The engine core's `Scheduler.make_stats()` also short-circuits inside upstream (`if not self.log_stats: return None`), so no `SchedulerStats` is produced per step — the per-iteration cost is bounded by the existing upstream gate.
 
@@ -162,10 +175,43 @@ second.
 
 | Metric | Type | Labels | Description |
 |--------|------|--------|-------------|
-| `vllm:omni_num_requests_running` | Gauge | `model_name` | Requests currently executing across all stages |
-| `vllm:omni_num_requests_waiting` | Gauge | `model_name` | Requests queued but not yet scheduled |
-| `vllm:omni_requests_success_total` | Counter | `model_name`, `finished_reason` | Total requests by completion reason ({stop, length, abort, ...}); aborts cover client-disconnect / cancellation paths in addition to upstream `FinishReason.ABORT` |
-| `vllm:omni_e2e_request_latency_s` | Histogram | `model_name` | Pipeline-global end-to-end latency in seconds |
+| `vllm_omni:num_requests_running` | Gauge | `model_name` | Requests currently executing across all stages |
+| `vllm_omni:num_requests_waiting` | Gauge | `model_name` | Requests queued but not yet scheduled |
+| `vllm_omni:requests_success_total` | Counter | `model_name`, `finished_reason` | Total requests by completion reason ({stop, length, abort, ...}); aborts cover client-disconnect / cancellation paths in addition to upstream `FinishReason.ABORT` |
+| `vllm_omni:e2e_request_latency_s` | Histogram | `model_name` | Pipeline-global end-to-end latency in seconds |
+
+### Image and diffusion service-level metrics
+
+| Metric | Type | Labels | Description |
+|--------|------|--------|-------------|
+| `vllm_omni:stage_gen_time_s` | Histogram | `model_name`, `stage`, `stage_type` | Stage submit to finished output; includes in-stage queueing |
+| `vllm_omni:request_queue_wait_s` | Histogram | `model_name` | Orchestration-layer queue wait; a present zero is recorded |
+| `vllm_omni:stage_waiting_requests` | Gauge | `model_name`, `stage` | Sum of the latest waiting snapshots across live replicas in the stage |
+| `vllm_omni:stage_in_queue_s` | Histogram | `model_name`, `stage` | Diffusion scheduler wait from enqueue to initial execution admission |
+| `vllm_omni:num_inference_steps` | Histogram | `model_name` | Denoising-step distribution for diffusion stages |
+| `vllm_omni:image_count_total` | Counter | `model_name` | Cumulative number of generated image outputs |
+| `vllm_omni:image_pixels` | Histogram | `model_name` | Pixel-count distribution for image outputs |
+| `vllm_omni:peak_memory_mb` | Gauge | `model_name`, `stage` | Positive per-stage peak-memory samples reported by engine outputs |
+| `vllm_omni:requests_failed_total` | Counter | `model_name`, `reason` | Request failures with bounded reason values (`client_abort`, `client_disconnect`, `stage_error`, `unknown`) |
+| `vllm_omni:kv_wait_s` | Histogram | `model_name`, `connector_type` | Completed AR KV-transfer wait intervals, grouped by connector backend |
+
+### Diffusion execution breakdown
+
+Labels: `{model_name, stage, replica}`.
+
+| Metric | Type | Description |
+|--------|------|-------------|
+| `vllm_omni:diffusion_exec_s` | Histogram | Core diffusion-step execution time per request |
+| `vllm_omni:diffusion_exec_per_step_s` | Histogram | Core diffusion execution divided by inference-step count |
+| `vllm_omni:diffusion_preprocess_s` | Histogram | Diffusion input preprocessing time |
+| `vllm_omni:diffusion_postprocess_s` | Histogram | Diffusion output postprocessing time |
+| `vllm_omni:vae_decode_s` | Histogram | VAE decode time when the engine reports it |
+| `vllm_omni:diffusion_forward_s` | Histogram | Forward-only denoise-loop time when pipeline profiling is enabled |
+| `vllm_omni:diffusion_kv_load_s` | Histogram | Diffusion-side KV receive/load time when upstream KV is used |
+| `vllm_omni:image_ttfp_s` | Histogram | Image-stage submit to first materialized image output |
+| `vllm_omni:denoise_step_latency_s` | Histogram | Mean diffusion forward time per denoise step for image output stages |
+
+The optional breakdown and memory families are sparse by design: when an engine does not produce the source measurement, no sample is emitted. This differs from queue timings, where an explicitly measured zero is meaningful and is retained.
 
 ### Audio (7)
 
@@ -173,13 +219,13 @@ Labels: `{model_name, stage, replica}` plus the listed extra label.
 
 | Metric | Type | Extra label | Description |
 |--------|------|-------------|-------------|
-| `vllm:omni_audio_ttfp_s` | Histogram | — | Time from request arrival to first audio packet/frame |
-| `vllm:omni_audio_duration_s` | Histogram | — | Audio content duration (`audio_frames / sample_rate`) |
-| `vllm:omni_audio_rtf` | Histogram | — | Real-time factor `stage_gen_time_s / audio_duration_s` (SLO `< 1`); uses `RTF_BUCKETS` |
-| `vllm:omni_audio_frames_total` | Counter | — | Cumulative audio frames generated |
-| `vllm:omni_audio_underrun_s` | Histogram | — | Per-request worst-case player deficit; `> 0` indicates listener heard silent gaps |
-| `vllm:omni_audio_continuity_ok_total` | Counter | `threshold_ms` | Incremented when the request's worst underrun stayed below `threshold_ms` |
-| `vllm:omni_audio_skipped_requests_total` | Counter | `reason` | Silent-loss counter — code2wav rejected malformed codec input and returned `200 OK` with empty audio |
+| `vllm_omni:audio_ttfp_s` | Histogram | — | Time from request arrival to first audio packet/frame |
+| `vllm_omni:audio_duration_s` | Histogram | — | Audio content duration (`audio_frames / sample_rate`) |
+| `vllm_omni:audio_rtf` | Histogram | — | Real-time factor `stage_gen_time_s / audio_duration_s` (SLO `< 1`); uses `RTF_BUCKETS` |
+| `vllm_omni:audio_frames_total` | Counter | — | Cumulative audio frames generated |
+| `vllm_omni:audio_underrun_s` | Histogram | — | Per-request worst-case player deficit; `> 0` indicates listener heard silent gaps |
+| `vllm_omni:audio_continuity_ok_total` | Counter | `threshold_ms` | Incremented when the request's worst underrun stayed below `threshold_ms` |
+| `vllm_omni:audio_skipped_requests_total` | Counter | `reason` | Silent-loss counter — code2wav rejected malformed codec input and returned `200 OK` with empty audio |
 
 ### Cross-stage transfer (4)
 
@@ -187,10 +233,10 @@ Labels: `{model_name, from_stage, from_replica, to_stage, to_replica}`.
 
 | Metric | Type | Description |
 |--------|------|-------------|
-| `vllm:omni_transfer_size_bytes` | Histogram | Per-transfer payload size in bytes |
-| `vllm:omni_transfer_tx_s` | Histogram | Sender-side time (serialize + submit to connector) |
-| `vllm:omni_transfer_rx_s` | Histogram | Receiver-side time (recv + deserialize) |
-| `vllm:omni_transfer_in_flight_s` | Histogram | Network in-flight time (TX done → RX recv start) |
+| `vllm_omni:transfer_size_bytes` | Histogram | Per-transfer payload size in bytes |
+| `vllm_omni:transfer_tx_s` | Histogram | Sender-side time (serialize + submit to connector) |
+| `vllm_omni:transfer_rx_s` | Histogram | Receiver-side time (recv + deserialize) |
+| `vllm_omni:transfer_in_flight_s` | Histogram | Network in-flight time (TX done → RX recv start) |
 
 ### LLM stage-level (wrapped `vllm:*`)
 
@@ -203,7 +249,7 @@ After the wrap, every upstream `vllm:*` family — TTFT, ITL, TPOT, e2e latency,
   - `SECONDS_FAST_BUCKETS` (0.001 s – 60 s) for fine-grained cross-stage transfer and audio-underrun values that need millisecond-level resolution.
 - Counters use the `_total` suffix (auto-appended by `prometheus_client`).
 - Sizes use the `_bytes` suffix.
-- All omni-specific families are prefixed `vllm:omni_`. The upstream `unregister_vllm_metrics()` function is monkey-patched to a scoped version that still strips upstream `vllm:*` collectors (so multi-engine init within one process does not crash on duplicate registration) but preserves anything prefixed `vllm:omni_` / `vllm_omni`.
+- All omni-specific families are prefixed `vllm_omni:`. The upstream `unregister_vllm_metrics()` function is monkey-patched to a scoped version that still strips upstream `vllm:*` collectors (so multi-engine init within one process does not crash on duplicate registration) but preserves anything prefixed `vllm_omni:` / `vllm_omni`.
 
 ## Logging vs. Prometheus
 

docs/design/module/diffusion/offloader.md

@@ -18,7 +18,7 @@ validation_paths:
   - tests/diffusion/offloader/**
 upstream_refs:
   - torch.nn.Module.to
-last_reviewed: 2026-07-16
+last_reviewed: 2026-08-25
 ---
 
 # Diffusion offloader
@@ -48,6 +48,11 @@ and correctness required by the selected execution mode.
 **Rule:** Retained host and device copies MUST respect configured limits and
 provide deterministic teardown.
 
+Allocator-cache retention MUST remain local to an explicit component owner,
+declare both cache and physical-free-memory bounds, release conservatively
+when telemetry is unavailable, and force release on failure or memory
+pressure. It MUST NOT replace unconditional executor shutdown cleanup.
+
 ## Safe-change guide
 
 Test enabled and disabled paths, repeated execution, asynchronous transfer,

docs/design/module/host_weight_runtime.md

@@ -16,7 +16,7 @@ validation_paths:
   - tests/host_weight_runtime/**
 upstream_refs:
   - https://github.com/vllm-project/vllm-omni/issues/6414
-last_reviewed: 2026-08-21
+last_reviewed: 2026-08-25
 ---
 
 # Host Weight Runtime
@@ -84,6 +84,13 @@ another. `WeightArtifactIdentity` includes:
 - producer version, implementation fingerprint, manifest schema, and restorer
   schema.
 
+Loader integrations may persist content digests for canonical source shards in
+the selected storage domain. Reuse requires the same path, inode, size,
+timestamps, symlink target, and cache-record checksum. Corrupt records are
+rebuilt, and coordination or cache-I/O failure falls back to hashing the
+canonical source directly. This is an identity-computation optimization, not a
+second source of truth or an artifact substitution rule.
+
 The identity excludes transfer policy such as registered mmap versus private
 pinned staging. Those paths move the same host representation and must not
 produce duplicate artifacts.
@@ -237,6 +244,12 @@ Every publication computes full file and tensor hashes. Local lookup supports:
   and
 - `full_checksum`, which additionally reads and hashes every payload file.
 
+An ordered producer may compute file and tensor digests while payload bytes are
+written and overlap each closed payload's `fsync` with later producer work. An
+unordered producer retains parallel readback hashing. In both cases manifest
+and `READY` publication wait for every payload digest and `fsync`; the atomic
+publication and validation contract is unchanged.
+
 Filesystem verity is modeled but unsupported. Invalid artifacts receive an
 external deny marker before their shared lookup lock is released; a build owner
 then quarantines the immutable directory under the exclusive lock.

docs/mkdocs/hooks/generate_examples.py

@@ -37,7 +37,6 @@ GENERAL_EXAMPLE_SLUGS = frozenset(
         "text_to_speech",
         "text_to_video",
         "x_to_text",
-        "x_to_video_audio",
     }
 )
 

docs/models/supported_models.md

@@ -52,11 +52,12 @@ th {
 | `LTX2DistilledOneStagePipeline` | LTX-2.5 merged-distilled one-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LTX2DistilledTwoStagePipeline` | LTX-2.5 merged-distilled two-stage T2V and I2V | `Lightricks/LTX-2.5-Diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/LTX/LTX-2.5.md) |
 | `LingBotVideoPipeline` | LingBot-Video dense and MoE T2I, T2V, TI2V | `robbyant/lingbot-video-dense-1.3b`, `robbyant/lingbot-video-moe-30b-a3b` | ✅︎ | | | | — |
+| `SanaVideoPipeline` | SANA-Video-2B T2V (native 480p/720p; Diffusers adapter validated at both resolutions) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-Video-2B.md) |
+| `SanaImageToVideoPipeline` | SANA-Video-2B I2V (native and Diffusers adapter validated at 480p/720p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/NVIDIA/SANA-Video-2B.md) |
 | `MiniMaxH3Pipeline` | MiniMax H3 T2VA, first/last-keyframe FL2VA, and mixed-reference Ref2VA | `MiniMaxAI/MiniMax-H3` | ✅︎ | ✅︎ | | | [Published](https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3) |
 | `DreamZeroPipeline` | DreamZero-DROID | `GEAR-Dreams/DreamZero-DROID` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `Pi0Pipeline` | π0 (Pi-Zero) VLA | `lerobot/pi0_base` | ✅︎ | | | | [Repository](https://github.com/vllm-project/vllm-omni/blob/main/recipes/lerobot/Pi0.md) |
 | `HeliosPipeline`, `HeliosPyramidPipeline` | Helios | `BestWishYsh/Helios-Base`, `BestWishYsh/Helios-Mid`, `BestWishYsh/Helios-Distilled` | ✅︎ | ✅︎ | ✅︎ | | — |
-| `MagiHumanPipeline` | MagiHuman | `SII-GAIR/daVinci-MagiHuman-Base-1080p` | ✅︎ | ✅︎ | | | — |
 | `OvisImagePipeline` | Ovis-Image | `OvisAI/Ovis-Image` | ✅︎ | ✅︎ | | ✅︎ | — |
 | `LongcatImagePipeline` | LongCat-Image | `meituan-longcat/LongCat-Image` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
 | `LongCatImageEditPipeline` | LongCat-Image-Edit | `meituan-longcat/LongCat-Image-Edit` | ✅︎ | ✅︎ | ✅︎ | ✅︎ | — |
@@ -92,7 +93,6 @@ 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` | ✅︎ | ✅︎ | | ✅︎ | — |
-| `DreamIDOmniPipeline` | DreamID-Omni | `XuGuo699/DreamID-Omni` | ✅︎ | ✅︎ | | | — |
 | `SenseNovaU1Pipeline` | SenseNova-U1 (DiT-only) | `SenseNova/SenseNova-U1-8B-MoT` | ✅︎ | | | | — |
 | `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` | ✅︎ | ✅︎ | | | — |

docs/serving/speech_api.md

@@ -131,7 +131,7 @@ Content-Type: application/json
 
 | Parameter | Type | Default | Description |
 |-----------|------|---------|-------------|
-| `ref_audio` | string | null | Reference audio (HTTP URL, base64 data URL, or `file://` URI with `--allowed-local-media-path`) |
+| `ref_audio` | string | null | Reference audio (HTTP URL, base64 data URL, or `file://` URI with `--allowed-local-media-path`). Local files fold `mtime_ns` and `size` into cache keys to automatically reload on-disk edits; HTTP URLs and base64 URIs remain cached by string locator. |
 | `ref_text` | string | null | Transcript of reference audio |
 | `x_vector_only_mode` | bool | null | Use speaker embedding only (no ICL) |
 

docs/user_guide/diffusion/attention_backends.md

@@ -20,6 +20,7 @@ For the internal selector, registry, and platform contract, see
 | Install and use SageAttention 2.2 or SageAttention3 | [SageAttention](attention_backends/sage.md) |
 | Match training or rollout kernels loaded from Hugging Face | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | Use block-sparse video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
+| Use FastVideo VSA with FastWan2.2-TI2V-5B on CUDA | [FastVideo VSA](attention_backends/fastvideo_vsa.md) |
 
 ## Backend options
 
@@ -35,6 +36,7 @@ For the internal selector, registry, and platform contract, see
 | `FLASH_ATTN_HUB` | Hub kernel | FlashAttention 2 from Hugging Face `kernels` | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | `FLASH_ATTN_3_HUB` | Hub kernel | FlashAttention 3 from Hugging Face `kernels` on Hopper or newer | [Hugging Face Hub Backends](attention_backends/huggingface_hub.md) |
 | `RAINFUSION_ATTN` | Block sparse | MindIE-SD RainFusion video attention on Ascend NPU | [RainFusion](attention_backends/rainfusion.md) |
+| `FASTVIDEO_VSA` | Block sparse | FastVideo variable sparse self-attention for FastWan2.2-TI2V-5B on CUDA | [FastVideo VSA](attention_backends/fastvideo_vsa.md) |
 
 ## Configuration
 
@@ -105,6 +107,7 @@ Backend-specific typed blocks are documented with their consumers:
   and [TRTLLM SAGE](attention_backends/trtllm.md#sage-quantization).
 - `skip_softmax`: [TRTLLM Skip-Softmax](attention_backends/trtllm.md#skip-softmax).
 - `block_sparse`: [RainFusion](attention_backends/rainfusion.md#configuration).
+- `fastvideo_vsa_topk`: [FastVideo VSA](attention_backends/fastvideo_vsa.md#choose-top-k).
 
 ## Platform defaults
 

docs/user_guide/diffusion/attention_backends/fastvideo_vsa.md

@@ -0,0 +1,125 @@
+# FastVideo VSA
+
+FastVideo Variable Sparse Attention (VSA) accelerates the `FastVideo/FastWan2.2-TI2V-5B-Diffusers` model's
+self-attention by partitioning the post-patch latent grid into spatiotemporal
+blocks. For every query block, VSA scores the key/value blocks and computes
+attention only against the selected top-k blocks.
+
+The supported checkpoint provides both text-to-video and image-to-video modes through `Wan22Pipeline`; the separate Wan I2V-14B, S2V, and VACE pipelines are outside this backend's supported scope.
+
+VSA is a CUDA-only, explicitly selected backend. It requires the
+`fastvideo-kernel` package and currently supports non-causal self-attention
+with equal query and key/value sequence lengths. Unsupported shapes, masks,
+dtypes, sequence-parallel execution, or kernel failures fall back to
+`TORCH_SDPA` and emit a warning with the reason.
+
+## Enable the backend
+
+For online serving, select the backend with the existing attention backend
+flag. Use `--fastvideo-vsa-topk` to set the number of key/value blocks retained
+for every query block:
+
+```bash
+vllm-omni serve <model> \
+  --diffusion-attention-backend FASTVIDEO_VSA \
+  --fastvideo-vsa-topk 64
+```
+
+The backwards-compatible environment variable selects the backend with the
+default `topk=64`:
+
+```bash
+export DIFFUSION_ATTENTION_BACKEND=FASTVIDEO_VSA
+vllm-omni serve <model>
+```
+
+To tune top-k, pass the CLI backend and top-k flags together as shown above,
+or use the equivalent structured configuration:
+
+```bash
+vllm-omni serve <model> \
+  --diffusion-attention-config \
+  '{"default":{"backend":"FASTVIDEO_VSA","fastvideo_vsa_topk":64}}'
+```
+
+Do not combine `--diffusion-attention-backend` with an explicit
+`diffusion_attention_config.default.backend`. The top-k value must be positive
+and is valid only when the default backend is `FASTVIDEO_VSA`.
+
+For a deploy YAML stage, use either the shorthand fields:
+
+```yaml
+stages:
+  - stage_id: 0
+    diffusion_attention_backend: FASTVIDEO_VSA
+    fastvideo_vsa_topk: 64
+```
+
+or the structured configuration:
+
+```yaml
+stages:
+  - stage_id: 0
+    diffusion_attention_config:
+      default:
+        backend: FASTVIDEO_VSA
+        fastvideo_vsa_topk: 64
+```
+
+## Choose top-k
+
+At runtime the backend logs the sequence shape and derived block count:
+
+```text
+FASTVIDEO_VSA routing: seq_len=27280, dit_seq_shape=(31, 22, 40),
+block_size=(4, 8, 8), num_blocks=120, topk=64, keep_ratio=53.3%,
+checkpoint_mode=native, route=VSA
+```
+
+Use `num_blocks` as the upper bound when tuning top-k. Top-k is the number of
+key/value blocks retained for each query block, not the number of tokens or
+the total number of blocks processed by the layer.
+
+- A smaller top-k increases sparsity and may reduce attention computation, but
+  it can remove relevant blocks and reduce visual quality or temporal
+  consistency. Routing and padding overhead can also make a smaller value
+  slower for some shapes.
+- A larger top-k retains more context and generally approaches dense-attention
+  quality, but reduces the potential speedup and increases memory traffic.
+- `topk > num_blocks` is invalid for the runtime shape and falls back to SDPA.
+- For a native checkpoint, `topk == num_blocks` routes to SDPA because scoring
+  every block provides no sparsity benefit.
+- For a FastVideo DMD checkpoint, `topk == num_blocks` stays on the VSA
+  all-block path so the checkpoint keeps its trained compensation semantics.
+
+There is no universally optimal value. Resolution, frame count, GPU, kernel
+version, and checkpoint all affect both quality and latency. Start from the
+logged `num_blocks`, test several keep ratios on the target workload, and
+compare output quality against the same checkpoint running dense attention.
+
+## Checkpoint behavior
+
+The backend does not expose a user-selectable gate mode. Wan checkpoints that
+contain learned `to_gate_compress` weights use that projection automatically.
+When those weights are absent, the unused projection is removed and VSA uses
+the sparse branch without learned compensation.
+
+FastVideo DMD checkpoints use their fixed distilled timestep schedule. Native
+Wan checkpoints keep their normal scheduler and inference-step configuration;
+selecting VSA does not turn a native checkpoint into a distilled model.
+
+## Verify routing and fallback
+
+Check the startup and first-forward logs instead of assuming that selecting
+the backend guarantees sparse execution:
+
+- `route=VSA` means top-k block selection is active.
+- `route=VSA_ALL_BLOCKS` means the FastVideo DMD checkpoint retained all
+  blocks through the VSA kernel.
+- `route=SDPA` or `FASTVIDEO_VSA falling back to SDPA: ...` means dense SDPA
+  executed; the warning includes the reason.
+
+The current kernel requires CUDA tensors in FP16 or BF16, 256-token blocks,
+standard `head_size**-0.5` scaling, equal Q/K/V head counts, no attention mask,
+and no active sequence-parallel context. NPU and XPU paths do not execute the
+FastVideo VSA CUDA kernel.

docs/user_guide/diffusion/execution_modes.md

@@ -101,8 +101,16 @@ set `DIFFUSION_ATTENTION_BACKEND=TORCH_SDPA` or configure
 `--max-num-seqs >1`. See the
 [HunyuanImage-3.0 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/Tencent/HunyuanImage-3.0-Instruct.md)
 for its validated configuration. Helios supports single-request step execution only: use
-`--step-execution --max-num-seqs 1` for Helios. Consult the selected pipeline's
-documentation and source for the latest support status.
+`--step-execution --max-num-seqs 1` for Helios. MiniMax H3 supports step-wise
+continuous batching by packing co-batched requests into one sequence that keeps
+a separate attention document per request; that layout needs a backend which
+honors the packed `cu_seqlens` metadata, so run it with
+`--diffusion-attention-backend FLASH_ATTN` (other backends stay correct but
+fall back to one transformer forward per request). Batching H3 does not improve
+its throughput — see the measured numbers in the
+[MiniMax-H3 recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/MiniMaxAI/MiniMax-H3.md)
+— so keep `--max-num-seqs 1` unless you need step-level scheduling. Consult the
+selected pipeline's documentation and source for the latest support status.
 
 ## Streaming Output
 

docs/user_guide/diffusion/offloader/distributed_layerwise_offload.md

@@ -63,6 +63,9 @@ omni = Omni(
 | `--dlo-use-allgather` | Shard host weights and reconstruct with AllGather | `true` |
 | `--dlo-no-use-allgather` | Stream complete rank-local blocks without a DLO weight collective | `false` |
 | `--dlo-resident-layers N` | Keep N leading main-DiT blocks on device; requires no-AllGather and model-declared resident paths | `0` |
+| `--host-weight-runtime-mode {disabled,preferred,required}` | HWR policy: no interaction, populate on a miss, or require an exact hit | `disabled` |
+| `--host-weight-runtime-root PATH` | Writable node-local HWR store shared by workers in one storage domain; required for `preferred` and `required` | unset |
+| `--dlo-host-registration-limit-gib N` | Optional per-worker ceiling for registering an HWR mmap; zero adds no ceiling | `0` |
 
 ## Host-weight loading
 
@@ -104,6 +107,130 @@ process.
 When the effective DLO group size is one, `dlo_use_allgather=True` does not
 perform a collective and uses the same rank-local transfer behavior.
 
+### Final-layout Host Weight Runtime
+
+HWR is an opt-in startup optimization for models that declare the final-layout
+BF16 restore contract. Use it only with no-AllGather DLO:
+
+The validated BF16 model contracts currently cover MiniMax H3 and
+`black-forest-labs/FLUX.2-klein-4B`. FLUX.2-klein-9B shares the same model
+class but has not been validated against this contract. FLUX.2-dev, online
+FP8, HSDP, LoRA/adapted weights, and non-default load formats remain outside
+the validated scope.
+
+```bash
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --dlo-no-use-allgather \
+  --host-weight-runtime-mode preferred \
+  --host-weight-runtime-root /var/cache/vllm-omni/hwr
+```
+
+#### Choosing a mode
+
+| Mode | Exact local artifact hit | Miss or recoverable artifact/store problem | Intended use |
+| --- | --- | --- | --- |
+| `disabled` | HWR is not consulted | Use the existing checkpoint-mmap or ordinary-loader path | Default compatibility path |
+| `preferred` | Restore the final-layout artifact | Load canonically, serve with those tensors, and attempt to publish an artifact for the next startup | Normal deployment and store population |
+| `required` | Restore the final-layout artifact | Fail startup without canonical DiT fallback or publication | Enforce a pre-populated store in controlled rollouts or CI |
+
+Both enabled modes still fail on non-retryable configuration, identity, or
+compatibility errors. `preferred` is a fallback policy for a cache miss or a
+recoverable cache problem; it does not hide an invalid deployment.
+
+#### Populating a store for `required` mode
+
+`required` is deliberately consume-only. PR2 does not include a separate
+prewarm command, so populate each node-local storage domain with one matching
+`preferred` producer cohort:
+
+1. Choose a persistent, writable root visible to every diffusion worker in the
+   storage domain. Do not use a process-private temporary directory.
+2. Start the deployment with `preferred`, using the exact model revision,
+   dtype, TP size, SP configuration, and other weight-layout settings intended
+   for serving. No inference request is needed; publication happens during
+   model startup. Wait for the engine to become healthy, then shut it down
+   cleanly.
+3. Restart with the same arguments and root, changing only the mode to
+   `required`. A successful startup proves that every worker acquired a valid
+   artifact; a miss, corrupt artifact, or incompatible identity fails startup.
+4. Repeat the population start on every node or storage domain because this
+   store is node-local.
+
+For example:
+
+```bash
+# First startup: canonically load and populate the exact artifacts.
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --dlo-no-use-allgather \
+  --host-weight-runtime-mode preferred \
+  --host-weight-runtime-root /var/cache/vllm-omni/hwr
+
+# After a healthy startup and clean shutdown, enforce cache hits.
+vllm serve /path/to/model --omni \
+  --enable-distributed-layerwise-offload \
+  --dlo-no-use-allgather \
+  --host-weight-runtime-mode required \
+  --host-weight-runtime-root /var/cache/vllm-omni/hwr
+```
+
+Include the same TP and SP layout flags in both commands. DP rank and DP size
+are excluded from artifact identity, so the population and serving DP sizes may
+differ and equivalent DP replicas share artifacts. TP rank is included, so a
+TP-N deployment normally needs N rank-specific artifacts in each storage
+domain; launching the matching TP cohort creates that set. If the model
+revision or layout changes, run `preferred` again for the new identity before
+returning to `required`.
+
+On a cold start, the canonical loader remains authoritative and publishes a
+validated final-layout artifact for later workers. A warm start restores the
+DiT final tensors without ordinary DiT materialization. DLO then attempts to
+register the immutable mapping for direct H2D; if registration is unavailable,
+it streams through the same two bounded host staging slots. The artifact
+identity includes the TP rank/size and SP layout, so TP1, TP2 rank-local shards,
+and distinct SP layouts do not alias one another. Publication failure remains
+separate from a `preferred` serving startup; the next `required` startup
+provides the explicit artifact-availability check.
+
+For local canonical checkpoints, the first eligible worker may hash source
+shards to establish immutable identity. HWR caches those digests in the same
+node-local storage domain and validates file metadata before reuse, so later
+workers normally avoid repeating that read. Cold BF16 publication also hashes
+ordered payloads as they are written and overlaps payload durability work with
+later shards; this changes startup work only, not artifact contents or runtime
+H2D behavior.
+
+#### Registered direct H2D
+
+Registration is attempted automatically only for an eligible warm HWR hit when
+`pin_cpu_memory` is enabled. A successful path is:
+
+```text
+shared read-only HWR mmap -> existing rotating HBM block buffer -> GPU kernel
+```
+
+It removes the recurrent mmap-to-private-staging CPU copy and does not allocate
+the two host staging slots. It does not reduce the H2D payload or make GPU
+kernels access host memory directly.
+
+CUDA requires read-only host-registration capability for the immutable HWR
+mapping. Unsupported capability, a positive registration limit smaller than
+the complete page-aligned mapping, or a safely rolled-back registration error
+falls back to bounded staging. Programmatic configurations can set
+`pin_cpu_memory=False` to disable registration explicitly. A successful
+registration locks the complete mapped range in host memory for that worker's
+lifetime, so use
+`--dlo-host-registration-limit-gib` when an operator-enforced ceiling is
+required.
+
+Shutdown drains H2D work, releases hook/source references, unregisters every
+range, and only then closes the HWR lease. AllGather and direct checkpoint mmap
+retain their existing transfer paths.
+
+When HWR is disabled, DLO is disabled, or AllGather is enabled, the loader does
+not resolve HWR sources or construct its store.
+
 ## Declarative topology
 
 Models may declare an `OffloadPlan` instead of embedding offload logic:
@@ -135,7 +262,9 @@ must enter each collective.
 - Direct checkpoint mmap currently requires TP1. TP greater than one is
   outside the Phase A shared-mmap support scope and falls back before model
   mutation to the ordinary TP-aware loader. DLO can stream that runtime layout,
-  but it provides no shared-mmap host-memory benefit or guarantee.
+  and eligible no-AllGather configurations may use HWR final-layout artifacts,
+  but direct checkpoint mmap provides no shared-mmap host-memory guarantee for
+  TP greater than one.
 - HSDP plus AllGather is rejected to avoid double sharding. HSDP without
   AllGather has limited end-to-end validation.
 - Per-tensor online FP8 linears use the ordinary loader and can run with either
@@ -147,10 +276,10 @@ must enter each collective.
   `OffloadPlan` that declares eligible `resident_dit_paths`.
 - DP concurrency requires an explicit, identical inference-step count.
 
-Sharing transformed TP or quantized runtime layouts through a normalized mmap
-cache is a follow-up design in
+Sharing quantized or otherwise unvalidated transformed layouts through a
+normalized HWR producer is a follow-up design in
 [RFC #6195](https://github.com/vllm-project/vllm-omni/issues/6195), not part of
-the direct-checkpoint path.
+the current BF16 final-layout path.
 
 See the [Cosmos3 DistOffload recipe](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-DistOffload.md)
 for an end-to-end example.

docs/user_guide/diffusion_features.md

@@ -129,7 +129,6 @@ The following tables show which models support each feature:
 | **Krea 2**               |     ❌     |     ✅      |           ❌           |       ❌        |         ❌         |          ❌          |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LongCat-Image**        |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **LongCat-Image-Edit**   |     ✅     |     ✅      |           ✅           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
-| **MagiHuman**            |     ❌     |     ❌      |           ❌           |       ❓        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Nextstep_1(T2I)**      |     ❓     |     ❓      |           ❌           |       ✅        |         ✅         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **OmniGen2**             |     ❌     |     ✅      |           ✅           |       ❌        |         ✅         |          ❌          |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **Ovis-Image**           |     ❌     |     ✅      |           ❌           |       ✅        |         ❌         |          ❌          |   ❌    |             ✅             |          ❌           |       ❌        |        ❌         |
@@ -147,7 +146,7 @@ The following tables show which models support each feature:
 
 > Notes:
 > 1. Nextstep_1(T2I) does not support cache acceleration methods such as TeaCache or Cache-DiT.
-> 2. `Tongyi-MAI/Z-Image-Turbo` and `SII-GAIR/daVinci-MagiHuman-Base-1080p` are distilled models with minimal NFEs; CFG-Parallel is not necessary.
+> 2. `Tongyi-MAI/Z-Image-Turbo` is a distilled model with minimal NFEs; CFG-Parallel is not necessary.
 > 3. Cosmos3 T2I uses `Cosmos3OmniDiffusersPipeline` with `modalities=["image"]`. Model-level CPU offload swaps the nested UND reasoner and GEN generator pathways; layerwise offload remains available for blockwise GEN/UND offload.
 > 4. Krea 2 currently supports single-GPU inference plus LoRA, Cache-DiT, HSDP, CPU/layerwise offload, and VAE-patch-parallel (decode). TP/SP/CFG-Parallel are not yet wired. The few-step distilled (Turbo) checkpoint uses `is_distilled=true` (fixed timestep shift `mu=1.15`); generate at 2048x2048 by default with `num_inference_steps≈8` and `guidance_scale=0`. The Raw checkpoint uses 1024x1024, `num_inference_steps=28`, and `guidance_scale=4.5`.
 > 5. HunyuanImage3 supports step execution. Multi-request step execution requires `TORCH_SDPA`; see [Diffusion Execution Modes](diffusion/execution_modes.md#step-execution).
@@ -162,9 +161,9 @@ The following tables show which models support each feature:
 | **LTX-2**                    |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |       ❌        |        ❌         |
 | **LTX-2.5**                  |     ❌     | ❓ (one-stage) | ❓ (Ulysses only) | ✅ (Full only) |         ❓         |         ❌         |   ✅    |             ✅             |      ✅ (decode)      |    ❓ (FP8)     |        ❌         |
+| **SANA-Video-2B T2V I2V**   |     ❓     |     ❓      |           ❓           |       ❓        |         ❓         |         ❓         |   ❓    |             ❓             |          ❓           |       ❓        |        ❌         |
 | **Helios**                   |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ✅*        |
 | **HunyuanVideo-1.5 T2V I2V** |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
-| **DreamID-Omni**             |     ❌     |     ❌      |           ❌           |       ✅        |         ❌         |         ❌         |   ✅    |             ✅             |          ❌           |       ❌        |        ❌         |
 | **Cosmos3**                  |     ❌     |     ✅      |           ✅           |       ✅        |         ✅         |         ❌         |   ✅    |             ✅             |  ✅ (encode/decode)   |       ✅        |        ❌         |
 | **LongCat-Video-Avatar-1.5** |     ❌     |     ❌      |           ❌           |       ❌        |         ❌         |         ❌         |   ❌    |             ❌             |          ❌           |       ❌        |        ❌         |
 | **MiniMax-H3**               | ✅ (FL2VA) |     ✅      |           ✅           |       ❌        |       ✅ (DiT/TE)  |         ❌         |   ✅    |             ✅             |       ✅ (tile)       |      ✅ (DiT)      |        ❌         |
@@ -182,8 +181,8 @@ The following tables show which models support each feature:
 
 **Frame Interpolation Support**
 
-- **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines
-- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni, SANA-WM
+- **Supported**: Wan2.2 text-to-video, image-to-video, and TI2V pipelines; SANA-Video-2B native text-to-video and image-to-video pipelines
+- **Not supported**: Wan2.1-VACE, LTX-2, LTX-2.3, LTX-2.5, Helios, HunyuanVideo-1.5, DreamID-Omni, SANA-WM; SANA-Video-2B Diffusers-adapter text-to-video and image-to-video pipelines
 
 ### AudioGen
 

docs/user_guide/examples/offline_inference/image_to_video.md

@@ -4,8 +4,8 @@ Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inf
 
 
 This example demonstrates how to generate videos from images using Wan2.2,
-LTX-2, and HunyuanVideo-1.5 Image-to-Video models with vLLM-Omni's
-offline inference API.
+LTX-2, HunyuanVideo-1.5, and SANA-Video Image-to-Video models with
+vLLM-Omni's offline inference API.
 
 ## Local CLI Usage
 
@@ -66,9 +66,49 @@ python image_to_video.py \
 See the [LTX-2 recipe](../../../../recipes/LTX/LTX-2.md) for all checkpoints,
 pipeline selection, T2V, defaults, and advanced options.
 
+### SANA-Video-2B
+
+SANA checkpoints identify their upstream T2V pipeline in `model_index.json`.
+Select the native I2V implementation explicitly:
+
+```bash
+python image_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  --model-class-name SanaImageToVideoPipeline \
+  --image cherry_blossom.jpg \
+  --prompt "Cherry blossoms sway in the breeze as petals drift past the camera." \
+  --negative-prompt "blurry, low quality, temporal artifacts" \
+  --height 480 \
+  --width 832 \
+  --num-frames 81 \
+  --num-inference-steps 50 \
+  --guidance-scale 6.0 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 \
+  --seed 42 \
+  --output sana_video_i2v_480p.mp4
+```
+
+For the 720p checkpoint, use
+`Efficient-Large-Model/SANA-Video_2B_720p_diffusers` with
+`--height 704 --width 1280`. The native I2V path supports both checkpoint
+variants: the 480p checkpoint uses the Wan VAE, while the 720p checkpoint uses
+the LTX-2 Video VAE.
+
+For SANA-Video, 81 frames at 16 FPS is the standard checkpoint request
+(approximately five seconds), not minute-scale long-video generation.
+Minute-scale SANA generation requires the separate LongSANA/LongLive
+autoregressive workflow, which this pipeline does not implement.
+
+See the [SANA-Video recipe](../../../../recipes/NVIDIA/SANA-Video-2B.md) for
+online serving, backend boundaries, and the tested hardware profile.
+
 Key arguments:
 
-- `--model`: Model ID (I2V-A14B for MoE, TI2V-5B for unified T2V+I2V, or LTX-2).
+- `--model`: Model ID or local path (for example Wan I2V/TI2V, LTX-2, or
+  SANA-Video).
+- `--model-class-name`: Explicit pipeline override. SANA-Video I2V requires
+  `SanaImageToVideoPipeline`.
 - `--image`: Path to input image (required).
 - `--prompt`: Text description of desired motion/animation.
 - `--height/--width`: Output resolution (auto-calculated from image if not set).

docs/user_guide/examples/offline_inference/text_to_video.md

@@ -16,6 +16,8 @@ For backend selection and SageAttention usage, see the [Diffusion Attention Back
 | `Lightricks/LTX-2` | 512x768 | 121 | 40 | video 3.0 / audio 7.0 | Model-dependent |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v` | 480x832 | 121 | 50 | 6.0 | 1×A100 80GB |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
+| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480x832 | 81 | 50 | 6.0 | BF16 DiT + FP32 Wan VAE |
+| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704x1280 | 81 | 50 | 6.0 | ~23.6 GiB |
 
 ## Local CLI Usage
 
@@ -122,6 +124,31 @@ python text_to_video.py \
   --output quick_test.mp4
 ```
 
+### SANA-Video-2B
+
+```bash
+python text_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  --model-class-name SanaVideoPipeline \
+  --prompt "A cinematic tracking shot of a sailboat crossing the ocean at sunset." \
+  --height 480 \
+  --width 832 \
+  --num-frames 81 \
+  --num-inference-steps 50 \
+  --guidance-scale 6.0 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 \
+  --output sana_video_480p.mp4
+```
+
+For the 720p checkpoint, switch the model to
+`Efficient-Large-Model/SANA-Video_2B_720p_diffusers`, use
+`--height 704 --width 1280`, and write to a different output path. SANA-Video
+accepts `motion_score`, `clean_caption`, and `use_resolution_binning` through
+`--extra-body`. For image-to-video, use the shared
+`examples/offline_inference/image_to_video/image_to_video.py` example with
+`--model-class-name SanaImageToVideoPipeline` and `--image <path>`.
+
 ## Key Arguments
 
 ### Common

docs/user_guide/examples/offline_inference/x_to_video_audio.md

@@ -1,115 +0,0 @@
-# X-To-Video-Audio
-
-Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/offline_inference/x_to_video_audio>.
-
-
-The `DreamID-Omni` pipeline generates short videos from text, image and video.
-
-## Local CLI Usage
-### Download the Model locally
-Since DreamID-Omni combine multiple models, and without any config, so we need to download them locally.
-
-```bash
-python download_dreamid_omni.py --output-dir ./dreamid_omni
-```
-After download, the model directory will look like this:
-
-```
-dreamid_omni/
-├── DreamID-Omni/
-│   ├── dreamid_omni.safetensors
-├── MMAudio/
-│   ├── ext_weights/
-│   │   ├── best_netG.pt
-│   │   ├── v1-16.pth
-├── Wan2.2-TI2V-5B/
-│   ├── google/*
-│   ├── models_t5_umt5-xxl-enc-bf16.pth
-│   ├── Wan2.2_VAE.pth
-│
-├── model_index.json # create by download_dreamid_omni.py
-```
-
-### Run the Inference
-```python
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "Two people walking together and singing happily" \
-  --image-path ./example0.png ./example1.png \
-  --audio-path ./example0.wav ./example1.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --output out_dreamid_omni_twoip.mp4
-```
-In the current test scenario (2 images + 2 audio inputs), the VRAM requirement is 72GB, regardless of whether cfg-parallel is enabled or disabled.
-The VRAM usage can be reduced by enabling CPU offload via --enable-cpu-offload.
-For multi-GPU memory reduction on the fused DreamID-Omni transformer, you can also enable HSDP:
-
-```python
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "Two people walking together and singing happily" \
-  --image-path ./example0.png ./example1.png \
-  --audio-path ./example0.wav ./example1.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --use-hsdp \
-  --hsdp-shard-size 2 \
-  --output out_dreamid_omni_twoip.mp4
-```
-
-
-You could take reference images/audios from the test cases in the official repo: https://github.com/Guoxu1233/DreamID-Omni
-
-For example, single IP ref resources can be found under https://github.com/Guoxu1233/DreamID-Omni/tree/main/test_case/oneip, you could download them correspondingly to your local and use them for testing.
-
-```python
-# Example usage for oneip, ref media from the official repo DreamID-Omni
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "<img1>: In the frame, a woman with black long hair is identified as <sub1>.\n**Overall Environment/Scene**: A lively open-kitchen café at night; stove flames flare, steam rises, and warm pendant lights swing slightly as staff move behind her. The shot is an upper-body close-up.\n**Main Characters/Subjects Appearance**: <sub1> is a young woman with thick dark wavy hair and a side part. She wears a fitted black top under a light apron, a thin gold chain necklace, and small stud earrings.\n**Main Characters/Subjects Actions**: <sub1> tastes the sauce with a spoon, then turns her face toward the camera while still holding the spoon, her expression shifting from focused to conflicted.\n<sub1> maintains eye contact, swallows as if choosing her words, and says, <S>I keep telling myself I’m fine,but some nights it feels like I’m just performing calm.<E>" \
-  --image-path 9.png \
-  --audio-path 9.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --output out_dreamid_omni_oneip.mp4
-```
-
-
-Key arguments:
-- `--prompt`: text description (string).
-- `--model`: path to the model local directory.
-- `--height/--width`: output resolution (defaults 704 * 1024).
-- `--image-path`: path to the input image list.
-- `--audio-path`: path to the input audio list, indicate the timbre of the output video.
-- `--cfg-parallel-size`: number of parallel cfg parallel (defaults 1).
-- `--use-hsdp`: enable HSDP weight sharding for DreamID-Omni fused blocks.
-- `--hsdp-shard-size`: number of GPUs used for HSDP sharding.
-- `--hsdp-replicate-size`: number of HSDP replica groups.
-- `--num-inference-steps`: number of denoising steps (defaults 45).
-- `--video-negative-prompt`: negative prompt for video generation.
-- `--audio-negative-prompt`: negative prompt for audio generation.
-- `--enable-cpu-offload`: enable CPU offload (defaults False).
-
-## Example materials
-
-??? abstract "download_dreamid_omni.py"
-    ``````py
-    --8<-- "examples/offline_inference/x_to_video_audio/download_dreamid_omni.py"
-    ``````
-??? abstract "x_to_video_audio.py"
-    ``````py
-    --8<-- "examples/offline_inference/x_to_video_audio/x_to_video_audio.py"
-    ``````

docs/user_guide/examples/online_serving/text_to_audio.md

@@ -1,193 +0,0 @@
-# Text-To-Audio
-
-Source <https://github.com/vllm-project/vllm-omni/tree/main/examples/online_serving/stable_audio>.
-
-This example demonstrates how to deploy Stable Audio models for online text-to-audio generation using vLLM-Omni.
-
-## Supported Models
-
-| Model | Description |
-|-------|-------------|
-| `stabilityai/stable-audio-open-1.0` | Open-source audio generation, up to ~47 seconds, 44.1 kHz stereo |
-
-## Start Server
-
-### Basic Start
-
-```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
-    --host 0.0.0.0 \
-    --port 8091 \
-    --gpu-memory-utilization 0.9 \
-    --trust-remote-code \
-    --enforce-eager \
-    --omni
-```
-
-## API Calls
-
-### Method 1: Using curl
-
-```bash
-# Run all curl examples
-bash curl_examples.sh
-
-# Or execute directly
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "The sound of a cat purring",
-        "audio_length": 10.0
-    }' --output cat.wav
-```
-
-### Method 2: Using Python Client
-
-```bash
-cd examples/online_serving/stable_audio
-
-# Simple generation
-python stable_audio_client.py \
-    --text "The sound of a cat purring"
-
-# With custom duration
-python stable_audio_client.py \
-    --text "A dog barking" \
-    --audio_length 5.0
-
-# With all parameters
-python stable_audio_client.py \
-    --text "Thunder and rain" \
-    --audio_length 15.0 \
-    --negative_prompt "Low quality" \
-    --guidance_scale 7.0 \
-    --num_inference_steps 100 \
-    --seed 42 \
-    --output thunder.wav
-```
-
-The Python client supports the following command-line arguments:
-
-- `--api_url`: API endpoint URL (default: `http://localhost:8091/v1/audio/generate`)
-- `--text`: Text prompt for audio generation (default: `"The sound of a cat purring"`)
-- `--audio_length`: Audio length in seconds (default: `10.0`, max ~47s for `stable-audio-open-1.0`)
-- `--audio_start`: Audio start time in seconds (default: `0.0`)
-- `--negative_prompt`: Negative prompt for classifier-free guidance (default: `"Low quality"`)
-- `--guidance_scale`: Guidance scale for diffusion (default: `7.0`)
-- `--num_inference_steps`: Number of inference steps (default: `100`)
-- `--seed`: Random seed for reproducibility (default: `None`)
-- `--response_format`: Audio output format (default: `wav`). Options: `wav`, `mp3`, `flac`, `pcm`
-- `--output`: Output file path (default: `stable_audio_output.wav`)
-
-### Method 3: Using Python httpx
-
-```python
-import httpx
-
-response = httpx.post(
-    "http://localhost:8091/v1/audio/generate",
-    json={
-        "input": "The sound of ocean waves crashing on a beach",
-        "audio_length": 10.0,
-        "negative_prompt": "Low quality, distorted",
-        "guidance_scale": 7.0,
-        "num_inference_steps": 100,
-    },
-    timeout=300.0,
-)
-
-with open("ocean.wav", "wb") as f:
-    f.write(response.content)
-```
-
-## Request Format
-
-### Simple Generation
-
-```json
-{
-    "input": "The sound of ocean waves"
-}
-```
-
-### Generation with Parameters
-
-```json
-{
-    "input": "A piano playing a gentle melody",
-    "audio_length": 10.0,
-    "negative_prompt": "Low quality, distorted, noisy",
-    "guidance_scale": 8.0,
-    "num_inference_steps": 150,
-    "seed": 42,
-    "response_format": "wav"
-}
-```
-
-## API Reference
-
-### Endpoint
-
-```
-POST /v1/audio/generate
-Content-Type: application/json
-```
-
-### Generation Parameters
-
-| Parameter | Type | Default | Description |
-|-----------|------|---------|-------------|
-| `input` | string | **required** | Text prompt describing the audio to generate |
-| `model` | string | server's model | Model to use (optional, should match server if specified) |
-| `response_format` | string | "wav" | Audio format: wav, mp3, flac, pcm, opus |
-| `speed` | float | 1.0 | Playback speed (0.25 - 4.0) |
-| `audio_length` | float | null | Audio duration in seconds (max ~47s for `stable-audio-open-1.0`) |
-| `audio_start` | float | 0.0 | Audio start time in seconds |
-| `negative_prompt` | string | null | Text describing what to avoid in generation |
-| `guidance_scale` | float | model default | Classifier-free guidance scale (higher = more adherence to prompt) |
-| `num_inference_steps` | int | model default | Number of denoising steps (higher = better quality, slower) |
-| `seed` | int | null | Random seed for reproducible generation |
-
-### Response Format
-
-Returns binary audio data with appropriate `Content-Type` header (e.g., `audio/wav`).
-
-## Tuning Tips
-
-1. **Audio Length**: Keep under 47 seconds for `stable-audio-open-1.0`.
-2. **Quality vs Speed**:
-   - 50 steps: Fast, decent quality (quick previews)
-   - 100 steps: Good balance (general purpose)
-   - 150+ steps: High quality, slower (final / critical audio)
-3. **Guidance Scale**:
-   - Lower (3 - 5): More creative / varied output
-   - Default (7): Good balance
-   - Higher (10+): Strict adherence to the prompt
-4. **Negative Prompts**: Use to avoid unwanted characteristics such as `"Low quality"`, `"distorted"`, `"noisy"`.
-5. **Seeds**: Set a fixed seed to get deterministic, reproducible results.
-
-## File Description
-
-| File | Description |
-|------|-------------|
-| `curl_examples.sh` | Curl examples covering common use cases |
-| `stable_audio_client.py` | Python client with full CLI argument support |
-
-## Troubleshooting
-
-1. **Audio generation model did not produce audio output**: Verify the server started successfully and the model loaded without errors.
-2. **Connection refused**: Make sure the server is running on the correct port.
-3. **Generation timeout**: Reduce `num_inference_steps` or `audio_length`, and check GPU memory with `nvidia-smi`.
-4. **Out of memory**: Lower `--gpu-memory-utilization` or reduce `audio_length`.
-5. **Audio quality issues**: Increase `num_inference_steps`, add a negative prompt, or raise `guidance_scale`.
-
-## Example materials
-
-??? abstract "stable_audio_client.py"
-    ``````py
-    --8<-- "examples/online_serving/stable_audio/stable_audio_client.py"
-    ``````
-??? abstract "curl_examples.sh"
-    ``````sh
-    --8<-- "examples/online_serving/stable_audio/curl_examples.sh"
-    ``````

examples/offline_inference/image_to_video/README.md

@@ -1,7 +1,7 @@
 # Image-To-Video
 
 This shared example generates videos from images with VACE, Wan2.2, LTX-2,
-HunyuanVideo-1.5, Cosmos3, and other compatible pipelines.
+HunyuanVideo-1.5, SANA-Video, Cosmos3, and other compatible pipelines.
 
 - `image_to_video.py`: command-line script for single video generation with advanced options.
 
@@ -27,6 +27,8 @@ This folder provides a unified CLI script for image-to-video generation using vL
 | `Wan-AI/Wan2.2-TI2V-5B-Diffusers` | 480 x 832 | 81 | 50 | 4.0 | Around 20–25 GiB BF16, smallest I2V model |
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_i2v` | 480 x 832 | 121 | 50 | 6.0 | Around 100 GiB at default settings; the example enables `--enable-cpu-offload` + VAE tiling/slicing to fit an 80 GiB card |
 | `Lightricks/LTX-2` | 512 x 768 | 121 | 40 | video 3.0 / audio 7.0 | Memory use depends on frame count and tensor parallelism |
+| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480 x 832 | 81 | 50 | 6.0 | Native `SanaImageToVideoPipeline`; Wan VAE |
+| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704 x 1280 | 81 | 50 | 6.0 | Native `SanaImageToVideoPipeline`; LTX-2 Video VAE |
 
 !!! info
     Peak VRAM: based on basic single-card usage, batch size = 1, without any acceleration/optimization features. Some model weights cannot fit into one card with 80 GiB VRAM, which may need to use CPU offloading.
@@ -227,6 +229,48 @@ python image_to_video.py \
 See the [LTX-2 recipe](../../../recipes/LTX/LTX-2.md) for all checkpoints,
 pipeline selection, T2V, defaults, and advanced options.
 
+### SANA-Video-2B
+
+The checkpoint metadata selects the upstream T2V pipeline, so pass
+`SanaImageToVideoPipeline` explicitly for I2V:
+
+```bash
+python image_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  --model-class-name SanaImageToVideoPipeline \
+  --image cherry_blossom.jpg \
+  --prompt "Cherry blossoms sway in the breeze as petals drift past the camera." \
+  --negative-prompt "blurry, low quality, temporal artifacts" \
+  --height 480 \
+  --width 832 \
+  --num-frames 81 \
+  --num-inference-steps 50 \
+  --guidance-scale 6.0 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 \
+  --seed 42 \
+  --output sana_video_i2v_480p.mp4
+```
+
+Use the 720p checkpoint with `--height 704 --width 1280`. In this example,
+81 frames at 16 FPS means the standard approximately five-second SANA-Video
+checkpoint request. It does not enable the separate LongSANA/LongLive
+minute-scale autoregressive workflow.
+
+The native `SanaImageToVideoPipeline` supports both checkpoint variants. The
+480p path uses vLLM-Omni's `DistributedAutoencoderKLWan` wrapper and the 720p
+path uses `DistributedAutoencoderKLLTX2Video`; both wrap their corresponding
+Diffusers VAE implementations. The native denoising loop intentionally keeps
+Diffusers' checkpoint-compatible `DPMSolverMultistepScheduler`.
+
+The Diffusers adapter has completed end-to-end I2V serving validation at both
+480p and 720p. For online adapter serving, select the matching checkpoint and
+use `WIDTH=1280 HEIGHT=704` for the 720p request. Adapter T2V is also validated
+at both resolutions; see the recipe for the exact backend matrix.
+
+See the [SANA-Video recipe](../../../recipes/NVIDIA/SANA-Video-2B.md) for
+online serving, native/adapter validation boundaries, and hardware notes.
+
 ## Advanced Features
 
 ### CFG Parallel

examples/offline_inference/image_to_video/image_to_video.py

@@ -3,13 +3,14 @@
 
 """
 Image-to-Video generation example using Wan2.2 I2V/TI2V models, LTX2/LTX-2.3,
-HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE.
+HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE.
 
 Supports:
 - Wan2.2-I2V-A14B-Diffusers: MoE model with CLIP image encoder
 - Wan2.2-TI2V-5B-Diffusers: Unified T2V+I2V model (dense 5B)
 - LTX2 image-to-video pipeline
 - HunyuanVideo-1.5 I2V: SigLIP + VAE dual image conditioning
+- SANA-Video 2B: first-frame latent conditioning at 480p or 720p
 - Wan2.1 VACE: first/last-frame, inpainting, and reference conditioning
 
 Usage:
@@ -38,6 +39,13 @@ Usage:
         --image input.jpg --prompt "A cat playing with yarn" \
         --flow-shift 5.0 --guidance-scale 6.0
 
+    # SANA-Video 2B I2V (480p)
+    python image_to_video.py --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+        --model-class-name SanaImageToVideoPipeline \
+        --image input.jpg --prompt "A cat turns toward the camera." \
+        --height 480 --width 832 --num-frames 81 --num-inference-steps 50 \
+        --guidance-scale 6.0
+
     # Cosmos3 I2V (image conditioning)
     python image_to_video.py --model nvidia/Cosmos3-Nano \
         --image input.jpg --prompt "The scene comes to life with smooth, natural motion." \
@@ -107,13 +115,16 @@ def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
         description=(
             "Generate a video from one or more images "
-            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE)."
+            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE)."
         )
     )
     parser.add_argument(
         "--model",
         default="Wan-AI/Wan2.2-I2V-A14B-Diffusers",
-        help="Diffusers I2V model ID or local path (Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, Cosmos3, or Wan2.1 VACE).",
+        help=(
+            "Diffusers I2V model ID or local path "
+            "(Wan2.2, LTX2/LTX-2.3, HunyuanVideo-1.5, SANA-Video, Cosmos3, or Wan2.1 VACE)."
+        ),
     )
     parser.add_argument(
         "--model-class-name",
@@ -377,6 +388,7 @@ def main():
     is_ltx2_distilled = "distilled" in model_class_name_lower or "distilled" in model_name
     is_ltx23 = "ltx23" in model_class_name_lower or "ltx-2.3" in model_name
     is_ltx2 = is_ltx2_distilled or is_ltx23 or "ltx2" in model_class_name_lower or "ltx-2" in model_name
+    is_sana = model_class_name == "SanaImageToVideoPipeline" or "sana-video" in model_name
     is_cosmos = "cosmos" in model_name or (model_class_name is not None and "cosmos" in model_class_name.lower())
     is_cosmos_edge = is_cosmos and ("edge" in model_name or "edge" in model_class_name_lower)
 
@@ -434,6 +446,17 @@ def main():
             512 * 768,
             32,
         )
+    elif is_sana:
+        is_720p = "720p" in model_name
+        d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = (
+            16,
+            6.0,
+            81,
+            50,
+            5.0,
+            (704 * 1280) if is_720p else (480 * 832),
+            32 if is_720p else 16,
+        )
     else:  # Wan2.2 / HunyuanVideo-1.5
         d_fps, d_guidance, d_num_frames, d_steps, d_flow_shift, d_max_area, d_mod = 16, 5.0, 81, 50, 5.0, 480 * 832, 16
 

examples/offline_inference/text_to_video/text_to_video.md

@@ -14,6 +14,8 @@ A unified script for text-to-video generation. Supports multiple models with mod
 | `hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-720p_t2v` | 720x1280 | 121 | 50 | 6.0 | FP8 + VAE tiling required |
 | `nvidia/Cosmos3-Nano` | 720x1280 | 189 | 35 | 6.0 | ~46 GiB (peak, 720p) |
 | `BestWishYsh/Helios-Base` / `Helios-Mid` / `Helios-Distilled` | 384x640 | 99 | 50 | 5.0 / 5.0 / 1.0 | — |
+| `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` | 480x832 | 81 | 50 | 6.0 | BF16 DiT + FP32 Wan VAE wrapper |
+| `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` | 704x1280 | 81 | 50 | 6.0 | BF16 DiT + LTX-2 Video VAE wrapper |
 
 ## Local CLI Usage
 
@@ -151,6 +153,40 @@ python text_to_video.py \
   --output quick_test.mp4
 ```
 
+### SANA-Video-2B
+
+Select the native T2V pipeline explicitly:
+
+```bash
+python text_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  --model-class-name SanaVideoPipeline \
+  --prompt "A cinematic tracking shot of a sailboat crossing the ocean at sunset." \
+  --height 480 \
+  --width 832 \
+  --num-frames 81 \
+  --num-inference-steps 50 \
+  --guidance-scale 6.0 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 \
+  --output sana_video_480p.mp4
+```
+
+For the 720p checkpoint, switch the model to
+`Efficient-Large-Model/SANA-Video_2B_720p_diffusers`, use
+`--height 704 --width 1280`, and write to a different output path. Both native
+checkpoint variants are supported. The Diffusers adapter T2V compatibility
+path is also validated at both resolutions through online serving.
+
+The native 480p path uses vLLM-Omni's `DistributedAutoencoderKLWan` wrapper;
+the native 720p path uses `DistributedAutoencoderKLLTX2Video`. Both wrappers
+retain their underlying Diffusers autoencoders. The pipeline intentionally
+uses Diffusers' checkpoint-compatible `DPMSolverMultistepScheduler`.
+
+For image-to-video, use the shared
+[`image_to_video.py`](../image_to_video/README.md#sana-video-2b) example with
+`--model-class-name SanaImageToVideoPipeline` and `--image <path>`.
+
 ### Cosmos3
 
 ```bash

examples/offline_inference/text_to_video/text_to_video.py

@@ -114,12 +114,32 @@ _MODEL_PRESETS = {
         "fps": 24,
         "output": "ltx23_output.mp4",
     },
+    "sana_480p": {
+        "height": 480,
+        "width": 832,
+        "num_frames": 81,
+        "num_inference_steps": 50,
+        "guidance_scale": 6.0,
+        "fps": 16,
+        "output": "sana_video_480p.mp4",
+    },
+    "sana_720p": {
+        "height": 704,
+        "width": 1280,
+        "num_frames": 81,
+        "num_inference_steps": 50,
+        "guidance_scale": 6.0,
+        "fps": 16,
+        "output": "sana_video_720p.mp4",
+    },
 }
 
 
 def _detect_preset(model: str, model_class_name: str | None = None) -> dict:
     model_lower = model.lower()
     class_lower = (model_class_name or "").lower()
+    if "sana-video" in model_lower or "sana_video" in model_lower or "sanavideo" in class_lower:
+        return _MODEL_PRESETS["sana_720p" if "720p" in model_lower else "sana_480p"]
     if "lingbot" in model_lower or "lingbotvideo" in class_lower:
         return _MODEL_PRESETS["lingbot"]
     if "ltx" in class_lower or "ltx" in model_lower:

examples/offline_inference/x_to_video_audio/download_dreamid_omni.py

@@ -1,108 +0,0 @@
-import argparse
-import fcntl
-import os
-import site
-import subprocess
-import tempfile
-import time
-from pathlib import Path
-
-from huggingface_hub import snapshot_download
-
-DEPENDENCY_REPO = "https://github.com/bytedance/DreamID-V.git"
-DEPENDENCY_BRANCH = "omni"
-CACHE_DIR = Path(tempfile.gettempdir()) / "vllm-omni-dependency"
-LOCK_FILE = CACHE_DIR / ".install.lock"
-DEPENDENCY_DIR = CACHE_DIR / "DreamID-Omni"
-
-
-def download_dependency():
-    CACHE_DIR.mkdir(parents=True, exist_ok=True)
-
-    with open(LOCK_FILE, "w") as f:
-        fcntl.flock(f, fcntl.LOCK_EX)
-        if not DEPENDENCY_DIR.exists():
-            print(f"Downloading DreamID-Omni to {DEPENDENCY_DIR} ...")
-            subprocess.run(
-                ["git", "clone", "--depth", "1", DEPENDENCY_REPO, "--branch", DEPENDENCY_BRANCH, str(DEPENDENCY_DIR)],
-                check=True,
-            )
-            print("Download finished.")
-        fcntl.flock(f, fcntl.LOCK_UN)
-
-    # write .pth to site-packages
-    site_packages = Path(site.getsitepackages()[0])
-    pth_file = site_packages / "vllm_omni_dependency.pth"
-    pth_file.write_text(str(DEPENDENCY_DIR))
-    print(f"Added {DEPENDENCY_DIR} to site-packages via {pth_file}")
-
-
-def timed_download(repo_id: str, local_dir: str, allow_patterns: list | None = None):
-    """Download files from HF repo and log time + destination."""
-    if os.path.exists(local_dir):
-        print(f"Directory {local_dir} already exists. Skipping download.")
-        return
-    print(f"Starting download from {repo_id} into {local_dir}")
-    start_time = time.time()
-
-    snapshot_download(
-        repo_id=repo_id,
-        local_dir=local_dir,
-        local_dir_use_symlinks=False,
-        allow_patterns=allow_patterns,
-    )
-
-    elapsed = time.time() - start_time
-    print(f"✅ Finished downloading {repo_id} in {elapsed:.2f} seconds. Files saved at: {local_dir}")
-
-
-def main(output_dir: str):
-    # Wan2.2
-    wan_dir = os.path.join(output_dir, "Wan2.2-TI2V-5B")
-    timed_download(
-        repo_id="Wan-AI/Wan2.2-TI2V-5B",
-        local_dir=wan_dir,
-        allow_patterns=["google/*", "models_t5_umt5-xxl-enc-bf16.pth", "Wan2.2_VAE.pth"],
-    )
-
-    # MMAudio
-    mm_audio_dir = os.path.join(output_dir, "MMAudio")
-    timed_download(
-        repo_id="hkchengrex/MMAudio",
-        local_dir=mm_audio_dir,
-        allow_patterns=["ext_weights/best_netG.pt", "ext_weights/v1-16.pth"],
-    )
-
-    dreamid_dir = os.path.join(output_dir, "DreamID-Omni")
-
-    timed_download(repo_id="XuGuo699/DreamID-Omni", local_dir=dreamid_dir)
-
-    # Now we construct the config file
-    import json
-
-    data = {
-        "_class_name": "DreamIDOmniPipeline",
-    }
-
-    with open(os.path.join(output_dir, "model_index.json"), "w", encoding="utf-8") as f:
-        json.dump(data, f, indent=2)
-
-    print(f"model_index.json created at {os.path.join(output_dir, 'model_index.json')}")
-
-    transformer_dir = os.path.join(output_dir, "transformer")
-    os.makedirs(transformer_dir, exist_ok=True)
-    with open(os.path.join(transformer_dir, "config.json"), "w", encoding="utf-8") as f:
-        json.dump({"fusion": "DreamID-Omni/dreamid_omni.safetensors"}, f)
-    print(f"transformer/config.json created at {os.path.join(transformer_dir, 'config.json')}")
-
-    # now we download the dependency code
-    download_dependency()
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Download models from Hugging Face")
-    parser.add_argument(
-        "--output-dir", type=str, default="./dreamid_omni", help="Base directory to save downloaded models"
-    )
-    args = parser.parse_args()
-    main(args.output_dir)

examples/offline_inference/x_to_video_audio/x_to_video_audio.md

@@ -1,163 +0,0 @@
-# X-To-Video-Audio
-
-The `DreamID-Omni` pipeline generates short videos from text, image and video.
-
-## Local CLI Usage
-### Download the Model locally
-Since DreamID-Omni combine multiple models, and without any config, so we need to download them locally.
-
-```bash
-python download_dreamid_omni.py --output-dir ./dreamid_omni
-```
-After download, the model directory will look like this:
-
-```
-dreamid_omni/
-├── DreamID-Omni/
-│   ├── dreamid_omni.safetensors
-├── MMAudio/
-│   ├── ext_weights/
-│   │   ├── best_netG.pt
-│   │   ├── v1-16.pth
-├── Wan2.2-TI2V-5B/
-│   ├── google/*
-│   ├── models_t5_umt5-xxl-enc-bf16.pth
-│   ├── Wan2.2_VAE.pth
-│
-├── model_index.json
-└── transformer/
-    └── config.json   # create by download_dreamid_omni.py
-```
-
-### Run the Inference
-```python
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "Two people walking together and singing happily" \
-  --image-path ./example0.png ./example1.png \
-  --audio-path ./example0.wav ./example1.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --output out_dreamid_omni_twoip.mp4
-```
-In the current test scenario (2 images + 2 audio inputs), the VRAM requirement is 72GB, regardless of whether cfg-parallel is enabled or disabled.
-The VRAM usage can be reduced by enabling CPU offload via --enable-cpu-offload.
-For multi-GPU memory reduction on the fused DreamID-Omni transformer, you can also enable HSDP:
-
-```python
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "Two people walking together and singing happily" \
-  --image-path ./example0.png ./example1.png \
-  --audio-path ./example0.wav ./example1.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --use-hsdp \
-  --hsdp-shard-size 2 \
-  --output out_dreamid_omni_twoip.mp4
-```
-
-
-You could take reference images/audios from the test cases in the official repo: https://github.com/Guoxu1233/DreamID-Omni
-
-For example, single IP ref resources can be found under https://github.com/Guoxu1233/DreamID-Omni/tree/main/test_case/oneip, you could download them correspondingly to your local and use them for testing.
-
-```python
-# Example usage for oneip, ref media from the official repo DreamID-Omni
-python x_to_video_audio.py \
-  --model /path/to/dreamid_omni \
-  --prompt "<img1>: In the frame, a woman with black long hair is identified as <sub1>.\n**Overall Environment/Scene**: A lively open-kitchen café at night; stove flames flare, steam rises, and warm pendant lights swing slightly as staff move behind her. The shot is an upper-body close-up.\n**Main Characters/Subjects Appearance**: <sub1> is a young woman with thick dark wavy hair and a side part. She wears a fitted black top under a light apron, a thin gold chain necklace, and small stud earrings.\n**Main Characters/Subjects Actions**: <sub1> tastes the sauce with a spoon, then turns her face toward the camera while still holding the spoon, her expression shifting from focused to conflicted.\n<sub1> maintains eye contact, swallows as if choosing her words, and says, <S>I keep telling myself I’m fine,but some nights it feels like I’m just performing calm.<E>" \
-  --image-path 9.png \
-  --audio-path 9.wav \
-  --video-negative-prompt "jitter, bad hands, blur, distortion" \
-  --audio-negative-prompt "robotic, muffled, echo, distorted" \
-  --cfg-parallel-size 2 \
-  --num-inference-steps 45 \
-  --height 704 \
-  --width 1280 \
-  --output out_dreamid_omni_oneip.mp4
-```
-
-
-## MagiHuman (`--model-type magi-human`)
-
-MagiHuman is a text → video+audio model with a DiT MoE backbone and a ~9B-param
-T5Gemma text encoder. A detailed text prompt is the only required input; an optional
-image and/or audio file may be supplied for conditioning. Natively supports Tensor
-Parallelism. For an 80GB node, `--tensor-parallel-size 4` is recommended to shard
-the MoE weights and the text encoder.
-
-> Install [MagiCompiler](https://github.com/SandAI-org/MagiCompiler) for correct
-> attention-kernel behaviour (the pipeline otherwise falls back to stubs).
-
-Text-only generation:
-
-```bash
-python x_to_video_audio.py \
-  --model-type magi-human \
-  --model /path/to/daVinci-MagiHuman \
-  --prompt "A young woman with long, wavy golden blonde hair... <dialogue and background sound>" \
-  --tensor-parallel-size 4 \
-  --height 256 --width 448 \
-  --num-inference-steps 8 \
-  --seed 52 \
-  --extra-body '{"seconds": 5, "sr_height": 1080, "sr_width": 1920, "sr_num_inference_steps": 5}' \
-  --output output_magihuman.mp4
-```
-
-With optional image and audio conditioning:
-
-```bash
-python x_to_video_audio.py \
-  --model-type magi-human \
-  --model /path/to/daVinci-MagiHuman \
-  --prompt "A young woman..." \
-  --tensor-parallel-size 4 \
-  --height 256 --width 448 \
-  --num-inference-steps 8 \
-  --seed 52 \
-  --extra-body '{"image_path": "/path/to/ref.jpg", "audio_path": "/path/to/ref.wav", "sr_height": 1080, "sr_width": 1920, "sr_num_inference_steps": 5}' \
-  --output output_magihuman.mp4
-```
-
-MagiHuman-specific arguments are passed as a JSON dict via `--extra-body` (declared
-in `vllm_omni/model_extras/magi_human.py`, routed via `extra_args`):
-
-- `seconds`: output duration in seconds (default 10; ignored when `audio_path` is set,
-  because the audio length then determines the number of frames).
-- `audio_path`: path to an audio file for audio-to-video conditioning. When provided,
-  the audio drives the frame count and is encoded as a latent condition into the DiT
-  (`is_a2v` mode). Omit for pure text-to-video+audio generation.
-- `image_path`: path to an image file for visual conditioning. Applied at both the
-  base-resolution (BR) and super-resolution (SR) stages when SR is enabled.
-- `sr_height` / `sr_width`: super-resolution output resolution. SR stage is skipped
-  when these are omitted.
-- `sr_num_inference_steps`: denoising steps for the SR stage.
-
-## DreamID-Omni arguments
-
-Key arguments:
-- `--prompt`: text description (string).
-- `--model`: path to the model local directory.
-- `--height/--width`: output resolution (defaults 704 * 1024).
-- `--image-path`: path to the input image list.
-- `--audio-path`: path to the input audio list, indicate the timbre of the output video.
-- `--cfg-parallel-size`: number of parallel cfg parallel (defaults 1).
-- `--use-hsdp`: enable HSDP weight sharding for DreamID-Omni fused blocks.
-- `--hsdp-shard-size`: number of GPUs used for HSDP sharding.
-- `--hsdp-replicate-size`: number of HSDP replica groups.
-- `--num-inference-steps`: number of denoising steps (defaults 45).
-- `--video-negative-prompt`: negative prompt for video generation.
-- `--audio-negative-prompt`: negative prompt for audio generation.
-- `--enable-cpu-offload`: enable CPU offload (defaults False).
-- `--cache-backend`: enable `cache_dit` for acceleration.
-- `--quantization`: online (dynamic) quantization method — `fp8` or `int8`.
-  (VAEs, the T5 text encoder, norms and modulation stay bf16.)

examples/offline_inference/x_to_video_audio/x_to_video_audio.py

@@ -1,279 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import argparse
-import json
-import re
-import time
-from typing import Any
-
-import numpy as np
-from PIL import Image
-from vllm.multimodal.media.audio import load_audio
-
-from vllm_omni.diffusion.utils.media_utils import mux_video_audio_bytes
-from vllm_omni.entrypoints.omni import Omni
-from vllm_omni.inputs.data import OmniDiffusionSamplingParams
-
-IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".webp", ".gif"}
-AUDIO_EXTENSIONS = {".wav", ".mp3", ".flac", ".m4a", ".aac", ".ogg"}
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description="Offline inference for X -> video+audio models.")
-    parser.add_argument("--model", required=True, help="Model ckpt root directory.")
-    parser.add_argument(
-        "--model-type",
-        default="dreamid-omni",
-        choices=["dreamid-omni", "magi-human"],
-        help="Model type.",
-    )
-    parser.add_argument("--prompt", default=None, help="Text prompt.")
-
-    parser.add_argument("--image-path", type=str, nargs="+", help="list of image-path")
-    parser.add_argument("--audio-path", type=str, nargs="+", help="list of audio-path")
-    parser.add_argument("--prompt-file", type=str, default=None, help="Text prompt in json format.")
-    parser.add_argument(
-        "--extra-body",
-        type=str,
-        default=None,
-        help="[magi-human] JSON dict of model-specific extra params (declared in vllm_omni/model_extras/), "
-        'merged into sampling extra_args. Example: \'{"image_path": "/path/to/img.jpg", "seconds": 5}\'.',
-    )
-
-    parser.add_argument("--height", type=int, default=704, help="Video height.")
-    parser.add_argument("--width", type=int, default=1280, help="Video width.")
-    parser.add_argument("--num-inference-steps", type=int, default=45, help="Sampling steps.")
-    parser.add_argument("--solver-name", default="unipc", help="Solver name: unipc|dpm++|euler.")
-    parser.add_argument("--shift", type=float, default=5.0, help="Scheduler shift.")
-    parser.add_argument("--seed", type=int, default=103, help="Random seed for reproducible generation.")
-    parser.add_argument(
-        "--cfg-parallel-size",
-        type=int,
-        default=1,
-        choices=[1, 2, 3, 4],
-        help="Number of GPUs used for classifier free guidance parallel size (max 4 branches).",
-    )
-    parser.add_argument(
-        "--tensor-parallel-size",
-        type=int,
-        default=1,
-        help=("Number of GPUs used for tensor parallelism (TP) inside the DiT."),
-    )
-    parser.add_argument(
-        "--video-negative-prompt",
-        default="jitter, bad hands, blur, distortion",
-        help="Negative prompt for video.",
-    )
-    parser.add_argument(
-        "--audio-negative-prompt",
-        default="robotic, muffled, echo, distorted",
-        help="Negative prompt for audio.",
-    )
-    parser.add_argument("--output", default="dreamid_output.mp4", help="Output video path.")
-    parser.add_argument(
-        "--quantization",
-        type=str,
-        default=None,
-        choices=["fp8", "int8"],
-        help=("Online (dynamic) quantization method for the DreamID-Omni transformer. "),
-    )
-    parser.add_argument(
-        "--enable-cpu-offload",
-        action="store_true",
-        default=False,
-        help="Enable CPU offloading for diffusion models.",
-    )
-    parser.add_argument(
-        "--enable-layerwise-offload",
-        action="store_true",
-        help="Enable layerwise (blockwise) offloading on DiT modules.",
-    )
-    parser.add_argument("--cache-backend", type=str, default=None, choices=["cache_dit"], help="Cache backend.")
-    parser.add_argument(
-        "--use-hsdp",
-        action="store_true",
-        help="Enable HSDP for DreamID-Omni fused transformer blocks.",
-    )
-    parser.add_argument(
-        "--hsdp-shard-size",
-        type=int,
-        default=1,
-        help="Number of GPUs used for HSDP sharding when HSDP is enabled.",
-    )
-    parser.add_argument(
-        "--hsdp-replicate-size",
-        type=int,
-        default=1,
-        help="Number of HSDP replica groups. Default 1 means pure sharding.",
-    )
-    return parser.parse_args()
-
-
-def load_image_and_audio(image_paths, audio_paths):
-    image = []
-    audio = []
-
-    for path in image_paths:
-        with Image.open(path) as img:
-            img = img.convert("RGB")
-            image.append(img)
-
-    for path in audio_paths:
-        audio_array, sr = load_audio(path, sr=16000)
-        audio_array = audio_array[int(sr * 1) : int(sr * 3)]
-        audio.append(audio_array)
-    return image, audio
-
-
-def _extract_peak_memory_mb(result: Any) -> float:
-    """Pull worker-reported peak VRAM (MiB) generation result.
-
-    Mirrors vllm_omni/entrypoints/openai/serving_video.py:_extract_peak_memory_mb.
-    """
-    if isinstance(result, list):
-        result = result[0] if result else None
-    if result is None:
-        return 0.0
-    val = getattr(result, "peak_memory_mb", 0.0)
-    if not val:
-        inner = result
-        if isinstance(inner, list):
-            inner = inner[0] if inner else None
-        val = getattr(inner, "peak_memory_mb", 0.0)
-    try:
-        return float(val or 0.0)
-    except (TypeError, ValueError):
-        return 0.0
-
-
-def main() -> None:
-    args = parse_args()
-    if args.prompt is None and args.prompt_file is None:
-        raise ValueError("Either --prompt or --prompt-file must be provided.")
-
-    text_prompt = args.prompt
-    if args.prompt_file:
-        with open(args.prompt_file) as f:
-            text_prompt = json.load(f)
-            text_prompt = re.sub(
-                r"\[SPEAKER_TIMESTAMPS_START\].*?\[SPEAKER_TIMESTAMPS_END\]", "", text_prompt, flags=re.DOTALL
-            ).strip()
-            text_prompt = re.sub(
-                r"\[AUDIO_DESCRIPTION_START].*?\[AUDIO_DESCRIPTION_END]", "", text_prompt, flags=re.DOTALL
-            ).strip()
-            text_prompt = re.sub(r"\[[A-Z_]+\]", "", text_prompt)
-            text_prompt = re.sub(r"\n\s*\n", "\n", text_prompt).strip()
-
-    if args.model_type == "magi-human":
-        prompt = text_prompt
-        extra_args = json.loads(args.extra_body) if args.extra_body else {}
-        sampling_params = OmniDiffusionSamplingParams(
-            height=args.height,
-            width=args.width,
-            num_inference_steps=args.num_inference_steps,
-            seed=args.seed,
-            extra_args=extra_args,
-        )
-    else:
-        image, audio = load_image_and_audio(args.image_path, args.audio_path)
-
-        prompt = {
-            "prompt": text_prompt,
-            "video_negative_prompt": args.video_negative_prompt,
-            "audio_negative_prompt": args.audio_negative_prompt,
-            "multi_modal_data": {"image": image, "audio": audio},
-        }
-
-        sampling_params = OmniDiffusionSamplingParams(
-            height=args.height,
-            width=args.width,
-            num_inference_steps=args.num_inference_steps,
-            seed=args.seed,
-            extra_args={
-                "solver_name": args.solver_name,
-                "shift": args.shift,
-            },
-        )
-
-    cache_config = None
-    if args.cache_backend == "cache_dit":
-        cache_config = {
-            "Fn_compute_blocks": 1,
-            "Bn_compute_blocks": 0,
-            "max_warmup_steps": 4,
-            "max_cached_steps": 20,
-            "residual_diff_threshold": 0.24,
-            "max_continuous_cached_steps": 3,
-            "enable_taylorseer": False,
-            "taylorseer_order": 1,
-            "scm_steps_mask_policy": None,
-            "scm_steps_policy": "dynamic",
-        }
-
-    omni_kwargs = dict(
-        model=args.model,
-        cfg_parallel_size=args.cfg_parallel_size,
-        tensor_parallel_size=args.tensor_parallel_size,
-        use_hsdp=args.use_hsdp,
-        hsdp_shard_size=args.hsdp_shard_size,
-        hsdp_replicate_size=args.hsdp_replicate_size,
-        model_type=args.model_type,
-        enable_cpu_offload=args.enable_cpu_offload,
-        enable_layerwise_offload=args.enable_layerwise_offload,
-        cache_backend=args.cache_backend,
-        cache_config=cache_config,
-    )
-    if args.quantization is not None:
-        omni_kwargs["quantization"] = args.quantization
-    omni = Omni(**omni_kwargs)
-    start = time.perf_counter()
-    outputs = omni.generate(prompt, sampling_params)
-    elapsed = time.perf_counter() - start
-
-    if not outputs:
-        raise RuntimeError("No output returned from the model.")
-    result = outputs[0]
-    if not result.images:
-        raise RuntimeError("No video frames found in model output.")
-    generated_video = result.images[0]
-    mm = result.multimodal_output or {}
-    generated_audio = mm.get("audio")
-
-    if args.model_type == "magi-human":
-        # MagiHuman returns frames already as (F, H, W, C) uint8, ready to mux.
-        fps = float(mm.get("fps", 25))
-        sample_rate = int(mm.get("audio_sample_rate", 24000))
-        frames = generated_video
-    else:
-        fps = float(mm.get("fps", 24))
-        sample_rate = int(mm.get("audio_sample_rate", 16000))
-        # DreamID-Omni returns video as (C, F, H, W) float32 in [-1, 1].
-        # mux_video_audio_bytes expects (F, H, W, C) uint8.
-        if not isinstance(generated_video, np.ndarray) or generated_video.ndim != 4:
-            raise RuntimeError(f"Unexpected video shape: {getattr(generated_video, 'shape', None)}")
-        frames = generated_video.transpose(1, 2, 3, 0)
-        frames = (np.clip((frames + 1.0) / 2.0, 0.0, 1.0) * 255.0).round().astype(np.uint8)
-
-    audio_np = None
-    if generated_audio is not None:
-        audio_np = np.squeeze(np.asarray(generated_audio)).astype(np.float32)
-
-    output_path = args.output
-    video_bytes = mux_video_audio_bytes(
-        frames,
-        audio_np,
-        fps=float(fps),
-        audio_sample_rate=sample_rate,
-    )
-    with open(output_path, "wb") as f:
-        f.write(video_bytes)
-    print(f"Saved generated video to {output_path}")
-    print(f"Total time: {elapsed:.2f}s")
-    peak_mb = _extract_peak_memory_mb(outputs)
-    if peak_mb:
-        print(f"Worker peak GPU memory (reserved): {peak_mb:.2f} MiB ({peak_mb / 1024:.2f} GiB)")
-
-
-if __name__ == "__main__":
-    main()

examples/online_serving/image_to_video/README.md

@@ -1,5 +1,19 @@
 # Image-To-Video
 
+This example demonstrates how to deploy image-to-video models for online
+video generation using vLLM-Omni.
+
+## Supported Models
+
+| Model | Model ID |
+|-------|----------|
+| Wan2.2 I2V | `Wan-AI/Wan2.2-I2V-A14B-Diffusers` |
+| SANA-Video 2B (480p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` |
+| SANA-Video 2B (720p) | `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` |
+
+For SANA-Video 2B image-to-video online serving, see the
+[SANA-Video 2B recipe](../../../recipes/NVIDIA/SANA-Video-2B.md).
+
 This example demonstrates online image-to-video generation with vLLM-Omni.
 The startup script defaults to Wan2.2, but `MODEL` can select any supported
 image-to-video model. The existing curl helper remains a concrete Wan2.2

examples/online_serving/image_to_video/run_curl_sana_video.sh

@@ -0,0 +1,67 @@
+#!/bin/bash
+# Submit and download a standard SANA-Video-2B image-to-video request.
+
+set -euo pipefail
+
+INPUT_IMAGE="${INPUT_IMAGE:-}"
+BASE_URL="${BASE_URL:-http://localhost:8099}"
+OUTPUT_PATH="${OUTPUT_PATH:-sana_video_i2v.mp4}"
+WIDTH="${WIDTH:-832}"
+HEIGHT="${HEIGHT:-480}"
+POLL_INTERVAL="${POLL_INTERVAL:-2}"
+
+if [ -z "$INPUT_IMAGE" ] || [ ! -f "$INPUT_IMAGE" ]; then
+    echo "Set INPUT_IMAGE to an existing image file."
+    exit 1
+fi
+
+create_response="$(
+    curl -sS -X POST "${BASE_URL}/v1/videos" \
+        -H "Accept: application/json" \
+        -F "prompt=A cat turns toward the camera with smooth, natural motion." \
+        -F "negative_prompt=blurry, low quality, temporal artifacts" \
+        -F "input_reference=@${INPUT_IMAGE}" \
+        -F "width=${WIDTH}" \
+        -F "height=${HEIGHT}" \
+        -F "num_frames=81" \
+        -F "fps=16" \
+        -F "num_inference_steps=50" \
+        -F "guidance_scale=6.0" \
+        -F "seed=42" \
+        -F 'extra_params={"motion_score":30}'
+)"
+
+video_id="$(echo "${create_response}" | jq -r '.id')"
+if [ -z "${video_id}" ] || [ "${video_id}" = "null" ]; then
+    echo "Failed to create SANA-Video job:"
+    echo "${create_response}" | jq .
+    exit 1
+fi
+
+while true; do
+    status_response="$(curl -sS "${BASE_URL}/v1/videos/${video_id}")"
+    status="$(echo "${status_response}" | jq -r '.status')"
+
+    case "${status}" in
+        queued|in_progress)
+            echo "Video job ${video_id} status: ${status}"
+            sleep "${POLL_INTERVAL}"
+            ;;
+        completed)
+            break
+            ;;
+        failed)
+            echo "SANA-Video generation failed:"
+            echo "${status_response}" | jq .
+            exit 1
+            ;;
+        *)
+            echo "Unexpected status response:"
+            echo "${status_response}" | jq .
+            exit 1
+            ;;
+    esac
+done
+
+curl -sS -L "${BASE_URL}/v1/videos/${video_id}/content" -o "${OUTPUT_PATH}"
+echo "Saved SANA-Video to ${OUTPUT_PATH}"

examples/online_serving/image_to_video/run_server_sana_video.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+# SANA-Video-2B native image-to-video serving.
+
+set -euo pipefail
+
+MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
+PORT="${PORT:-8099}"
+
+vllm serve "$MODEL" \
+    --omni \
+    --model-class-name SanaImageToVideoPipeline \
+    --dtype bfloat16 \
+    --port "$PORT"

examples/online_serving/image_to_video/run_server_sana_video_diffusers.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+# SANA-Video-2B image-to-video serving through the Diffusers adapter.
+# Both 480p and 720p checkpoints are validated. For 720p:
+# MODEL=Efficient-Large-Model/SANA-Video_2B_720p_diffusers bash run_server_sana_video_diffusers.sh
+# INPUT_IMAGE=input.jpg WIDTH=1280 HEIGHT=704 bash run_curl_sana_video.sh
+
+set -euo pipefail
+
+MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
+PORT="${PORT:-8099}"
+
+vllm serve "$MODEL" \
+    --omni \
+    --model-class-name SanaImageToVideoPipeline \
+    --diffusion-load-format diffusers \
+    --diffusion-attention-backend TORCH_SDPA \
+    --dtype bfloat16 \
+    --port "$PORT"

examples/online_serving/minicpmo/README.md

@@ -32,12 +32,14 @@ GPU 0 (90%) and colocates the Talker (55%) and Code2Wav (35%) on GPU 1. That
 profile admits at most four concurrent sequences per stage.
 
 | deploy config | GPUs | Notes |
-|---|---|---|
+| --- | --- | --- |
 | `minicpmo_4_5.yaml` (default) | 1 | Memory-constrained compatibility layout. |
 | `minicpmo_4_5_2gpu.yaml` | 2 | Recommended continuous-batching layout; Talker and Code2Wav share GPU 1. |
 | `minicpmo_4_5_3gpu.yaml` | 3 | One GPU per stage. |
 | `minicpmo_4_5_8x4090.yaml` | 8 | Full 8x4090 layout. |
-| `minicpmo_4_5_duplex.yaml` | 1 | Experimental native full-duplex overlay. |
+
+Every profile sets `session_mode: duplex`, so native full-duplex is served
+from the same process as `/v1/chat/completions`.
 
 The split pipeline preserves native-duplex epoch/turn identity, segment text,
 turn completion, reference voice, and terminal-audio metadata through
@@ -55,8 +57,8 @@ vllm-omni serve openbmb/MiniCPM-o-4_5 \
 ```
 
 For local ModelScope checkpoints, replace `openbmb/MiniCPM-o-4_5` with the
-checkpoint path. To start the experimental native duplex backend, use
-`vllm_omni/deploy/minicpmo_4_5_duplex.yaml`.
+checkpoint path. For native full-duplex, connect `/v1/realtime?duplex=1` on
+this same server.
 
 ### Per-stage overrides
 
@@ -145,7 +147,7 @@ assistant template, so it is not an apples-to-apples accuracy run.
 
 ## Run the Realtime duplex CLI demo
 
-After the duplex backend is running, stream one WAV through the Realtime
+After the server is running, stream one WAV through the Realtime
 WebSocket endpoint:
 
 ```bash

examples/online_serving/minicpmo/realtime_duplex_demo.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """Minimal single-input MiniCPM-o 4.5 Realtime duplex demo.
 
 Run this after starting the duplex server. Strict lifecycle, overlap, and
@@ -19,8 +22,8 @@ if str(REPO_ROOT) not in sys.path:
     sys.path.insert(0, str(REPO_ROOT))
 
 from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
-    PCM16_BYTES_PER_SAMPLE,
-    PCM16_SAMPLE_RATE,
+    PCM16_BYTES_PER_SAMPLE,  # noqa: F401 - compatibility export
+    PCM16_SAMPLE_RATE,  # noqa: F401 - compatibility export
     RealtimeDuplexClient,
     RealtimeEventCollector,
     build_realtime_url,
@@ -28,6 +31,13 @@ from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
     wait_for,
     write_pcm16_wav,
 )
+from vllm_omni.experimental.fullduplex.client import chunk_period_ms as _chunk_period_ms  # noqa: E402
+from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
+    has_residual_model_unit as _has_residual_model_unit,
+)
+from vllm_omni.experimental.fullduplex.client import (  # noqa: E402
+    reference_audio_data_url as _ref_audio_data_url,
+)
 
 
 class _StreamingOutputWriter:
@@ -133,25 +143,6 @@ def _latest_model_decision(
     return decision
 
 
-def _chunk_period_ms(events: list[dict[str, object]]) -> int:
-    for event in reversed(events):
-        session = event.get("session")
-        if not isinstance(session, dict):
-            continue
-        capabilities = session.get("capabilities")
-        if not isinstance(capabilities, dict):
-            continue
-        chunk_period_ms = capabilities.get("chunk_period_ms")
-        if isinstance(chunk_period_ms, int) and chunk_period_ms > 0:
-            return chunk_period_ms
-    return 1000
-
-
-def _has_residual_model_unit(pcm16: bytes, *, chunk_period_ms: int) -> bool:
-    unit_bytes = PCM16_SAMPLE_RATE * PCM16_BYTES_PER_SAMPLE * chunk_period_ms // 1000
-    return bool(unit_bytes > 0 and len(pcm16) % unit_bytes)
-
-
 def _response_in_progress(events: list[dict[str, object]]) -> bool:
     return sum(event.get("type") == "response.created" for event in events) > sum(
         event.get("type") == "response.done" for event in events
@@ -168,13 +159,6 @@ def _event_count_after(
     return sum(event.get("type") == event_type for event in events[index + 1 :])
 
 
-def _ref_audio_data_url(path: str | None) -> str | None:
-    if path is None:
-        return None
-    ref_path = Path(path).expanduser()
-    return "data:audio/wav;base64," + base64.b64encode(ref_path.read_bytes()).decode("ascii")
-
-
 async def run_demo(args: argparse.Namespace) -> dict[str, object]:
     input_pcm16 = read_pcm16_wav(Path(args.input_wav))
     if not input_pcm16:

examples/online_serving/stable_audio/README.md

@@ -1,234 +0,0 @@
-# Stable Audio Online Serving
-
-Generate audio from text prompts using Stable Audio models via an OpenAI-compatible API endpoint.
-
-## Features
-
-- **OpenAI-compatible API**: Use `/v1/audio/generate` endpoint
-- **Flexible control**: Adjust audio length, guidance scale, inference steps
-- **Quality control**: Use negative prompts to avoid unwanted characteristics
-- **Reproducible**: Set random seed for deterministic generation
-
-## Quick Start
-
-### 1. Start the Server
-
-```bash
-vllm-omni serve stabilityai/stable-audio-open-1.0 \
-    --host 0.0.0.0 \
-    --port 8091 \
-    --gpu-memory-utilization 0.9 \
-    --trust-remote-code \
-    --enforce-eager \
-    --omni
-```
-
-### 2. Generate Audio
-
-#### Using curl
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "The sound of a cat purring",
-        "audio_length": 10.0
-    }' --output cat.wav
-```
-
-#### Using Python Client
-
-```bash
-python stable_audio_client.py \
-    --text "The sound of a cat purring" \
-    --audio_length 10.0 \
-    --output cat.wav
-```
-
-#### Using Bash Script
-
-```bash
-bash curl_examples.sh
-```
-
-## API Reference
-
-### Endpoint
-
-```
-POST /v1/audio/generate
-```
-
-### Request Body
-
-```json
-{
-    "input": "Text description of the audio",
-    "audio_length": 10.0,
-    "audio_start": 0.0,
-    "negative_prompt": "Low quality",
-    "guidance_scale": 7.0,
-    "num_inference_steps": 100,
-    "seed": 42,
-    "response_format": "wav"
-}
-```
-
-### Parameters
-
-| Parameter | Type | Default | Description |
-|-----------|------|---------|-------------|
-| `input` | string | **required** | Text prompt describing the audio to generate |
-| `audio_length` | float | ~47s | Audio duration in seconds (max ~47s for stable-audio-open-1.0) |
-| `audio_start` | float | 0.0 | Audio start time in seconds |
-| `negative_prompt` | string | null | Text describing what to avoid in generation |
-| `guidance_scale` | float | 7.0 | Classifier-free guidance scale (higher = more adherence to prompt) |
-| `num_inference_steps` | int | 50 | Number of denoising steps (higher = better quality, slower) |
-| `seed` | int | null | Random seed for reproducibility |
-| `response_format` | string | "wav" | Output format: wav, mp3, flac, pcm |
-
-### Response
-
-Returns audio data in the requested format (default: WAV).
-
-## Usage Examples
-
-### Basic Generation
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "The sound of ocean waves"
-    }' --output ocean.wav
-```
-
-### Custom Duration
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "A dog barking",
-        "audio_length": 5.0
-    }' --output dog_5s.wav
-```
-
-### High Quality with Negative Prompt
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "A piano playing a gentle melody",
-        "audio_length": 10.0,
-        "negative_prompt": "Low quality, distorted, noisy",
-        "guidance_scale": 8.0,
-        "num_inference_steps": 150
-    }' --output piano_hq.wav
-```
-
-### Reproducible Generation
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "Thunder and rain sounds",
-        "audio_length": 15.0,
-        "seed": 42
-    }' --output thunder.wav
-```
-
-### Quick Generation (Fewer Steps)
-
-For faster generation with slightly lower quality:
-
-```bash
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "Birds chirping in a forest",
-        "audio_length": 8.0,
-        "num_inference_steps": 50
-    }' --output birds_quick.wav
-```
-
-## Python Client Examples
-
-### Simple Generation
-
-```bash
-python stable_audio_client.py \
-    --text "The sound of a cat purring"
-```
-
-### Custom Parameters
-
-```bash
-python stable_audio_client.py \
-    --text "Thunder and rain" \
-    --audio_length 15.0 \
-    --negative_prompt "Low quality" \
-    --guidance_scale 7.0 \
-    --num_inference_steps 100 \
-    --seed 42 \
-    --output thunder.wav
-```
-
-### Different Output Format
-
-```bash
-python stable_audio_client.py \
-    --text "Guitar playing" \
-    --response_format mp3 \
-    --output guitar.mp3
-```
-
-## Tips
-
-1. **Audio Length**: Keep under 47 seconds for `stable-audio-open-1.0`
-2. **Quality vs Speed**:
-   - 50 steps: Fast, decent quality
-   - 100 steps: Good balance (default)
-   - 150+ steps: High quality, slower
-3. **Guidance Scale**:
-   - Lower (3-5): More creative/varied
-   - Default (7): Good balance
-   - Higher (10+): More literal to prompt
-4. **Negative Prompts**: Use to avoid "Low quality", "distorted", "noisy", etc.
-5. **Seeds**: Use same seed for reproducible results
-
-## Performance
-
-| Inference Steps | Quality | Speed | Use Case |
-|----------------|---------|-------|----------|
-| 50 | Good | Fast | Quick previews |
-| 100 (default) | Very Good | Medium | Production |
-| 150+ | Excellent | Slow | Final/critical audio |
-
-## Troubleshooting
-
-### Server not responding
-- Check if server is running: `curl http://localhost:8091/health`
-- Check server logs for errors
-
-### Audio quality issues
-- Increase `num_inference_steps` (e.g., 150)
-- Add negative prompts: `"Low quality, distorted, noisy"`
-- Increase `guidance_scale` for more prompt adherence
-
-### Generation timeout
-- Reduce `num_inference_steps`
-- Reduce `audio_length`
-- Check GPU memory with `nvidia-smi`
-
-### Wrong audio length
-- Ensure `audio_length` is within model limits (~47s max)
-- Adjust `audio_start` if trimming is needed
-
-## See Also
-
-- [Offline Inference Example](../../offline_inference/text_to_audio/README.md)
-- [Stable Audio Model Card](https://huggingface.co/stabilityai/stable-audio-open-1.0)
-- [vLLM-Omni Documentation](https://github.com/vllm-project/vllm-omni)

examples/online_serving/stable_audio/curl_examples.sh

@@ -1,54 +0,0 @@
-#!/bin/bash
-# Examples for using Stable Audio with curl via /v1/audio/generate endpoint
-
-# Example 1: Simple request with default parameters
-echo "Example 1: Simple request with default parameters"
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "The sound audience clapping and cheering in a stadium"
-    }' --output stadium.wav
-
-# Example 2: Request with custom audio_length
-echo "Example 2: Custom audio length (5 seconds)"
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "The sound of a dog barking",
-        "audio_length": 5.0
-    }' --output dog_5s.wav
-
-# Example 3: Request with negative prompt for quality control
-echo "Example 3: With negative prompt"
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "A piano playing a gentle melody",
-        "audio_length": 10.0,
-        "negative_prompt": "Low quality, distorted, noisy"
-    }' --output piano.wav
-
-# Example 4: Full control with all parameters
-echo "Example 4: Full control (custom length, guidance, steps, seed)"
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "Thunder and rain sounds",
-        "audio_length": 15.0,
-        "negative_prompt": "Low quality",
-        "guidance_scale": 7.0,
-        "num_inference_steps": 100,
-        "seed": 42
-    }' --output thunder_rain.wav
-
-# Example 5: Quick generation with fewer steps (faster but lower quality)
-echo "Example 5: Quick generation (fewer steps)"
-curl -X POST http://localhost:8091/v1/audio/generate \
-    -H "Content-Type: application/json" \
-    -d '{
-        "input": "Ocean waves crashing on a beach",
-        "audio_length": 8.0,
-        "num_inference_steps": 50
-    }' --output ocean.wav
-
-echo "All examples completed!"

examples/online_serving/stable_audio/stable_audio_client.py

@@ -1,170 +0,0 @@
-#!/usr/bin/env python3
-"""
-OpenAI-compatible client for Stable Audio via /v1/audio/generate endpoint.
-
-This script demonstrates how to use the OpenAI-compatible speech API
-to generate audio from text using Stable Audio models.
-
-Examples:
-    # Simple generation
-    python stable_audio_client.py --text "The sound of a cat purring"
-
-    # With custom duration
-    python stable_audio_client.py --text "A dog barking" --audio_length 5.0
-
-    # With all parameters
-    python stable_audio_client.py --text "Thunder and rain" \
-        --audio_length 15.0 \
-        --negative_prompt "Low quality" \
-        --guidance_scale 7.0 \
-        --num_inference_steps 100 \
-        --seed 42 \
-        --output thunder.wav
-"""
-
-import argparse
-import sys
-
-import requests
-
-
-def parse_args():
-    parser = argparse.ArgumentParser(description="Generate audio with Stable Audio via OpenAI-compatible API")
-    parser.add_argument(
-        "--api_url",
-        default="http://localhost:8091/v1/audio/generate",
-        help="API endpoint URL",
-    )
-    parser.add_argument(
-        "--text",
-        default="The sound of a cat purring",
-        help="Text prompt for audio generation",
-    )
-    parser.add_argument(
-        "--audio_length",
-        type=float,
-        default=10.0,
-        help="Audio length in seconds (max ~47s for stable-audio-open-1.0)",
-    )
-    parser.add_argument(
-        "--audio_start",
-        type=float,
-        default=0.0,
-        help="Audio start time in seconds",
-    )
-    parser.add_argument(
-        "--negative_prompt",
-        default="Low quality",
-        help="Negative prompt for classifier-free guidance",
-    )
-    parser.add_argument(
-        "--guidance_scale",
-        type=float,
-        default=7.0,
-        help="Guidance scale for diffusion (higher = more adherence to prompt)",
-    )
-    parser.add_argument(
-        "--num_inference_steps",
-        type=int,
-        default=100,
-        help="Number of inference steps (higher = better quality, slower)",
-    )
-    parser.add_argument(
-        "--seed",
-        type=int,
-        default=None,
-        help="Random seed for reproducibility",
-    )
-    parser.add_argument(
-        "--output",
-        default="stable_audio_output.wav",
-        help="Output file path",
-    )
-    parser.add_argument(
-        "--response_format",
-        default="wav",
-        choices=["wav", "mp3", "flac", "pcm"],
-        help="Audio output format",
-    )
-    return parser.parse_args()
-
-
-def generate_audio(args):
-    """Generate audio using the API."""
-
-    # Build request payload
-    payload = {
-        "input": args.text,
-        "audio_length": args.audio_length,
-        "audio_start": args.audio_start,
-        "response_format": args.response_format,
-    }
-
-    # Add optional parameters
-    if args.negative_prompt:
-        payload["negative_prompt"] = args.negative_prompt
-    if args.guidance_scale:
-        payload["guidance_scale"] = args.guidance_scale
-    if args.num_inference_steps:
-        payload["num_inference_steps"] = args.num_inference_steps
-    if args.seed is not None:
-        payload["seed"] = args.seed
-
-    print(f"\n{'=' * 60}")
-    print("Stable Audio - Text-to-Audio Generation")
-    print(f"{'=' * 60}")
-    print(f"API URL: {args.api_url}")
-    print(f"Prompt: {args.text}")
-    print(f"Audio length: {args.audio_length}s")
-    print(f"Negative prompt: {args.negative_prompt}")
-    print(f"Guidance scale: {args.guidance_scale}")
-    print(f"Inference steps: {args.num_inference_steps}")
-    if args.seed is not None:
-        print(f"Seed: {args.seed}")
-    print(f"Output: {args.output}")
-    print(f"{'=' * 60}\n")
-
-    try:
-        # Make the API request
-        print("Generating audio...")
-        response = requests.post(
-            args.api_url,
-            json=payload,
-            headers={"Content-Type": "application/json"},
-            timeout=300,  # 5 minute timeout for long generations
-        )
-
-        # Check for errors
-        if response.status_code != 200:
-            print(f"Error: API returned status code {response.status_code}")
-            print(f"Response: {response.text}")
-            return False
-
-        # Save the audio
-        with open(args.output, "wb") as f:
-            f.write(response.content)
-
-        print(f"✓ Audio saved to {args.output}")
-        print(f"  File size: {len(response.content) / 1024:.1f} KB")
-        return True
-
-    except requests.exceptions.Timeout:
-        print("Error: Request timed out. Try reducing inference steps or audio length.")
-        return False
-    except requests.exceptions.ConnectionError:
-        print(f"Error: Could not connect to {args.api_url}")
-        print("Make sure the server is running.")
-        return False
-    except Exception as e:
-        print(f"Error: {e}")
-        return False
-
-
-def main():
-    args = parse_args()
-    success = generate_audio(args)
-    sys.exit(0 if success else 1)
-
-
-if __name__ == "__main__":
-    main()

examples/online_serving/text_to_video/README.md

@@ -11,6 +11,11 @@ This example demonstrates how to deploy text-to-video models for online video ge
 | Wan2.2 T2V | `Wan-AI/Wan2.2-T2V-A14B-Diffusers` |
 | LTX-2 | `Lightricks/LTX-2` |
 | Helios (Base / Mid / Distilled) | `BestWishYsh/Helios-Base`, `Helios-Mid`, `Helios-Distilled` |
+| SANA-Video 2B (480p) | `Efficient-Large-Model/SANA-Video_2B_480p_diffusers` |
+| SANA-Video 2B (720p) | `Efficient-Large-Model/SANA-Video_2B_720p_diffusers` |
+
+For SANA-Video 2B text-to-video online serving, see the
+[SANA-Video 2B recipe](../../../recipes/NVIDIA/SANA-Video-2B.md).
 
 ## Wan2.2 T2V
 

examples/online_serving/text_to_video/run_curl_sana_video.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+
+BASE_URL="${BASE_URL:-http://localhost:8091}"
+OUTPUT_PATH="${OUTPUT_PATH:-sana_video.mp4}"
+
+curl -sS -X POST "$BASE_URL/v1/videos/sync" \
+    -F "prompt=A cinematic tracking shot of a sailboat crossing the ocean at sunset" \
+    -F "height=480" \
+    -F "width=832" \
+    -F "num_frames=81" \
+    -F "fps=16" \
+    -F "num_inference_steps=50" \
+    -F "guidance_scale=6.0" \
+    -F "seed=42" \
+    -F 'extra_params={"motion_score":30}' \
+    --output "$OUTPUT_PATH"

examples/online_serving/text_to_video/run_server_sana_video.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+# SANA-Video-2B text-to-video serving.
+
+MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
+PORT="${PORT:-8091}"
+
+vllm serve "$MODEL" \
+    --omni \
+    --model-class-name SanaVideoPipeline \
+    --dtype bfloat16 \
+    --port "$PORT"

examples/online_serving/text_to_video/run_server_sana_video_diffusers.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+# SANA-Video through the black-box Diffusers backend adapter.
+
+MODEL="${MODEL:-Efficient-Large-Model/SANA-Video_2B_480p_diffusers}"
+PORT="${PORT:-8091}"
+
+vllm serve "$MODEL" \
+    --omni \
+    --diffusion-load-format diffusers \
+    --diffusion-attention-backend TORCH_SDPA \
+    --dtype bfloat16 \
+    --port "$PORT"

pyproject.toml

@@ -73,6 +73,8 @@ dev = [
     "FlagEmbedding==1.4.0", # for hunyuan_image3 tests
     "kernels==0.14.1", # for accuracy tests comparison with diffusers output (0.15.0 incompatible with diffusers 0.38.0)
     "s3tokenizer==0.3.0", # for tts test
+    # Provides top-level cosyvoice2; do not co-install stepaudio2-minicpmo,
+    # whose top-level stepaudio2 package conflicts with this distribution.
     "step-audio2==1.0.0", # Step-Audio2 / MiniCPM-o 4.5 Token2Wav (flashcosyvoice / hyperpyyaml)
 ]
 
@@ -266,6 +268,7 @@ markers = [
     # specified computation resources marks (auto-added)
     # Tag with [hardware-resource] for tests.helpers.mark.get_hardware_mark_list().
     "H100: [hardware-resource] Tests that require H100 GPU",
+    "H800: [hardware-resource] Tests that require H800 GPU",
     "H200: [hardware-resource] Tests that require H200 GPU",
     "L4: [hardware-resource] Tests that require L4 GPU",
     "B200: [hardware-resource] Tests that require B200 GPU",

recipes/MiniMaxAI/MiniMax-H3.md

@@ -80,7 +80,7 @@ between the legacy and direct planar paths
 ([full result](https://github.com/vllm-project/vllm-omni/pull/6288#issuecomment-5337546499)):
 
 | Metric | Legacy | Direct planar | Change |
-|---|---:|---:|---:|
+| --- | ---: | ---: | ---: |
 | Median wall time | 1.805 s | 1.394 s | -22.8% |
 | Median process CPU time | 3.613 s | 3.207 s | -11.2% |
 | Peak RSS | 3182 MiB | 2794 MiB | -387 MiB (-12.2%) |
@@ -183,7 +183,7 @@ vllm serve "${MODEL}" \
 Use the profile that matches the per-GPU memory capacity:
 
 | Profile | GPUs | Starting shape | Resident DiT blocks | Attention | Execution | Status |
-|---|---:|---:|---:|---|---|---|
+| --- | ---: | ---: | ---: | --- | --- | --- |
 | `rtx5090` | 2 x 32 GB | 1344x768 | 20 | cuDNN attention | eager | Target-hardware validated |
 | `rtx4090` | 2 x 24 GB | 1024x576 | 12 | cuDNN attention | eager | Capacity-proxy starting point |
 
@@ -202,8 +202,8 @@ increasing it on a different request shape.
 At vLLM-Omni commit `ae6577ea`, one full 50-step T2VA request completed on
 2 x RTX 5090 without OOM:
 
-| Shape | Frames | Client E2E | Sampled peak/GPU | Output validation |
-|---:|---:|---:|---:|---|
+| Shape    | Frames        | Client E2E | Sampled peak/GPU       | Output validation                                            |
+| -------: | ------------: | ---------: | ---------------------: | -----------------------------------------------------------: |
 | 1344x768 | 124 at 24 FPS | 8 min 38 s | approximately 22.6 GiB | H.264 video + 32 kHz stereo AAC; full `ffmpeg` decode passed |
 
 This is a single end-to-end validation run, not a warmed multi-run latency
@@ -272,7 +272,7 @@ vllm serve "${MODEL}" \
 Do not add `--enforce-eager` to this performance configuration. The first
 request includes regional compilation; warm the server once before measuring
 steady-state latency. H3 is CFG-distilled, so `--cfg-parallel-size` must remain
-1. The H3 VAE supports its native `tile` mode, not
+`1`. The H3 VAE supports its native `tile` mode, not
 `spatial_shard_height` or `spatial_shard_width`.
 
 ### Attention Backends
@@ -380,6 +380,63 @@ No restart is needed: `task=fl2va` routes to `FL2VA/transformer`, while
 `task=ref2va` routes to
 `Ref2VA/transformer`. T2VA uses the FL2VA DiT.
 
+### Step execution and continuous batching
+
+H3 implements the step-wise execution contract, so the scheduler can admit and
+retire requests between denoise steps instead of running one request end to
+end. Add the feature gate, then raise `--max-num-seqs` to co-batch:
+
+```bash
+--step-execution --max-num-seqs 4
+```
+
+Co-batched requests are packed into a single sequence that keeps one attention
+document per request, so a batch costs one DiT forward. That packing requires
+`--diffusion-attention-backend FLASH_ATTN`; other backends fall back to one
+forward per request. `--max-num-seqs 1` keeps the conservative single-request
+step path. Cache acceleration (`--cache-backend`) is not available in step mode.
+
+!!! warning "Co-batching does not improve H3 throughput for large simultaneous requests"
+    Keep `--max-num-seqs 1` unless you specifically need scheduler-level control
+    (admitting and retiring requests between denoise steps). Measured on two
+    H100s (TP2, BF16, 672x384, 209 frames, 30 steps, 4 requests at concurrency
+    4 with `--request-rate inf`, i.e. all four submitted at time 0; one packed
+    request is 16384 rows):
+
+    | Configuration | Wall time | Mean latency | Peak memory |
+    |---------------|-----------|--------------|-------------|
+    | request mode | 174.8 s | 111.5 s | 72.4 GB |
+    | `--step-execution --max-num-seqs 1` | 179.0 s | 113.8 s | 72.4 GB |
+    | `--step-execution --max-num-seqs 4` | 182.1 s | 175.7 s | 78.3 GB |
+
+    A single H3 denoise step is a compute-bound dense GEMM over an already long
+    packed sequence, so fusing N requests costs N times the FLOPs and buys almost
+    no amortization — unlike LLM decoding, which is memory-bandwidth bound.
+    Going from one request per step to four cuts the per-request denoise cost
+    only from 1.323 s to 1.291 s (2.4%), which the step-mode bookkeeping then
+    spends. Mean latency degrades further because co-batched requests finish
+    together instead of staggered. Quantization moves the absolute numbers
+    without changing this: with online `int8` the same workload runs in 153.3 s
+    at 56.9 GB (request mode), and `--max-num-seqs 4` is still 5.0% slower than
+    request mode.
+
+    Two workloads outside this table are unmeasured and may behave differently:
+
+    - **Staggered arrivals (admission latency).** A single 16384-row H3 request
+      already saturates the GPUs, so submitting all requests at time 0 is the
+      one arrival pattern where co-batching cannot win: it can only bunch
+      completions. In request mode a new request queues behind the whole
+      in-flight generation (~45 s at this size); step mode admits at the next
+      denoise-step boundary (~1.3 s). Whether that shows up as a wall-clock
+      benefit under Poisson arrivals is not yet measured for H3. To reproduce,
+      run request mode vs `--step-execution --max-num-seqs 4` with
+      `diffusion_benchmark_serving.py --request-rate 0.05` (roughly one request
+      every 20 s) and report mean / p95 latency, which then includes queueing.
+    - **Small requests.** The numbers above are drawn from 672x384 / 209-frame
+      requests. A short clip at lower resolution packs a few thousand rows and
+      may not saturate the hardware; co-batching may amortize better there.
+      This is also unmeasured.
+
 ### Online FP8 quantization
 
 MiniMax H3 supports online FP8 quantization of both the DiT and the Qwen3-VL
@@ -530,7 +587,7 @@ vLLM-Omni with MiniMax H3 support, BF16. gfx942 rows measured with the
 `0.26.0+rocm723` wheel (HIP 7.2).
 
 | Workload | Configuration | Observed result |
-|----------|---------------|-----------------|
+| ---------- | --------------- | ----------------- |
 | T2VA, 1344x768, 209 frames, 50 steps | 4x gfx942 (MI300X), FLASH_ATTN, USP4, text-enc TP4, VAE PP4 tile | encode 0.09 s, denoise 244.04 s, decode 4.15 s, 267.42 s client E2E; H.264 24 FPS + 32 kHz stereo AAC |
 | FL2VA, 1344x768, 209 frames, 50 steps | 4x gfx942 (MI300X), FLASH_ATTN, USP4, text-enc TP4, VAE PP4 tile | encode 13.98 s, denoise 257.58 s, decode 4.11 s, 287.07 s client E2E; H.264 24 FPS + 32 kHz stereo AAC |
 | T2VA, 832x480, ~4 s, 40 steps | 1x gfx950 (MI350), FLASH_ATTN, CPU offload | valid MP4 (H.264 + synced audio); ~0.73 s/denoise-step (~1.37 it/s), ~55 s client E2E incl. warmup |
@@ -712,7 +769,7 @@ at most 15 seconds combined.
 ## Official input matrix and limits
 
 | Task | Supported references | Limits |
-|------|----------------------|--------|
+| ------ | ---------------------- | -------- |
 | T2VA | text only | prompt must be non-empty |
 | FL2VA | first image, last image, or ordered first+last images | at most 2 images; `frame_indices` is `[0]`, `[-1]`, or `[0,-1]` |
 | Ref2VA | image-only, image+image, image+video, video+audio, and mixed image/video/audio | images ≤9, videos ≤3, audios ≤3, total references ≤12; audio requires a visual reference |
@@ -743,13 +800,17 @@ Omitting `quality` preserves the startup default: it uses the reference path
 normally, or the server-configured profile when the server was started with
 `--cache-backend cache_dit`.
 
+Turbo is independent of this quality switch: `quality` selects a Cache-DiT
+policy, while Turbo changes the active LoRA weights and sampling schedule. See
+[LoRA](#lora) below.
+
 The following result was measured on 4× NVIDIA H200 with SP4, text-encoder
 TP4, 1344×768, 124 frames, 24 FPS, and 50 inference steps. One full
 `lossless` warmup was excluded, followed by three fixed prompt/seed pairs in
 balanced switch order.
 
 | `quality` | Median inference latency | Speedup | SSIM vs `lossless` | PSNR vs `lossless` | Expected trade-off |
-|---|---:|---:|---:|---:|---|
+| --- | ---: | ---: | ---: | ---: | --- |
 | `lossless` | 85.49 s | 1.00× | 1.0000 | exact | Native reference path |
 | `high` | 63.36 s | 1.35× | 0.9709 | 34.98 dB | Faster with measured same-seed deviation |
 
@@ -759,10 +820,50 @@ balanced switch order.
 > workload. The values above apply to this deployment and are not universal
 > guarantees. `lossless` remains the exact reference path.
 
+## LoRA
+
+### Turbo LoRA
+
+Only the native Diffusers 4-step FL2VA/T2VA v1.0 artifact is supported:
+
+```text
+lightx2v/Minimax-h3-Turbo/minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
+```
+
+Download only that file:
+
+```bash
+export TURBO_DIR=/path/to/minimax-h3-turbo
+export TURBO_FILE=minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
+hf download lightx2v/Minimax-h3-Turbo "${TURBO_FILE}" --local-dir "${TURBO_DIR}"
+export TURBO_LORA="${TURBO_DIR}/${TURBO_FILE}"
+```
+
+Start from a non-offloaded or DLO FL2VA server command and add
+`--task-type fl2va --lora-backend peft --lora-path "${TURBO_LORA}"`.
+`--lora-path` preloads the adapter; each request still activates it and uses
+the published sampling settings:
+
+```bash
+-F 'num_inference_steps=5' \
+-F 'flow_shift=6' \
+-F 'extra_params={"task":"t2va","duration":4.4,"audio_flow_shift":3.0}' \
+-F "lora={\"name\":\"h3-turbo-v1.0\",\"path\":\"${TURBO_LORA}\",\"scale\":1.0}"
+```
+
+For FL2VA, change `task` and add `input_reference` as shown above. The 8-step,
+ComfyUI, Ref2VA, and v1.1 artifacts are not supported. This integration is
+dynamic-only and does not support prefusion or LoRA composition. DLO is
+supported by keeping the request-switchable LoRA A/B buffers resident on the
+accelerator while DLO streams only the base blocks; budget for this additional
+fixed HBM usage. Model-level and standard layerwise offload remain unsupported.
+The five requested sigma points produce the four denoiser evaluations expected
+by the Turbo artifact.
+
 ## Key parameters
 
 | Parameter | Recommended value | Notes |
-|-----------|-------------------|-------|
+| ----------- | ------------------- | ------- |
 | `quality` | omitted or `lossless` | Request-level quality intent; `high` dynamically installs H3's conservative Cache-DiT profile |
 | `extra_params.force_refresh_step_hint` | omitted | Optional positive 1-based denoising-step hint for an active Cache-DiT request; pair with `extra_params.force_refresh_step_policy`=`once` or `repeat` |
 | `task` | `t2va`, `fl2va`, or `ref2va` | Passed in `extra_params`; selects the task-specific DiT |
@@ -787,9 +888,9 @@ Users can also use a ComfyUI frontend to interact with a hosted MiniMax-H3 servi
 The four-GPU recommendation was measured on four NVIDIA B300 GPUs with one
 excluded warmup followed by three requests.
 
-| Workload | Configuration | Observed result |
-|----------|---------------|-----------------|
-| FL2VA, 209 frames, 1248x768 | no offload, U4, VPP4 tile, regional compile | 86.964 s mean HTTP client latency |
+| Workload                               | Configuration                               | Observed result                      |
+| -------------------------------------- | ------------------------------------------- | ------------------------------------ |
+| FL2VA, 209 frames, 1248x768            | no offload, U4, VPP4 tile, regional compile | 86.964 s mean HTTP client latency    |
 | Two-video Ref2VA, 362 frames, 1344x768 | no offload, U4, VPP4 tile, regional compile | 784.394 s accounted model-stage mean |
 
 These measurements describe the validated shapes rather than a general
@@ -881,7 +982,12 @@ vllm serve "${MODEL_ROOT}/FL2VA" \
 - TeaCache is calibrated for FL2VA only; Ref2VA requests run uncached.
 - Combined serving requires sibling `FL2VA` and `Ref2VA` directories, loads
   both task-specific DiTs, and loads shared components once from `FL2VA`.
-- H3 currently executes one generation request per diffusion batch.
+- Request mode executes one generation request per diffusion batch. Use
+  `--step-execution` with `--max-num-seqs N` to admit several requests at once
+  (see
+  [Execution modes](https://github.com/vllm-project/vllm-omni/blob/main/docs/user_guide/diffusion/execution_modes.md));
+  step
+  mode does not support `cache_backend`.
 - The first regional-compile request is a warmup and should not be included in
   steady-state performance measurements.
 - The serving path accepts fewer references than the model supports. H3 documents up

recipes/NVIDIA/SANA-Video-2B.md

@@ -0,0 +1,245 @@
+# SANA-Video 2B
+
+> Native text-to-video and image-to-video generation at 480p and 720p,
+> plus a validated Diffusers-adapter compatibility path
+
+## Summary
+
+- Vendor: NVIDIA
+- Model: `Efficient-Large-Model/SANA-Video_2B_480p_diffusers`,
+  `Efficient-Large-Model/SANA-Video_2B_720p_diffusers`
+- Task: Text-to-video and image-to-video
+- Mode: Offline inference and OpenAI-compatible online serving
+- Maintainer: Community
+
+## When to use this recipe
+
+Use the native `SanaVideoPipeline` for T2V and
+`SanaImageToVideoPipeline` for I2V. Both native pipelines support the 480p
+and 720p checkpoints. Use `--diffusion-load-format diffusers` when you need
+the black-box Diffusers compatibility baseline; adapter T2V and I2V are
+validated at both resolutions.
+
+The native pipeline loads the 480p checkpoint through
+`DistributedAutoencoderKLWan` and the 720p checkpoint through
+`DistributedAutoencoderKLLTX2Video`. These are vLLM-Omni distributed wrappers
+around the corresponding Diffusers autoencoders, not independent VAE
+implementations. The denoising loop also intentionally loads Diffusers'
+`DPMSolverMultistepScheduler` from the checkpoint to preserve its scheduler
+configuration and numerical behavior.
+
+## References
+
+- Upstream project: <https://github.com/NVlabs/Sana>
+- Model cards:
+  [480p](https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_480p_diffusers),
+  [720p](https://huggingface.co/Efficient-Large-Model/SANA-Video_2B_720p_diffusers)
+- Diffusers documentation: <https://huggingface.co/docs/diffusers/api/pipelines/sana_video>
+- Online serving guides:
+  [Text-to-Video](../../docs/user_guide/examples/online_serving/text_to_video.md),
+  [Image-to-Video](../../docs/user_guide/examples/online_serving/image_to_video.md)
+- Offline T2V example:
+  [`examples/offline_inference/text_to_video/text_to_video.py`](../../examples/offline_inference/text_to_video/text_to_video.py)
+- Offline I2V example:
+  [`examples/offline_inference/image_to_video/image_to_video.py`](../../examples/offline_inference/image_to_video/image_to_video.py)
+- Support discussion: [vLLM-Omni issue #5432](https://github.com/vllm-project/vllm-omni/issues/5432)
+
+## Hardware Support
+
+## GPU
+
+### 1x RTX 5090 32GB
+
+#### Environment
+
+- OS: Ubuntu 22.04.5 LTS
+- Python: 3.12.3
+- Driver / runtime: NVIDIA driver 580.95.05; PyTorch 2.11.0+cu130
+- Diffusers: 0.38.0
+- vLLM version: 0.26.0
+- vLLM-Omni version or commit: PR #5508, commit `22037901`
+
+#### Command
+
+##### Native text-to-video inference
+
+```bash
+python examples/offline_inference/text_to_video/text_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_720p_diffusers \
+  --model-class-name SanaVideoPipeline \
+  --prompt "A cat walking on the grass, facing the camera." \
+  --negative-prompt "blurry, low quality, temporal artifacts" \
+  --height 704 --width 1280 --num-frames 81 \
+  --num-inference-steps 50 --guidance-scale 6 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 --seed 42 --output sana_video_720p.mp4
+```
+
+For 480p, select `SANA-Video_2B_480p_diffusers` and use
+`--height 480 --width 832`.
+
+##### Native image-to-video inference
+
+SANA checkpoints declare `SanaVideoPipeline` in `model_index.json`, so I2V
+must be selected explicitly with `--model-class-name
+SanaImageToVideoPipeline`.
+
+```bash
+python examples/offline_inference/image_to_video/image_to_video.py \
+  --model Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  --model-class-name SanaImageToVideoPipeline \
+  --image input.jpg \
+  --prompt "A cat turns toward the camera with smooth, natural motion." \
+  --negative-prompt "blurry, low quality, temporal artifacts" \
+  --height 480 --width 832 --num-frames 81 \
+  --num-inference-steps 50 --guidance-scale 6 \
+  --extra-body '{"motion_score": 30}' \
+  --fps 16 --seed 42 --output sana_video_i2v_480p.mp4
+```
+
+The same pipeline class supports the 720p checkpoint through vLLM-Omni's
+distributed LTX-2 VAE wrapper; use `--height 704 --width 1280`.
+
+For online I2V serving:
+
+```bash
+MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  bash examples/online_serving/image_to_video/run_server_sana_video.sh
+
+INPUT_IMAGE=input.jpg OUTPUT_PATH=sana_video_i2v.mp4 \
+  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
+```
+
+##### Native online serving
+
+```bash
+MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  bash examples/online_serving/text_to_video/run_server_sana_video.sh
+
+bash examples/online_serving/text_to_video/run_curl_sana_video.sh
+```
+
+To run the black-box compatibility backend for T2V, replace the server script
+with `run_server_sana_video_diffusers.sh`. The same `/v1/videos` request
+works; `num_frames` is adapted to Diffusers' `frames` argument. The script
+selects `TORCH_SDPA` because SANA-Video uses an attention mask that the
+AITER-backed Diffusers attention path does not accept.
+
+##### Diffusers-adapter image-to-video serving
+
+The validated I2V adapter commands are:
+
+```bash
+# 480p
+MODEL=Efficient-Large-Model/SANA-Video_2B_480p_diffusers \
+  bash examples/online_serving/image_to_video/run_server_sana_video_diffusers.sh
+
+INPUT_IMAGE=input.jpg OUTPUT_PATH=sana_video_i2v_adapter.mp4 \
+  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
+
+# 720p
+MODEL=Efficient-Large-Model/SANA-Video_2B_720p_diffusers \
+  bash examples/online_serving/image_to_video/run_server_sana_video_diffusers.sh
+
+INPUT_IMAGE=input.jpg WIDTH=1280 HEIGHT=704 \
+  OUTPUT_PATH=sana_video_i2v_adapter_720p.mp4 \
+  bash examples/online_serving/image_to_video/run_curl_sana_video.sh
+```
+
+#### Verification
+
+Check the encoded 720p output metadata after running a generation command:
+
+```bash
+ffprobe -v error -select_streams v:0 \
+  -show_entries stream=width,height,r_frame_rate,nb_frames \
+  -of default=noprint_wrappers=1 sana_video_720p.mp4
+```
+
+The standard 720p request above should report:
+
+```text
+width=1280
+height=704
+r_frame_rate=16/1
+nb_frames=81
+```
+
+For 480p, expect `width=832` and `height=480`.
+
+The automated serving matrix covers both checkpoint variants:
+
+| Backend | 480p T2V | 720p T2V | 480p I2V | 720p I2V |
+| --- | --- | --- | --- | --- |
+| Native vLLM-Omni | Validated | Validated | Validated | Validated |
+| Diffusers adapter | Validated | Validated | Validated | Validated |
+
+Use the native `SanaVideoPipeline` and `SanaImageToVideoPipeline` for the
+primary SANA execution paths. The Diffusers adapter is retained as a
+validated compatibility/reference backend.
+
+#### Notes
+
+- Key flags: select I2V explicitly with `--model-class-name
+  SanaImageToVideoPipeline`; SANA checkpoint `model_index.json` files declare
+  the T2V class. Pass `motion_score` through `--extra-body`. The supplied
+  Diffusers-adapter server scripts select `TORCH_SDPA` because the AITER-backed
+  path does not accept SANA's attention mask.
+- Output profile: 81 frames at 16 FPS is the standard checkpoint profile and
+  produces approximately five seconds of video. Minute-scale generation
+  requires the separate LongSANA/LongLive block-autoregressive workflow.
+- Backend boundary: the native pipelines and Transformer are owned by
+  vLLM-Omni. The 480p Wan VAE and 720p LTX-2 VAE run through vLLM-Omni
+  distributed wrappers derived from the corresponding Diffusers VAE classes.
+  The denoising loop intentionally retains the checkpoint-compatible
+  Diffusers `DPMSolverMultistepScheduler`.
+- Known limitations:
+    - Sequence/tensor/CFG parallelism, Cache-DiT, TeaCache, and step execution
+    are not validated for the native pipeline.
+    - The Diffusers backend is a compatibility path and does not provide native
+    vLLM-Omni parallelism or continuous batching.
+    - Native describes pipeline and Transformer ownership, not a zero-Diffusers
+    dependency guarantee.
+
+### 1x NVIDIA L20X
+
+#### Environment
+
+- Driver: 570.133.20
+- PyTorch: 2.13.0+cu130
+- Diffusers: 0.38.0
+- vLLM: 0.27.0
+- Nsight Systems: 2026.1
+- vLLM-Omni commit: `54fcc6d`
+
+#### Same-workload T2V comparison
+
+Both backends used
+`Efficient-Large-Model/SANA-Video_2B_480p_diffusers`, the same T2V prompt
+and seeds, 832×480 output, 81 frames, and four inference steps. One warmup
+request was excluded, followed by three measured requests.
+
+| Metric | Native vLLM-Omni | Diffusers adapter | Native delta |
+| --- | ---: | ---: | ---: |
+| E2E latency | 7.1600 ± 0.0043 s | 6.9761 ± 0.0517 s | +2.636% |
+| Throughput | 0.1397 req/s | 0.1433 req/s | -2.568% |
+| Peak reserved VRAM | 23,936 MiB | 19,060 MiB | +25.582% |
+| Initialization | 11.669 s | 20.236 s | -42.336% |
+
+The peak-memory result is not a Transformer-only comparison. For the 480p
+checkpoint, the native pipeline deliberately loads and decodes with the Wan
+VAE in FP32. This matches the upstream SANA-Video reference and the recorded
+accuracy-golden setup for this integration. The adapter launch uses
+`--dtype bfloat16` as the Diffusers pipeline-wide dtype, so its Wan VAE also
+runs in BF16.
+
+Peak VRAM is the CUDA allocator's maximum reserved memory over the complete
+request. It includes VAE decode activations and allocator-reserved blocks, not
+only live Transformer tensors. The different VAE precision policies are a
+material code-level difference and a likely contributor to the higher native
+high-water mark; the available full-request measurements do not isolate how
+much of the delta they account for. The 25.582% result should therefore not be
+interpreted as an isolated native Transformer memory regression. Exact
+attribution requires both paths to use the same VAE dtype. Changing the native
+480p VAE to BF16 would instead change the recorded golden/reference setup and
+requires a new accuracy study.

recipes/README.md

@@ -57,6 +57,8 @@ recipes/
 | [`krea/Krea-2.md`](./krea/Krea-2.md) | Text-to-image (Turbo + Raw), offline + online, with LoRA | 1x H100 80GB |
 | [`LTX/LTX-2.md`](./LTX/LTX-2.md) | LTX-2/LTX-2.3 text-to-video and image-to-video with synchronized audio | H200 141GB / 96GB-class GPU |
 | [`LTX/LTX-2.5.md`](./LTX/LTX-2.5.md) | LTX-2.5-Diffusers: Full/SFT one-stage and distilled two-stage T2V/I2V with synchronized audio | NVIDIA B300; cuDNN-qualified |
+| [`MammothModa2/MammothModa2.md`](./MammothModa2/MammothModa2.md) | Preview and Dev text-to-image (AR → DiT); Dev text/image understanding | Preview: 1x L40S 48GB / 1x ≥40GB GPU; Dev: 1x NVIDIA GPU with sufficient cache headroom |
+| [`NVIDIA/SANA-Video-2B.md`](./NVIDIA/SANA-Video-2B.md) | Native and Diffusers-adapter 480p/720p T2V+I2V | 1x RTX 5090 32GB |
 | [`meituan-longcat/LongCat-Video-Avatar-1.5.md`](./meituan-longcat/LongCat-Video-Avatar-1.5.md) | Audio-driven avatar video generation (AT2V / AI2V, single- and multi-speaker, AVC continuation) | 1x H100 80GB |
 | [`mistralai/Voxtral-TTS.md`](./mistralai/Voxtral-TTS.md) | Online serving for TTS | 1x RTX 4090 24GB |
 | [`Robbyant/LingBot-Video.md`](./Robbyant/LingBot-Video.md) | Native dense and MoE text-to-video serving | Dense: 1x L20X; MoE: 1x L20X (~67.7 GiB peak) |
@@ -83,6 +85,7 @@ recipes/
 | [`Wan-AI/Wan2.2-I2V.md`](./Wan-AI/Wan2.2-I2V.md) | Image-to-video serving (Wan2.2 14B) | 8x Ascend NPU (A2/A3) |
 | [`Wan-AI/Wan2.2-S2V.md`](./Wan-AI/Wan2.2-S2V.md) | Speech-to-video serving (Wan2.2 14B) | 2x A100/H100 80GB |
 | [`Wan-AI/Wan2.1-VACE.md`](./Wan-AI/Wan2.1-VACE.md) | Unified T2V, I2V, V2LF, FLF2V, inpaint, and R2V | 1x RTX 5090 (1.3B) / 1x L40S 48GB with layerwise offload (14B) |
+| [`XiaomiMiMo/MiMo-Audio.md`](./XiaomiMiMo/MiMo-Audio.md) | Offline + online omni audio (TTS, ASR, dialogue) | 1x RTX 5090 / 5090D 32GB |
 | [`StabilityAI/Stable-Diffusion-3.5.md`](./StabilityAI/Stable-Diffusion-3.5.md) | Text-to-image serving (SD 3.5-medium and SD 3.5-large) | 1x RTX A6000 48GB |
 | [`zai-org/GLM-TTS.md`](./zai-org/GLM-TTS.md) | Online serving for Chinese/English zero-shot voice-cloned TTS | 1x A40 48GB |
 | [`GLM/GLM-Image.md`](./GLM/GLM-Image.md) | Online serving for image generation | 1x A800 80GB / 2x A800 80GB |

recipes/StabilityAI/Stable-Audio-Open.md

@@ -23,8 +23,6 @@ offline validation sample with TeaCache and online serving through the
 - Model: <https://huggingface.co/stabilityai/stable-audio-open-1.0>
 - Offline example:
   [`examples/offline_inference/text_to_audio`](../../examples/offline_inference/text_to_audio)
-- Online example:
-  [`examples/online_serving/stable_audio`](../../examples/online_serving/stable_audio)
 - Audio generation API:
   [`docs/serving/audio_generate_api.md`](../../docs/serving/audio_generate_api.md)
 - Related issue:
@@ -103,7 +101,7 @@ curl -X POST http://localhost:8091/v1/audio/generate \
     "seed": 42,
     "response_format": "wav"
   }' \
-  --output examples/online_serving/stable_audio/piano_10s.wav
+  --output piano_10s.wav
 ```
 
 ##### Verification
@@ -137,12 +135,12 @@ PY
 Validate the online output:
 
 ```bash
-ls -lh examples/online_serving/stable_audio/piano_10s.wav
+ls -lh piano_10s.wav
 
 python - <<'PY'
 import soundfile as sf
 
-path = "examples/online_serving/stable_audio/piano_10s.wav"
+path = "piano_10s.wav"
 audio, sample_rate = sf.read(path)
 print("sample_rate:", sample_rate)
 print("shape:", audio.shape)

recipes/XiaomiMiMo/MiMo-Audio.md

@@ -0,0 +1,161 @@
+# MiMo-Audio for omni audio generation and understanding
+
+> Offline + online TTS, ASR, audio understanding, and multi-turn dialogue
+
+## Summary
+
+- Vendor: Xiaomi MiMo
+- Model: `XiaomiMiMo/MiMo-Audio-7B-Instruct`
+- Task: Text-to-speech, speech-to-text, audio understanding, spoken dialogue
+- Mode: Offline end-to-end scripts and online OpenAI-compatible chat serving
+- Maintainer: Community
+
+## When to use this recipe
+
+Use this recipe when you want a known-good starting point for running
+MiMo-Audio on a single **32 GB Blackwell consumer GPU** (RTX 5090 / 5090D).
+The bundled deploy profile [`vllm_omni/deploy/mimo_audio_5090d.yaml`](../../vllm_omni/deploy/mimo_audio_5090d.yaml)
+splits VRAM between the AR thinker (stage 0) and code2wav (stage 1), avoids
+FlashAttention-2 on SM 12.x, and keeps the audio tokenizer on CPU to reduce
+Stage-1 GPU pressure.
+
+## References
+
+- Upstream model card: <https://huggingface.co/XiaomiMiMo/MiMo-Audio-7B-Instruct>
+- Audio tokenizer: <https://huggingface.co/XiaomiMiMo/MiMo-Audio-Tokenizer>
+- Offline example:
+  [`examples/offline_inference/mimo_audio/end2end.py`](../../examples/offline_inference/mimo_audio/end2end.py)
+- Online example:
+  [`examples/online_serving/mimo_audio/README.md`](../../examples/online_serving/mimo_audio/README.md)
+- Default deploy config (generic single GPU):
+  [`vllm_omni/deploy/mimo_audio.yaml`](../../vllm_omni/deploy/mimo_audio.yaml)
+- 5090D-tuned deploy config:
+  [`vllm_omni/deploy/mimo_audio_5090d.yaml`](../../vllm_omni/deploy/mimo_audio_5090d.yaml)
+- Pipeline / deploy docs:
+  [Stage configs](https://docs.vllm.ai/projects/vllm-omni/en/latest/configuration/stage_configs/)
+- Related issue or discussion:
+  [RFC: add recipes folder](https://github.com/vllm-project/vllm-omni/issues/2645)
+- Tokenizer device alignment fix (required for CPU tokenizer on stage 1):
+  [PR #6539](https://github.com/vllm-project/vllm-omni/pull/6539)
+
+## Hardware Support
+
+## GPU
+
+### 1x NVIDIA GeForce RTX 5090 / 5090D 32GB
+
+Two-stage async-chunk pipeline (thinker + code2wav) on a single GPU. Tested
+offline with `tts_sft`.
+
+#### Environment
+
+- OS: Ubuntu (Linux x86_64)
+- Python: 3.12
+- PyTorch: 2.11.0+cu129
+- Driver / runtime: NVIDIA driver 570.x, CUDA 12.9 runtime (Blackwell SM 12.x)
+- GPU: NVIDIA GeForce RTX 5090 D, 32 GB
+- vLLM: 0.26.0+cu129
+- vLLM-Omni: `main` (tested while [#6539](https://github.com/vllm-project/vllm-omni/pull/6539) was open for CPU-tokenizer support)
+
+Install vLLM-Omni from source and align the vLLM wheel with your CUDA stack.
+On Blackwell, prefer the **cu129** vLLM build over cu13 defaults when using
+PyTorch 2.11+cu129.
+
+Download or cache:
+
+- `XiaomiMiMo/MiMo-Audio-7B-Instruct`
+- `XiaomiMiMo/MiMo-Audio-Tokenizer` (or a local checkout of the tokenizer repo)
+
+Export the runtime knobs below before offline or online runs:
+
+```bash
+export MIMO_AUDIO_TOKENIZER_PATH="${MIMO_AUDIO_TOKENIZER_PATH:-XiaomiMiMo/MiMo-Audio-Tokenizer}"
+export MIMO_AUDIO_TOKENIZER_DEVICE=cpu
+export MIMO_AUDIO_TOKENIZER_CUDA_GRAPH=0
+export VLLM_USE_FLASHINFER_SAMPLER=0
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+```
+
+Optional (mirror / download stability):
+
+```bash
+export HF_HUB_DISABLE_XET=1
+export HF_ENDPOINT=https://hf-mirror.com   # if you use a Hugging Face mirror
+```
+
+Do **not** set `VLLM_ATTENTION_BACKEND` on vLLM 0.26.x; configure
+`attention_backend: TRITON_ATTN` in the deploy YAML instead.
+
+#### Command
+
+Offline TTS validation from the repository root:
+
+```bash
+cd examples/offline_inference/mimo_audio
+
+python3 -u end2end.py \
+  --deploy-config vllm_omni/deploy/mimo_audio_5090d.yaml \
+  --model-name XiaomiMiMo/MiMo-Audio-7B-Instruct \
+  --query-type tts_sft \
+  --text "The weather is so nice today."
+```
+
+Online serving (requires MiMo chat template):
+
+```bash
+export MIMO_AUDIO_TOKENIZER_PATH="XiaomiMiMo/MiMo-Audio-Tokenizer"
+
+vllm serve XiaomiMiMo/MiMo-Audio-7B-Instruct --omni \
+  --deploy-config vllm_omni/deploy/mimo_audio_5090d.yaml \
+  --served-model-name "MiMo-Audio-7B-Instruct" \
+  --port 18091 \
+  --chat-template ./examples/online_serving/mimo_audio/chat_template.jinja
+```
+
+Pass the same environment exports to the server process when using a CPU
+tokenizer or disabling FlashInfer sampling.
+
+#### Verification
+
+Offline:
+
+```bash
+ls -lh output_audio/tts_sft/*.wav
+```
+
+Expect a non-empty `.wav` under `output_audio/tts_sft/` and a companion `.txt`
+with the decoded text. A successful run completes both stages without a
+Stage-1 `decode_vq` device mismatch.
+
+Online:
+
+```bash
+curl -s -o /tmp/mimo_health.txt -w '%{http_code}' http://127.0.0.1:18091/health
+```
+
+Then use the client under
+[`examples/online_serving/mimo_audio/`](../../examples/online_serving/mimo_audio/).
+
+#### Notes
+
+- Memory usage (observed offline): Stage-0 weights ~16.6 GiB; KV cache ~4.9 GiB
+  with `gpu_memory_utilization: 0.78` on stage 0 and `0.12` on stage 1.
+- Key flags:
+  - `--deploy-config vllm_omni/deploy/mimo_audio_5090d.yaml` for 5090D memory
+    and `TRITON_ATTN`.
+  - `MIMO_AUDIO_TOKENIZER_PATH` is mandatory.
+  - `MIMO_AUDIO_TOKENIZER_DEVICE=cpu` saves VRAM; requires
+    [#6539](https://github.com/vllm-project/vllm-omni/pull/6539) (or equivalent
+    fix) so code2wav decode uses the tokenizer device.
+  - `VLLM_USE_FLASHINFER_SAMPLER=0` avoids FlashInfer top-p/top-k sampler
+    issues on consumer GPUs.
+- Known limitations:
+  - Blackwell may log `SM 12.x requires CUDA >= 12.9` from some optional
+    tooling; the tested path uses TRITON attention and cu129 wheels.
+  - Without a working `flash-attn` build, audio quality may be slightly
+    metallic; install FlashAttention when a compatible wheel is available.
+  - First request triggers Triton JIT for rotary / attention kernels (latency
+    spike); subsequent requests are faster.
+  - For longer outputs or batching, you may need to raise `max_model_len` in
+    the deploy YAML and match `max_position_embeddings` in the model config
+    (see the offline example README).

tests/benchmarks/metrics/test_metrics.py

@@ -1,10 +1,12 @@
 # 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 metrics.py
 """
 
+import math
+
 import pytest
 from vllm.benchmarks.serve import TaskType
 
@@ -66,6 +68,22 @@ def _make_output(prompt_len: int, output_tokens: int = 10) -> MixRequestFuncOutp
     return output
 
 
+def _calculate_test_metrics(outputs, goodput=None):
+    return calculate_metrics(
+        input_requests=[],
+        outputs=outputs,
+        dur_s=1.0,
+        tokenizer=None,
+        selected_percentiles=[50.0],
+        goodput_config_dict=goodput or {},
+        task_type=TaskType.GENERATION,
+        selected_percentile_metrics=[],
+        max_concurrency=None,
+        request_rate=float("inf"),
+        benchmark_duration=1.0,
+    )[0]
+
+
 # ============================================================================
 # total_input Tests
 # ============================================================================
@@ -126,6 +144,166 @@ def test_audio_continuity_aggregation():
     assert p99 is not None and p99 > 0.4
 
 
+def test_unmeasured_duplex_latency_does_not_add_zero_samples():
+    measured = _make_output(100, output_tokens=2)
+    measured.ttft = 0.1
+    measured.audio_ttfp = 0.2
+    measured.audio_rtf = 0.5
+    measured.duplex_session_metrics = {
+        "mean_ttft_ms": 100.0,
+        "mean_ttfp_ms": 200.0,
+        "mean_rtf": 0.5,
+    }
+    listen_only = _make_output(100, output_tokens=0)
+    listen_only.ttft = listen_only.audio_ttfp = listen_only.audio_rtf = 0.0
+    listen_only.duplex_session_metrics = {
+        "mean_ttft_ms": None,
+        "mean_ttfp_ms": None,
+        "mean_rtf": None,
+    }
+
+    metrics = _calculate_test_metrics([measured, listen_only], {"ttft": 150.0})
+
+    assert metrics.mean_ttft_ms == 100.0
+    assert metrics.mean_audio_ttfp_ms == 200.0
+    assert metrics.mean_audio_rtf == 0.5
+    assert (metrics.num_ttft_samples, metrics.num_audio_ttfp_samples, metrics.num_audio_rtf_samples) == (1, 1, 1)
+    assert metrics.request_goodput == 1.0
+
+
+def test_all_unmeasured_duplex_latency_is_not_reported_as_zero():
+    listen_only = _make_output(100, output_tokens=0)
+    listen_only.duplex_session_metrics = {
+        "mean_ttft_ms": None,
+        "mean_ttfp_ms": None,
+        "mean_rtf": None,
+    }
+
+    metrics = _calculate_test_metrics(
+        [listen_only],
+        {"ttft": float("inf"), "audio_ttft": float("inf")},
+    )
+
+    assert (metrics.num_ttft_samples, metrics.num_audio_ttfp_samples, metrics.num_audio_rtf_samples) == (0, 0, 0)
+    assert math.isnan(metrics.mean_ttft_ms)
+    assert math.isnan(metrics.mean_audio_ttfp_ms)
+    assert math.isnan(metrics.mean_audio_rtf)
+    assert metrics.request_goodput == 0.0
+
+
+def test_unmeasured_duplex_tpot_does_not_add_zero_or_misalign_goodput():
+    missing_tpot = _make_output(100, output_tokens=5)
+    missing_tpot.itl = []
+    missing_tpot.text_latency = missing_tpot.ttft = 0.1
+    missing_tpot.tpot_measured = False
+    missing_tpot.duplex_session_metrics = {"mean_ttft_ms": 100.0}
+
+    slow_ttft = _make_output(100, output_tokens=5)
+    slow_ttft.ttft = 1.0
+    slow_ttft.text_latency = 1.4
+    slow_ttft.itl = [0.1] * 4
+    slow_ttft.duplex_session_metrics = {"mean_ttft_ms": 1000.0}
+
+    metrics = _calculate_test_metrics(
+        [missing_tpot, slow_ttft],
+        {"ttft": 500.0, "tpot": 200.0},
+    )
+
+    assert metrics.num_tpot_samples == 1
+    assert metrics.mean_tpot_ms == pytest.approx(100.0)
+    assert metrics.request_goodput == 0.0
+
+
+def test_all_unmeasured_duplex_token_timing_is_not_reported_as_zero():
+    output = _make_output(100, output_tokens=5)
+    output.itl = []
+    output.text_latency = output.ttft
+    output.tpot_measured = False
+    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}
+
+    metrics = _calculate_test_metrics([output], {"tpot": float("inf")})
+
+    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (0, 0)
+    assert math.isnan(metrics.mean_tpot_ms)
+    assert math.isnan(metrics.mean_itl_ms)
+    assert metrics.request_goodput == 0.0
+
+
+def test_duplex_response_timings_do_not_build_a_session_token_timeline():
+    output = _make_output(100, output_tokens=5)
+    output.latency = 101.0
+    output.duplex_request_metrics = [
+        {"response_id": "r1", "stage0_tokens": {"itls_ms": [100.0, 100.0]}},
+        {"response_id": "r2", "stage0_tokens": {"itls_ms": [100.0, 100.0]}},
+    ]
+    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}
+
+    metrics = _calculate_test_metrics([output])
+
+    assert math.isnan(metrics.max_output_tokens_per_s)
+    assert metrics.max_concurrent_requests == 1
+    assert metrics.mean_tpot_ms == pytest.approx(100.0)
+    assert metrics.mean_itl_ms == pytest.approx(100.0)
+
+
+def test_unmeasured_tpot_stays_missing_after_tokenizer_fallback():
+    output = _make_output(100, output_tokens=0)
+    output.generated_text = "timing metadata missing"
+    output.itl = []
+    output.text_latency = output.ttft
+    output.tpot_measured = False
+    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}
+
+    def tokenizer(text, *, add_special_tokens):
+        assert text == output.generated_text
+        assert add_special_tokens is False
+        return type("Tokenized", (), {"input_ids": [1, 2, 3]})()
+
+    metrics, _ = calculate_metrics(
+        input_requests=[],
+        outputs=[output],
+        dur_s=1.0,
+        tokenizer=tokenizer,
+        selected_percentiles=[50.0],
+        goodput_config_dict={"tpot": float("inf")},
+        task_type=TaskType.GENERATION,
+        selected_percentile_metrics=[],
+        max_concurrency=None,
+        request_rate=float("inf"),
+        benchmark_duration=1.0,
+    )
+
+    assert metrics.total_output == 3
+    assert metrics.num_tpot_samples == 0
+    assert math.isnan(metrics.mean_tpot_ms)
+    assert metrics.request_goodput == 0.0
+
+
+def test_measured_zero_itl_is_not_treated_as_missing():
+    output = _make_output(100, output_tokens=3)
+    output.itl = [0.0, 0.0]
+    output.duplex_session_metrics = {"mean_ttft_ms": 100.0}
+
+    metrics = _calculate_test_metrics([output], {"tpot": 1.0})
+
+    assert (metrics.num_tpot_samples, metrics.num_itl_samples) == (1, 2)
+    assert metrics.mean_tpot_ms == 0.0
+    assert metrics.mean_itl_ms == 0.0
+    assert metrics.request_goodput == 1.0
+
+
+def test_duplex_goodput_does_not_pair_measurements_from_different_requests():
+    text_only, audio_only = _make_output(100), _make_output(100)
+    text_only.ttft, text_only.audio_ttfp = 0.1, 0.0
+    text_only.duplex_session_metrics = {"mean_ttft_ms": 100.0, "mean_ttfp_ms": None, "mean_rtf": None}
+    audio_only.ttft, audio_only.audio_ttfp = 0.0, 0.2
+    audio_only.duplex_session_metrics = {"mean_ttft_ms": None, "mean_ttfp_ms": 200.0, "mean_rtf": None}
+
+    metrics = _calculate_test_metrics([text_only, audio_only], {"ttft": 500.0, "audio_ttft": 500.0})
+
+    assert metrics.request_goodput == 0.0
+
+
 # ============================================================================
 # TTFT suppression for pure-audio (TTS) benchmarks
 # ============================================================================

tests/benchmarks/test_omniinteract.py

@@ -0,0 +1,778 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import base64
+import io
+import json
+import subprocess
+import tarfile
+import threading
+import time
+import wave
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+
+import pytest
+from vllm.benchmarks.lib.endpoint_request_func import RequestFuncInput
+from vllm.benchmarks.serve import TaskType
+
+from vllm_omni.benchmarks import omniinteract as oi
+from vllm_omni.benchmarks import serve as benchmark_serve
+from vllm_omni.benchmarks.data_modules import omniinteract_dataset as data
+from vllm_omni.benchmarks.metrics.metrics import calculate_metrics
+from vllm_omni.benchmarks.patch import patch as benchmark_patch
+from vllm_omni.entrypoints.cli.benchmark.cli_args import preprocess_serve_args
+from vllm_omni.entrypoints.cli.benchmark.serve import OmniBenchmarkServingSubcommand
+from vllm_omni.experimental.fullduplex.client import RealtimeEventCollector
+from vllm_omni.utils.tracking_parser import TrackingArgumentParser
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.benchmark]
+MODEL = "openbmb/MiniCPM-o-4_5"
+
+
+def _case(tmp_path: Path, *, subset: str = "1q1a", name: str = "video.mp4") -> data.OmniInteractCase:
+    video, annotation = tmp_path / name, tmp_path / f"{Path(name).stem}.json"
+    video.parent.mkdir(parents=True, exist_ok=True)
+    video.touch()
+    annotation.write_text("{}")
+    return data.OmniInteractCase(subset, name, video, annotation, "multi_turn")
+
+
+def _collector(*events: tuple[dict[str, object], float]) -> RealtimeEventCollector:
+    collector = RealtimeEventCollector()
+    for event, received_at in events:
+        collector.add(event, received_at_s=received_at)
+    return collector
+
+
+def _created(response_id: str = "r1") -> dict[str, object]:
+    return {"type": "response.created", "response": {"id": response_id}}
+
+
+def _done(response_id: str = "r1", status: str = "completed") -> dict[str, object]:
+    return {"type": "response.done", "response": {"id": response_id, "status": status}}
+
+
+def _audio(
+    response_id: str = "r1",
+    *,
+    samples: int = 2400,
+    rate: int | None = 24_000,
+    value: int = 1,
+) -> dict[str, object]:
+    event: dict[str, object] = {
+        "type": "response.audio.delta",
+        "response_id": response_id,
+        "format": "pcm16",
+        "delta": base64.b64encode(bytes((value, 0)) * samples).decode(),
+    }
+    if rate is not None:
+        event["sample_rate_hz"] = rate
+    return event
+
+
+def _text(response_id: str = "r1", value: str = "hello") -> dict[str, object]:
+    return {"type": "response.audio_transcript.delta", "response_id": response_id, "delta": value}
+
+
+def _listen(*, buffering: bool = False) -> dict[str, object]:
+    return {
+        "type": "response.listen",
+        "response": {"metadata": {"model_listen": True, "buffering": buffering}},
+    }
+
+
+def _session_options(tmp_path: Path) -> data.OmniInteractSessionOptions:
+    return data.OmniInteractSessionOptions(tmp_path / "artifacts", 10.0, 5.0, "reference.wav", True)
+
+
+def _sample(case: data.OmniInteractCase, options: data.OmniInteractSessionOptions):
+    return data.OmniInteractSampleRequest(
+        prompt="",
+        prompt_len=0,
+        expected_output_len=0,
+        multi_modal_data=None,
+        request_id="measured-0",
+        omniinteract_case=case,
+        omniinteract_options=options,
+    )
+
+
+def _write_dataset(root: Path) -> Path:
+    data_root = root / "data"
+    for subset in ("1q1a", "1q1a_math"):
+        subset_root = data_root / subset
+        (subset_root / "videos").mkdir(parents=True)
+        (subset_root / "annotations").mkdir()
+        (subset_root / "videos" / f"{subset}.mp4").touch()
+        (subset_root / "annotations" / f"{subset}.json").write_text("{}")
+        (subset_root / "video_json_map.json").write_text(
+            json.dumps(
+                {
+                    "entries": [
+                        {
+                            "video": f"videos/{subset}.mp4",
+                            "annotation": f"annotations/{subset}.json",
+                            "scene_type": "multi_turn",
+                        }
+                    ]
+                }
+            )
+        )
+    one_to_many = data_root / "1qna"
+    (one_to_many / "videos_bench" / "nested").mkdir(parents=True)
+    (one_to_many / "annotations" / "nested").mkdir(parents=True)
+    (one_to_many / "videos_bench" / "nested" / "guide.mp4").touch()
+    (one_to_many / "annotations" / "nested" / "guide.json").write_text("{}")
+    return data_root
+
+
+def _successful_output(
+    tmp_path: Path,
+    *,
+    audio_time: float = 10.0,
+    status: str = "completed",
+) -> tuple[data.OmniInteractCase, RealtimeEventCollector, oi.OmniInteractCaseResult]:
+    case = _case(tmp_path)
+    collector = _collector(
+        (_created(), 9.9),
+        (_audio(), audio_time),
+        (_text(), audio_time + 0.01),
+        (_done(status=status), audio_time + 0.1),
+    )
+    result = oi.OmniInteractCaseResult(case.subset, str(case.video_path), "")
+    return case, collector, result
+
+
+def _write_success(
+    root: Path,
+    case: data.OmniInteractCase,
+    collector: RealtimeEventCollector,
+    result: oi.OmniInteractCaseResult,
+    **kwargs,
+):
+    require_response = kwargs.pop("require_response", True)
+    return oi.prepare_success_artifacts(
+        root,
+        case,
+        collector,
+        stream_start=10.0,
+        video_duration_s=1.0,
+        require_response=require_response,
+        result=result,
+        **kwargs,
+    )
+
+
+def test_dataset_discovers_official_layouts_and_total_selection(tmp_path: Path):
+    root = _write_dataset(tmp_path)
+    all_cases = data.discover_omniinteract_cases(root, data.OMNIINTERACT_SUBSETS, num_prompts=0, disable_shuffle=True)
+    selected = data.discover_omniinteract_cases(root, data.OMNIINTERACT_SUBSETS, num_prompts=2, disable_shuffle=True)
+    assert [case.subset for case in all_cases] == ["1q1a", "1q1a_math", "1qna"]
+    assert all_cases[-1].video_rel == "videos_bench/nested/guide.mp4"
+    assert [case.subset for case in selected] == ["1q1a", "1q1a_math"]
+
+
+def _archive(path: Path, member: str) -> None:
+    with tarfile.open(path, "w") as archive:
+        content = b"{}"
+        item = tarfile.TarInfo(member)
+        item.size = len(content)
+        archive.addfile(item, io.BytesIO(content))
+
+
+def test_archive_is_safe_and_atomically_shared(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    bad = tmp_path / "bad.tar"
+    _archive(bad, "../escape")
+    with pytest.raises(ValueError, match="Unsafe path"):
+        data._extract_archive(bad, tmp_path / "bad-cache")
+
+    good = tmp_path / "good.tar"
+    _archive(good, "data/1q1a/video_json_map.json")
+    target = tmp_path / "cache"
+    barrier, safe_extract = threading.Barrier(2), data._safe_extract
+    monkeypatch.setattr(data, "_safe_extract", lambda *args: (barrier.wait(), safe_extract(*args)))
+    with ThreadPoolExecutor(2) as executor:
+        roots = list(executor.map(lambda _: data._extract_archive(good, target), range(2)))
+    assert roots[0] == roots[1]
+    assert (roots[0] / "1q1a" / "video_json_map.json").is_file()
+    assert not list(target.glob(".tmp-*"))
+
+
+def test_hub_archive_uses_vllm_filesystem(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    source = tmp_path / "source.tar"
+    _archive(source, "data/1q1a/video_json_map.json")
+
+    class FS:
+        def get_file(self, remote: str, local: str) -> None:
+            assert remote == "datasets/org/repo/data.tar.gz"
+            Path(local).write_bytes(source.read_bytes())
+
+    monkeypatch.setenv("HF_HOME", str(tmp_path / "hf"))
+    monkeypatch.setattr(data, "hf_fs", lambda: FS())
+    assert (data.resolve_omniinteract_root(None, "org/repo") / "1q1a").is_dir()
+
+
+def test_media_commands_are_bounded(monkeypatch: pytest.MonkeyPatch, tmp_path: Path):
+    calls: list[float] = []
+    commands: list[list[str]] = []
+    bounded_run = oi._run_media_command
+
+    def run(command, *, timeout_s, text=False):
+        calls.append(timeout_s)
+        commands.append(command)
+        if command[0] == "ffprobe":
+            return subprocess.CompletedProcess(command, 0, "1.0", "")
+        return subprocess.CompletedProcess(command, 0, b"", b"")
+
+    monkeypatch.setattr(oi, "_run_media_command", run)
+    duration, pcm, frames = oi.prepare_media(tmp_path / "video.mp4", 1.0, timeout_s=3.0, max_duration_s=1.0)
+    assert (duration, len(pcm), frames, calls) == (1.0, 32_000, [None], [3.0, 3.0, 3.0])
+    assert commands[1][commands[1].index("-t") + 1] == "1.0"
+    assert commands[1][commands[1].index("-fs") + 1] == "32000"
+
+    monkeypatch.setattr(subprocess, "run", lambda *a, **k: (_ for _ in ()).throw(subprocess.TimeoutExpired("x", 3)))
+    with pytest.raises(TimeoutError, match="timed out"):
+        bounded_run(["ffmpeg"], timeout_s=3.0)
+
+
+def test_media_duration_is_bounded_before_decode(monkeypatch: pytest.MonkeyPatch, tmp_path: Path):
+    calls = 0
+
+    def run(command, *, timeout_s, text=False):
+        nonlocal calls
+        calls += 1
+        return subprocess.CompletedProcess(command, 0, "601.0", "")
+
+    monkeypatch.setattr(oi, "_run_media_command", run)
+    with pytest.raises(ValueError, match="Invalid video duration"):
+        oi.prepare_media(tmp_path / "video.mp4", 1.0, timeout_s=3.0, max_duration_s=600.0)
+    assert calls == 1
+
+
+def test_final_commit_keeps_a_partial_model_unit():
+    """Exact units retain a sample for the server's correlated final flush."""
+
+    events = [{"session": {"capabilities": {"chunk_period_ms": 1000}}}]
+    exact_unit = b"\0\0" * 16_000
+    assert oi._ensure_final_commit_tail(exact_unit, events) == exact_unit[:-2]
+    assert oi._ensure_final_commit_tail(exact_unit + b"\0\0", events) == exact_unit + b"\0\0"
+
+
+@pytest.mark.parametrize(
+    ("events", "match"),
+    [
+        (((_done(), 0.0),), "without response.created"),
+        (((_created(), 0.0), (_created(), 0.1)), "duplicate response.created"),
+        (((_created(), 0.0), (_done(), 0.1), (_done(), 0.2)), "duplicate response.done"),
+        (
+            (
+                (_created(), 0.0),
+                ({"type": "response.done", "response": {"id": "r1", "status_details": {"type": "failed"}}}, 0.1),
+            ),
+            "reports failure",
+        ),
+    ],
+)
+def test_response_ledger_rejects_identity_errors(events, match: str):
+    with pytest.raises(ValueError, match=match):
+        oi.response_ledger(_collector(*events))
+
+
+class _CompletionClient:
+    def __init__(self, collector: RealtimeEventCollector):
+        self.events, self.acks = collector, []
+
+    def raise_if_reader_stopped(self) -> None:
+        return None
+
+    async def send_playback_ack(self, response_id: str, played_ms: int) -> None:
+        self.acks.append((response_id, played_ms))
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("tail", "succeeds"),
+    [
+        (((_created(), 1.1), (_done(), 1.2)), True),
+        (((_listen(), 1.1),), True),
+        (((_listen(buffering=True), 1.1),), False),
+        ((({"type": "session.closed", "reason": "timeout"}, 1.1),), False),
+    ],
+)
+async def test_completion_requires_a_final_model_decision(tail, succeeds: bool):
+    events = _collector(({"type": "input_audio_buffer.committed"}, 1.0), *tail)
+    coroutine = oi.wait_for_session_completion(
+        _CompletionClient(events), oi._Playback(), commit_from=0, timeout_s=0.08, settle_s=0.0
+    )
+    if succeeds:
+        assert await coroutine == 0
+    else:
+        with pytest.raises((TimeoutError, RuntimeError)):
+            await coroutine
+
+
+@pytest.mark.asyncio
+async def test_deferred_commit_ignores_the_prior_response_terminal():
+    events = _collector(
+        (_created("old"), 0.0),
+        ({"type": "input_audio_buffer.committed", "event": {"overlap_deferred": True}}, 1.0),
+        (_done("old"), 1.1),
+        (_listen(), 1.2),
+    )
+    assert (
+        await oi.wait_for_session_completion(
+            _CompletionClient(events), oi._Playback(), commit_from=1, timeout_s=0.1, settle_s=0.0
+        )
+        == 1
+    )
+
+
+@pytest.mark.asyncio
+async def test_playback_acks_only_after_serial_audio_drain():
+    collector = _collector((_audio(samples=2400), 10.0), (_done(), 10.01))
+    client, playback = _CompletionClient(collector), oi._Playback()
+    await playback.acknowledge(client, now=10.05)
+    assert client.acks == []
+    await playback.acknowledge(client, now=10.1)
+    assert client.acks == [("r1", 100)]
+
+
+@pytest.mark.parametrize(
+    ("event", "match"),
+    [
+        (_audio(rate=16_000), "24000 Hz"),
+        ({**_audio(), "format": "opus"}, "pcm16"),
+        ({**_audio(), "delta": "!"}, "base64"),
+    ],
+)
+def test_playback_rejects_non_official_audio(event: dict[str, object], match: str):
+    playback = oi._Playback()
+    with pytest.raises(ValueError, match=match):
+        playback.ingest(_collector((event, 1.0)))
+
+
+def test_artifacts_publish_official_bundle_and_sparse_deferred_state(tmp_path: Path):
+    case, collector, result = _successful_output(tmp_path)
+    summary = _write_success(
+        tmp_path / "out",
+        case,
+        collector,
+        result,
+    )
+    context = result._artifact_context
+    assert result.success and result.eligible_for_official_eval
+    assert summary["scene_type"] == "multi_turn"
+    assert context and context.spans
+    assert all("delta" not in event for event in context.events if event["type"] == "response.audio.delta")
+    oi.publish_deferred_case_artifacts(tmp_path / "out", case, result)
+    directory = oi._output_dir(tmp_path / "out", case)
+    assert all((directory / name).is_file() for name in oi.SUCCESS_ARTIFACTS)
+    with wave.open(str(directory / "output.wav")) as output:
+        assert (output.getframerate(), output.getnframes()) == (24_000, 24_000)
+    transcript = json.loads((directory / "wav_transcript.json").read_text())
+    assert transcript["text"] == "hello" and transcript["chunks"][0]["timestamp"] == [0.0, 0.1]
+    assert "serialized playback queue" in transcript["timestamp_semantics"]
+    assert result._artifact_context is None
+
+
+@pytest.mark.parametrize(
+    ("audio_time", "status", "reason"),
+    [(10.95, "completed", "audio_clipped"), (10.0, "cancelled", "cancelled_response")],
+)
+def test_ineligible_outputs_are_excluded_from_manifest(tmp_path: Path, audio_time: float, status: str, reason: str):
+    case, collector, result = _successful_output(tmp_path, audio_time=audio_time, status=status)
+    _write_success(tmp_path / "out", case, collector, result, require_response=False)
+    assert reason in result.official_eval_ineligible_reasons
+    oi.write_batch_artifacts(tmp_path / "out", [case], [result])
+    assert (tmp_path / "out" / "official_eval_manifest.jsonl").read_text() == ""
+
+
+def test_artifacts_require_complete_audio_and_transcript_response(tmp_path: Path):
+    case, _, result = _successful_output(tmp_path)
+    collector = _collector((_created(), 10.0), (_audio(), 10.0), (_done(), 10.1))
+    with pytest.raises(ValueError, match="audio and transcript"):
+        _write_success(tmp_path, case, collector, result)
+
+
+def test_publication_failure_removes_partial_success_bundle(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case, collector, result = _successful_output(tmp_path)
+    _write_success(tmp_path, case, collector, result)
+    original = oi._atomic_write_json
+
+    def fail(path: Path, value: object) -> None:
+        if path.name == "events.json":
+            raise OSError("disk full")
+        original(path, value)
+
+    monkeypatch.setattr(oi, "_atomic_write_json", fail)
+    with pytest.raises(OSError, match="disk full"):
+        oi.publish_deferred_case_artifacts(tmp_path, case, result)
+    directory = oi._output_dir(tmp_path, case)
+    assert not any((directory / name).exists() for name in oi.SUCCESS_ARTIFACTS)
+    assert not result.success and "artifact_write_failed" in result.official_eval_ineligible_reasons
+
+
+def test_atomic_write_preserves_destination_on_failure(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    destination = tmp_path / "output.wav"
+    destination.write_bytes(b"old")
+    monkeypatch.setattr(oi, "write_pcm16_wav", lambda *a, **k: (_ for _ in ()).throw(OSError("write")))
+    with pytest.raises(OSError):
+        oi._atomic_write_wav(destination, b"\0\0", 24_000)
+    assert destination.read_bytes() == b"old"
+    assert not list(tmp_path.glob(".*.tmp"))
+
+
+class _RealtimeClient:
+    instances: list[_RealtimeClient] = []
+
+    def __init__(self, url: str, **kwargs):
+        self.url, self.events, self.acks, self.configure_kwargs = url, RealtimeEventCollector(), [], {}
+        self.instances.append(self)
+
+    async def __aenter__(self):
+        return self
+
+    async def __aexit__(self, *args):
+        return None
+
+    async def configure(self, model: str, **kwargs) -> None:
+        self.configure_kwargs = kwargs
+        self.events.add({"type": "session.created", "session": {"capabilities": {"chunk_period_ms": 1000}}})
+
+    async def send(self, event: dict[str, object]) -> None:
+        return None
+
+    async def commit(self) -> None:
+        now = time.monotonic()
+        for event, offset in (
+            ({"type": "input_audio_buffer.committed"}, 0.0),
+            (_created(), 0.001),
+            (_audio(samples=24), 0.002),
+            (_text(value="answer"), 0.003),
+            (_done(), 0.004),
+        ):
+            self.events.add(event, received_at_s=now + offset)
+
+    async def send_playback_ack(self, response_id: str, played_ms: int) -> None:
+        self.acks.append((response_id, played_ms))
+
+    async def close_session(self, **kwargs) -> None:
+        self.events.add({"type": "session.closed"})
+
+    def raise_if_reader_stopped(self) -> None:
+        return None
+
+
+@pytest.mark.asyncio
+async def test_public_runner_executes_one_prepared_session(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case = _case(tmp_path)
+    config = oi.OmniInteractBenchmarkConfig(
+        model=MODEL,
+        output_root=tmp_path / "out",
+        ref_audio="ref.wav",
+        require_response=True,
+        extra_body={"custom": "value"},
+    )
+    prepared = data.OmniInteractPreparedInput(1.0, b"\0\0\0\0", ("frame",), "data:audio/wav;base64,ref")
+    monkeypatch.setattr(oi, "RealtimeDuplexClient", _RealtimeClient)
+    monkeypatch.setattr(oi, "_COMPLETION_SETTLE_S", 0.0)
+    result = await oi.run_omniinteract_case(
+        case,
+        config,
+        request_index=0,
+        prepared_input=prepared,
+    )
+    assert result.success and result.transcript == "answer" and result._artifact_context is not None
+    assert not config.output_root.exists()
+    assert "autostart=0" in _RealtimeClient.instances[-1].url
+    assert _RealtimeClient.instances[-1].configure_kwargs["extra_body"] == {"custom": "value"}
+    assert _RealtimeClient.instances[-1].acks == [("r1", 1)]
+
+
+def test_standard_sample_loading_prepares_media_before_timing(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    _write_dataset(tmp_path)
+    ref = tmp_path / "ref.wav"
+    ref.touch()
+    parser = TrackingArgumentParser()
+    OmniBenchmarkServingSubcommand.add_cli_args(parser)
+    args = parser.parse_args(
+        [
+            "--backend",
+            "openai-realtime-duplex",
+            "--dataset-name",
+            "omniinteract",
+            "--dataset-path",
+            str(tmp_path),
+            "--model",
+            MODEL,
+            "--endpoint",
+            "/v1/realtime",
+            "--num-prompts",
+            "0",
+            "--omniinteract-ref-audio",
+            str(ref),
+        ]
+    )
+    preprocess_serve_args(args)
+    assert args.endpoint == "/v1/realtime"
+    monkeypatch.setattr(benchmark_patch, "reference_audio_data_url", lambda _: "data:audio/wav;base64,ref")
+    monkeypatch.setattr(benchmark_patch, "prepare_media", lambda *a, **k: (1.0, b"pcm", ["frame"]))
+    samples = benchmark_patch.get_samples(args, None)
+    assert len(samples) == args.num_prompts == 3
+    assert all(sample.omniinteract_prepared_input.video_frames == ("frame",) for sample in samples)
+
+
+def _request(case: data.OmniInteractCase, options: data.OmniInteractSessionOptions) -> RequestFuncInput:
+    request = RequestFuncInput(
+        model=MODEL,
+        model_name=MODEL,
+        prompt="",
+        api_url="http://server:8000/v1/realtime",
+        prompt_len=0,
+        output_len=0,
+        logprobs=None,
+        multi_modal_content=None,
+        ignore_eos=False,
+        request_id="measured-0",
+    )
+    request.omniinteract_case, request.omniinteract_options = case, options
+    return request
+
+
+@pytest.mark.asyncio
+async def test_adapter_requires_and_forwards_exact_prepared_payload(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case, options = _case(tmp_path), _session_options(tmp_path)
+    request = _request(case, options)
+    output = await benchmark_patch.async_request_openai_realtime_duplex(request, session=None)
+    assert not output.success and "prepared before benchmark request timing" in output.error
+
+    prepared = data.OmniInteractPreparedInput(1.0, b"pcm", ("frame",), "data:audio/wav;base64,ref")
+    request.omniinteract_prepared_input = prepared
+    request.extra_headers = {"X-Proxy": "test"}
+    request.extra_body = {"custom": "value"}
+    monkeypatch.setenv("OPENAI_API_KEY", "secret")
+    seen = {}
+
+    async def run(_case_arg, config, **kwargs):
+        seen["headers"] = config.extra_headers
+        seen["extra_body"] = config.extra_body
+        seen.update(kwargs)
+        return oi.OmniInteractCaseResult(
+            case.subset,
+            str(case.video_path),
+            str(options.output_root),
+            success=True,
+            transcript="timing metadata missing",
+            output_tokens=0,
+            duplex_request_metrics=[{"request_metrics": {"ttft_ms": 1.0}}],
+            duplex_session_metrics={"mean_ttft_ms": 10.0},
+        )
+
+    monkeypatch.setattr(benchmark_patch, "run_omniinteract_case", run)
+    output = await benchmark_patch.async_request_openai_realtime_duplex(request, session=None)
+    assert output.success and seen["prepared_input"] is prepared
+    assert seen["headers"] == {"Authorization": "Bearer secret", "X-Proxy": "test", "x-request-id": "measured-0"}
+    assert seen["extra_body"] == {"custom": "value"}
+    assert output.itl == []
+    assert output.tpot_measured is False
+    assert seen["capture_artifacts"] is True
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("request_metrics", "expected_itl", "expected_text_latency"),
+    [
+        (
+            [
+                {"stage0_tokens": {"output_token_count": 3, "itls_ms": [10.0, 20.0]}},
+                {"stage0_tokens": {"output_token_count": 2, "itls_ms": [30.0]}},
+            ],
+            [0.01, 0.02, 0.03],
+            0.16,
+        ),
+        (
+            [
+                {"stage0_tokens": {"output_token_count": 3, "tpot_ms": 10.0}},
+                {"stage0_tokens": {"output_token_count": 2, "tpot_ms": 40.0}},
+            ],
+            [],
+            0.18,
+        ),
+        (
+            [
+                {"stage0_tokens": {"output_token_count": 3, "tpot_ms": 0.0}},
+                {"stage0_tokens": {"output_token_count": 2, "tpot_ms": 0.0}},
+            ],
+            [],
+            0.1,
+        ),
+    ],
+)
+async def test_adapter_reports_exact_or_weighted_token_timing(
+    tmp_path: Path,
+    monkeypatch: pytest.MonkeyPatch,
+    request_metrics: list[dict[str, object]],
+    expected_itl: list[float],
+    expected_text_latency: float,
+):
+    case, options = _case(tmp_path), _session_options(tmp_path)
+    request = _request(case, options)
+    request.omniinteract_prepared_input = data.OmniInteractPreparedInput(
+        1.0, b"pcm", ("frame",), "data:audio/wav;base64,ref"
+    )
+
+    async def run(*args, **kwargs):
+        return oi.OmniInteractCaseResult(
+            case.subset,
+            str(case.video_path),
+            str(options.output_root),
+            success=True,
+            output_tokens=5,
+            duplex_request_metrics=request_metrics,
+            duplex_session_metrics={"mean_ttft_ms": 100.0},
+        )
+
+    monkeypatch.setattr(benchmark_patch, "run_omniinteract_case", run)
+    output = await benchmark_patch.async_request_openai_realtime_duplex(request, session=None)
+
+    assert output.success
+    assert output.itl == expected_itl
+    assert output.text_latency == pytest.approx(expected_text_latency)
+    assert output.tpot_measured is True
+
+
+def test_batch_finalization_publishes_only_measured_results(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case, options = _case(tmp_path), _session_options(tmp_path)
+    options.output_root.mkdir()
+    sample = _sample(case, options)
+    result = oi.OmniInteractCaseResult(
+        case.subset, str(case.video_path), str(options.output_root), success=True, eligible_for_official_eval=True
+    )
+    output = benchmark_patch.MixRequestFuncOutput(success=True)
+    output.omniinteract_case_result = result
+    published = []
+    monkeypatch.setattr(benchmark_patch, "publish_deferred_case_artifacts", lambda *args: published.append(args))
+    summary = benchmark_patch._finalize_omniinteract_batch([sample], [output])
+    assert summary and summary["total"] == summary["success"] == summary["eligible_for_official_eval"] == 1
+    assert summary["failed"] == summary["successful_but_ineligible"] == summary["audio_clipped_bytes"] == 0
+    assert published == [(options.output_root, case, result)]
+
+
+def test_case_artifact_failures_do_not_discard_serving_metrics(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case, options = _case(tmp_path), _session_options(tmp_path)
+    sample = _sample(case, options)
+    result = oi.OmniInteractCaseResult(
+        case.subset, str(case.video_path), str(options.output_root), success=True, eligible_for_official_eval=True
+    )
+    output = benchmark_patch.MixRequestFuncOutput(success=True, error="inference warning")
+    output.omniinteract_case_result = result
+    monkeypatch.setattr(
+        benchmark_patch,
+        "publish_deferred_case_artifacts",
+        lambda *args: (_ for _ in ()).throw(OSError("case disk full")),
+    )
+    monkeypatch.setattr(
+        benchmark_patch,
+        "write_failure_artifacts",
+        lambda *args: (_ for _ in ()).throw(OSError("failure disk full")),
+    )
+    monkeypatch.setattr(benchmark_patch, "write_omniinteract_batch_artifacts", lambda *args: None)
+
+    summary = benchmark_patch._finalize_omniinteract_batch([sample], [output])
+    metrics, _ = calculate_metrics(
+        [],
+        [output],
+        1.0,
+        None,
+        [50.0],
+        {},
+        TaskType.GENERATION,
+        [],
+        None,
+        float("inf"),
+        1.0,
+    )
+
+    assert summary and summary["failed"] == 1
+    assert summary["artifacts_complete"] is False
+    assert len(summary["artifact_errors"]) == 2
+    assert output.success and metrics.completed == 1
+    assert output.error.startswith("inference warning\n")
+    assert "Artifact publication failed: case disk full" in output.error
+
+
+def test_batch_artifact_failure_removes_partial_batch_files(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    case, options = _case(tmp_path), _session_options(tmp_path)
+    sample = _sample(case, options)
+    result = oi.OmniInteractCaseResult(
+        case.subset, str(case.video_path), str(options.output_root), success=True, eligible_for_official_eval=True
+    )
+    output = benchmark_patch.MixRequestFuncOutput(success=True)
+    output.omniinteract_case_result = result
+    monkeypatch.setattr(benchmark_patch, "publish_deferred_case_artifacts", lambda *args: None)
+
+    def fail_batch(root, *args):
+        root.mkdir(parents=True, exist_ok=True)
+        (root / "batch_summary.json").write_text("partial")
+        raise OSError("batch disk full")
+
+    monkeypatch.setattr(benchmark_patch, "write_omniinteract_batch_artifacts", fail_batch)
+
+    summary = benchmark_patch._finalize_omniinteract_batch([sample], [output])
+
+    assert summary and summary["success"] == 1 and summary["artifacts_complete"] is False
+    assert summary["artifact_errors"] == ["Batch artifact publication failed: batch disk full"]
+    assert output.success
+    assert not (options.output_root / "batch_summary.json").exists()
+
+
+@pytest.mark.parametrize(
+    ("field", "value", "match"),
+    [
+        ("omniinteract_ref_audio", None, "ref-audio"),
+        ("endpoint", "/v1/completions", "v1/realtime"),
+        ("max_concurrency", 0, "positive"),
+        ("skip_tokenizer_init", True, "skip-tokenizer-init"),
+    ],
+)
+def test_cli_rejects_incompatible_omniinteract_options(field: str, value, match: str):
+    args = argparse.Namespace(
+        dataset_name="omniinteract",
+        backend="openai-realtime-duplex",
+        endpoint="/v1/realtime",
+        omniinteract_ref_audio="ref.wav",
+        max_concurrency=1,
+        skip_tokenizer_init=False,
+    )
+    setattr(args, field, value)
+    with pytest.raises(ValueError, match=match):
+        preprocess_serve_args(args)
+
+
+def test_whole_benchmark_lock_serializes_a_shared_output_root(tmp_path: Path, monkeypatch: pytest.MonkeyPatch):
+    entered, release = threading.Event(), threading.Event()
+    calls: list[int] = []
+
+    async def fake_main_async(args):
+        calls.append(len(calls))
+        if len(calls) == 1:
+            entered.set()
+            await asyncio.to_thread(release.wait)
+        return {"total": 1}
+
+    monkeypatch.setattr(benchmark_serve, "main_async", fake_main_async)
+    args = argparse.Namespace(dataset_name="omniinteract", omniinteract_output_dir=tmp_path, extra_body=None)
+    with ThreadPoolExecutor(2) as executor:
+        first = executor.submit(benchmark_serve.main, args)
+        assert entered.wait(1)
+        second = executor.submit(benchmark_serve.main, args)
+        time.sleep(0.05)
+        assert calls == [0]
+        release.set()
+        assert first.result() == second.result() == {"total": 1}
+    assert calls == [0, 1]

tests/benchmarks/test_serve_cli.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 import argparse
 import json
 import os
@@ -22,6 +25,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu, pytest.mark.benchmark]
     "dataset_name",
     [
         "daily-omni",
+        "omniinteract",
         "seed-tts",
         "seed-tts-text",
         "seed-tts-design",
@@ -55,6 +59,7 @@ def test_extend_omni_choices_updates_tracking_parser_shadow(dataset_name: str) -
     [
         (["--print-stage"], "print_stage", True),
         (["--daily-omni-input-mode", "audio"], "daily_omni_input_mode", "audio"),
+        (["--omniinteract-subsets", "1qna"], "omniinteract_subsets", ["1qna"]),
         (["--seed-tts-locale", "zh"], "seed_tts_locale", "zh"),
     ],
 )
@@ -80,12 +85,14 @@ def test_add_omni_args_preserves_implicit_defaults() -> None:
     args = parser.parse_args([])
     assert args.print_stage is False
     assert args.daily_omni_input_mode == "all"
+    assert args.omniinteract_subsets == ["1q1a", "1q1a_math", "1qna"]
     assert args.seed_tts_locale == "en"
     assert args.explicit_keys == set()
 
 
 def test_update_omni_help_updates_upstream_actions() -> None:
     parser = TrackingArgumentParser()
+    parser.add_argument("--num-prompts", default=1000, help="Number of prompts.")
     parser.add_argument("--percentile-metrics", help="Upstream percentile help.")
     parser.add_argument("--random-mm-limit-mm-per-prompt", help="Upstream limit help.")
     parser.add_argument("--random-mm-bucket-config", help="Upstream bucket help.")
@@ -93,6 +100,7 @@ def test_update_omni_help_updates_upstream_actions() -> None:
     update_omni_help(parser)
 
     actions = {action.dest: action for action in parser._actions}
+    assert "OmniInteract uses 3" in actions["num_prompts"].help
     assert all(metric in actions["percentile_metrics"].help for metric in ("ttfc", "tpop", "audio_rtf"))
     assert "probabilities are renormalized" in actions["random_mm_limit_mm_per_prompt"].help
     assert "Currently allows for 3 modalities" in actions["random_mm_bucket_config"].help
@@ -120,6 +128,31 @@ def test_preprocess_serve_args_merges_bot_task_without_overriding_extra_body(
     assert args.extra_body == expected
 
 
+@pytest.mark.parametrize(
+    ("argv", "expected"),
+    [
+        ([], 3),
+        (["--num-prompts", "0"], 0),
+        (["--num-prompts", "12"], 12),
+    ],
+)
+def test_preprocess_serve_args_applies_safe_omniinteract_prompt_default(
+    argv: list[str],
+    expected: int,
+) -> None:
+    parser = TrackingArgumentParser()
+    parser.add_argument("--dataset-name", default="omniinteract")
+    parser.add_argument("--backend", default="openai-realtime-duplex")
+    parser.add_argument("--endpoint", default="/v1/realtime")
+    parser.add_argument("--omniinteract-ref-audio", default="reference.wav")
+    parser.add_argument("--num-prompts", type=int, default=1000)
+    args = parser.parse_args(argv)
+
+    preprocess_serve_args(args)
+
+    assert args.num_prompts == expected
+
+
 @pytest.mark.parametrize(
     ("argv", "expected_extra_body", "expected_explicit"),
     [

tests/config/test_config_factory.py

@@ -1109,8 +1109,35 @@ class TestDeployConfigLoading:
         with pytest.raises(ValueError, match=r"stage_args.*PipelineConfig.*stages"):
             load_deploy_config(deploy_path)
 
-    def test_load_minicpmo_duplex_deploy_config(self):
-        deploy_path = Path(get_deploy_config_path("minicpmo_4_5_duplex.yaml"))
+    @pytest.mark.parametrize(
+        ("filename", "max_sessions"),
+        [
+            ("minicpmo_4_5.yaml", 4),
+            ("minicpmo_4_5_2gpu.yaml", 4),
+            ("minicpmo_4_5_3gpu.yaml", 4),
+            ("minicpmo_4_5_8x4090.yaml", 1),
+            ("minicpmo_4_5_3gpu_stage1_replicas.yaml", 4),
+            ("minicpmo_4_5_4gpu_stage1_replicas.yaml", 4),
+            ("minicpmo_4_5_8x4090_stage1_replicas.yaml", 1),
+        ],
+    )
+    def test_minicpmo_deploy_configs_enable_duplex(self, filename: str, max_sessions: int):
+        """Every MiniCPM-o profile serves /v1/realtime, so the retired
+        minicpmo_4_5_duplex.yaml overlay has no replacement to fall back on.
+        """
+        deploy = load_deploy_config(Path(get_deploy_config_path(filename)))
+        pipeline = resolve_pipeline_config("minicpmo_4_5")
+        assert isinstance(pipeline, PipelineConfig)
+        stages = merge_pipeline_deploy(pipeline, deploy)
+
+        assert deploy.session_mode == "duplex"
+        assert deploy.active_stream_window == 1
+        assert deploy.duplex_session.max_sessions == max_sessions
+        assert [stage.session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
+        assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
+
+    def test_load_minicpmo_default_deploy_config(self):
+        deploy_path = Path(get_deploy_config_path("minicpmo_4_5.yaml"))
 
         deploy = load_deploy_config(deploy_path)
         pipeline = resolve_pipeline_config("minicpmo_4_5")
@@ -1118,17 +1145,15 @@ class TestDeployConfigLoading:
 
         stages = merge_pipeline_deploy(pipeline, deploy)
 
-        assert deploy.session_mode == "duplex"
         assert deploy.async_chunk is True
-        assert deploy.active_stream_window == 1
-        assert [stage.session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
-        assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
-        assert [stage.yaml_engine_args["async_scheduling"] for stage in stages] == [False, False, False]
-        assert all("Async" not in (stage.scheduler_cls or "") for stage in stages)
+        assert stages[0].yaml_engine_args["async_scheduling"] is True
+        assert stages[1].yaml_engine_args["async_scheduling"] is True
+        assert "Async" in (stages[0].scheduler_cls or "")
+        assert "Async" in (stages[1].scheduler_cls or "")
         assert [stage.devices for stage in deploy.stages] == ["0", "0", "0"]
         assert deploy.stages[1].enforce_eager is False
         assert stages[1].yaml_extras["default_sampling_params"]["max_tokens"] == 4096
-        assert stages[1].yaml_extras["default_sampling_params"]["min_tokens"] == 0
+        assert stages[1].yaml_extras["default_sampling_params"]["min_tokens"] == 50
         assert stages[1].yaml_extras["default_sampling_params"]["temperature"] == 0.8
         assert stages[1].yaml_extras["default_sampling_params"]["stop_token_ids"] == [6561]
         assert "codec_sampling_params" not in stages[1].yaml_engine_args
@@ -1304,6 +1329,7 @@ stages:
             ({"model_type": "step_audio_2"}, None, "step_audio_2"),
             (None, {"_class_name": "HunyuanVideo15Pipeline"}, "hunyuan_video_15"),
             (None, {"_class_name": "WanPipeline"}, "wan2_2_ti2v"),
+            (None, {"_class_name": "WanDMDPipeline"}, "wan2_2_ti2v"),
         ],
     )
     def test_migrated_models_are_discovered_without_explicit_deploy(
@@ -2115,7 +2141,8 @@ class TestBaseConfigInheritance:
     def test_minicpmo_overlays_inherit_talker_sampling_params(self):
         """Overlays must keep the base Talker codec Sampler knobs."""
         for filename in (
-            "minicpmo_4_5_duplex.yaml",
+            "minicpmo_4_5_2gpu.yaml",
+            "minicpmo_4_5_3gpu.yaml",
             "minicpmo_4_5_3gpu_stage1_replicas.yaml",
             "minicpmo_4_5_4gpu_stage1_replicas.yaml",
             "minicpmo_4_5_8x4090_stage1_replicas.yaml",
@@ -2127,9 +2154,6 @@ class TestBaseConfigInheritance:
             sampling = deploy.stages[1].default_sampling_params
             assert sampling is not None, f"{filename} stage 1 lost default_sampling_params"
             assert sampling["temperature"] == 0.8, filename
-            # Upstream utils.TTSSamplingParams. min_tokens is not checked here:
-            # the duplex overlay drops it to 0 because the model budgets each
-            # generate_chunk itself.
             assert sampling["top_k"] == 25, filename
             assert sampling["top_p"] == 0.85, filename
             assert sampling["repetition_penalty"] == 1.05, filename
@@ -2752,22 +2776,23 @@ class TestSentinelDefaultPrecedence:
         )
 
     def test_ming_flash_omni_topology(self):
-        """Guard ming_flash_omni's SIP cleanup: stage 0 has no full-payload
-        producer hook (arch is not in ``_FULL_PAYLOAD_INPUT_STAGES``), and
-        stage 1 uses only ``thinker2talker_token_only`` (sync_process_input_func).
+        """Guard ming_flash_omni's SIP cleanup: stage 0 has no
+        ``custom_process_next_stage_input_func``, and stage 1 does not set
+        ``requires_full_payload_input=True``. Stage 1 uses only
+        ``thinker2talker_token_only`` (sync_process_input_func).
         The dead ``thinker2talker`` (custom_process_input_func) was removed
         because ming_flash_omni has no async_chunk support and both functions
         called ``_build_talker_inputs`` identically.
-        Merge under either async_chunk mode must not re-introduce a
-        stage-0 full-payload hook."""
+        Merge under either async_chunk mode must not re-introduce either
+        full-payload transport gate."""
         pipeline = resolve_pipeline_config("ming_flash_omni")
         assert isinstance(pipeline, PipelineConfig)
 
         stage0, stage1 = pipeline.stages
         assert stage0.custom_process_next_stage_input_func is None, (
-            "ming_flash_omni stage 0 must not declare a full-payload producer "
-            "(connector path is not active for this arch)."
+            "ming_flash_omni stage 0 must not declare a full-payload producer hook."
         )
+        assert stage1.requires_full_payload_input is False
         assert stage1.custom_process_input_func is None
         assert stage1.sync_process_input_func is not None
         assert stage1.sync_process_input_func.endswith("thinker2talker_token_only")

tests/config/test_config_import_cycle.py

@@ -66,6 +66,8 @@ def test_npu_platform_defers_minimax_h3_encoder_patch() -> None:
     runtime_src = src[runtime_start:]
     assert "apply_minimax_h3_qwen3vl_patch" not in init_src
     assert "apply_minimax_h3_qwen3vl_patch" in runtime_src
+    assert "apply_minimax_h3_qwen3vl_sdpa_patch" not in init_src
+    assert "apply_minimax_h3_qwen3vl_sdpa_patch" in runtime_src
     assert "apply_minimax_h3_qwen3vl_swiglu_patch" in runtime_src
 
 

tests/config/test_environment_variables.py

@@ -0,0 +1,317 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+"""Drift checks for the reviewed environment-variable inventory."""
+
+import ast
+import builtins
+from collections import Counter
+from pathlib import Path
+
+import pytest
+
+from vllm_omni.config.environment_variable_inventory import (
+    ENVIRONMENT_VARIABLE_INVENTORY,
+    EnvironmentVariableCategory,
+    ModelEnvironmentVariableDisposition,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+_REPO_ROOT = Path(__file__).parents[2]
+_PACKAGE_ROOT = _REPO_ROOT / "vllm_omni"
+_INVENTORY_MODULE = _PACKAGE_ROOT / "config" / "environment_variable_inventory.py"
+_REFERENCE_PAGE = _REPO_ROOT / "docs" / "configuration" / "environment_variables.md"
+
+
+def _module_string_constants(tree: ast.Module) -> dict[str, str]:
+    constants: dict[str, str] = {}
+    for statement in tree.body:
+        targets: list[ast.expr] = []
+        value: ast.expr | None = None
+        if isinstance(statement, ast.Assign):
+            targets = statement.targets
+            value = statement.value
+        elif isinstance(statement, ast.AnnAssign):
+            targets = [statement.target]
+            value = statement.value
+        if isinstance(value, ast.Constant) and isinstance(value.value, str):
+            for target in targets:
+                if isinstance(target, ast.Name):
+                    constants[target.id] = value.value
+    return constants
+
+
+def _os_import_aliases(tree: ast.Module) -> tuple[set[str], set[str], set[str]]:
+    os_names: set[str] = set()
+    getenv_names: set[str] = set()
+    environ_names: set[str] = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            for imported in node.names:
+                if imported.name == "os":
+                    os_names.add(imported.asname or "os")
+        elif isinstance(node, ast.ImportFrom) and node.module == "os":
+            for imported in node.names:
+                if imported.name == "getenv":
+                    getenv_names.add(imported.asname or imported.name)
+                elif imported.name == "environ":
+                    environ_names.add(imported.asname or imported.name)
+    return os_names, getenv_names, environ_names
+
+
+def _is_environ(node: ast.expr, os_names: set[str], environ_names: set[str]) -> bool:
+    return (
+        isinstance(node, ast.Attribute)
+        and isinstance(node.value, ast.Name)
+        and node.value.id in os_names
+        and node.attr == "environ"
+    ) or (isinstance(node, ast.Name) and node.id in environ_names)
+
+
+def _environment_key_expressions(
+    node: ast.AST,
+    os_names: set[str],
+    getenv_names: set[str],
+    environ_names: set[str],
+) -> list[ast.expr]:
+    """Return key expressions used by one direct environment access."""
+    if isinstance(node, ast.Call) and node.args:
+        func = node.func
+        if (
+            isinstance(func, ast.Attribute)
+            and isinstance(func.value, ast.Name)
+            and func.value.id in os_names
+            and func.attr == "getenv"
+        ) or (isinstance(func, ast.Name) and func.id in getenv_names):
+            return [node.args[0]]
+        if (
+            isinstance(func, ast.Attribute)
+            and _is_environ(func.value, os_names, environ_names)
+            and func.attr in {"get", "pop", "setdefault"}
+        ):
+            return [node.args[0]]
+    elif isinstance(node, ast.Subscript) and _is_environ(node.value, os_names, environ_names):
+        return [node.slice]
+    elif isinstance(node, ast.Compare):
+        expressions: list[ast.expr] = []
+        left_values = [node.left, *node.comparators[:-1]]
+        for left, operator, right in zip(left_values, node.ops, node.comparators):
+            if isinstance(operator, (ast.In, ast.NotIn)) and _is_environ(right, os_names, environ_names):
+                expressions.append(left)
+        return expressions
+    return []
+
+
+def _resolve_environment_name(expression: ast.expr, constants: dict[str, str]) -> str | None:
+    value: str | None = None
+    if isinstance(expression, ast.Constant) and isinstance(expression.value, str):
+        value = expression.value
+    elif isinstance(expression, ast.Name):
+        value = constants.get(expression.id)
+    return value or None
+
+
+def _environment_wrapper_parameters(
+    tree: ast.Module,
+    os_names: set[str],
+    getenv_names: set[str],
+    environ_names: set[str],
+) -> dict[str, set[tuple[str, int | None]]]:
+    """Find local helpers that forward one parameter as an environment key."""
+    wrappers: dict[str, set[tuple[str, int | None]]] = {}
+    for function in (node for node in tree.body if isinstance(node, (ast.FunctionDef, ast.AsyncFunctionDef))):
+        positional = [*function.args.posonlyargs, *function.args.args]
+        parameter_indexes: dict[str, int | None] = {parameter.arg: index for index, parameter in enumerate(positional)}
+        parameter_indexes.update({parameter.arg: None for parameter in function.args.kwonlyargs})
+        forwarded: set[tuple[str, int | None]] = set()
+        for node in ast.walk(function):
+            for expression in _environment_key_expressions(node, os_names, getenv_names, environ_names):
+                if isinstance(expression, ast.Name) and expression.id in parameter_indexes:
+                    forwarded.add((expression.id, parameter_indexes[expression.id]))
+        if forwarded:
+            wrappers[function.name] = forwarded
+    return wrappers
+
+
+def _environment_accesses(path: Path) -> set[str]:
+    """Return statically resolvable names accessed through ``os``."""
+    tree = ast.parse(path.read_text(encoding="utf-8"), filename=str(path))
+    constants = _module_string_constants(tree)
+    os_names, getenv_names, environ_names = _os_import_aliases(tree)
+    wrappers = _environment_wrapper_parameters(tree, os_names, getenv_names, environ_names)
+    names: set[str] = set()
+
+    for node in ast.walk(tree):
+        for expression in _environment_key_expressions(node, os_names, getenv_names, environ_names):
+            if name := _resolve_environment_name(expression, constants):
+                names.add(name)
+
+        if isinstance(node, ast.Call) and isinstance(node.func, ast.Name):
+            for parameter_name, position in wrappers.get(node.func.id, set()):
+                argument = node.args[position] if position is not None and position < len(node.args) else None
+                if argument is None:
+                    argument = next(
+                        (keyword.value for keyword in node.keywords if keyword.arg == parameter_name),
+                        None,
+                    )
+                if argument is not None and (name := _resolve_environment_name(argument, constants)):
+                    names.add(name)
+
+    return names
+
+
+def test_inventory_matches_reviewed_snapshot_counts():
+    """Make an inventory expansion an explicit review decision."""
+    category_counts = Counter(item.category for item in ENVIRONMENT_VARIABLE_INVENTORY.values())
+    assert category_counts == {
+        EnvironmentVariableCategory.PUBLIC_OMNI: 23,
+        EnvironmentVariableCategory.INHERITED_VLLM: 20,
+        EnvironmentVariableCategory.PLATFORM_EXTERNAL: 27,
+        EnvironmentVariableCategory.MODEL_SPECIFIC: 56,
+        EnvironmentVariableCategory.BENCHMARK_TRANSITIONAL: 20,
+        EnvironmentVariableCategory.INTERNAL: 2,
+    }
+
+    disposition_counts = Counter(
+        item.model_disposition
+        for item in ENVIRONMENT_VARIABLE_INVENTORY.values()
+        if item.category is EnvironmentVariableCategory.MODEL_SPECIFIC
+    )
+    assert {disposition: disposition_counts[disposition] for disposition in ModelEnvironmentVariableDisposition} == {
+        ModelEnvironmentVariableDisposition.PROMOTE: 33,
+        ModelEnvironmentVariableDisposition.REQUEST_SCOPE: 6,
+        ModelEnvironmentVariableDisposition.EXTERNAL: 0,
+        ModelEnvironmentVariableDisposition.INTERNALIZE: 11,
+        ModelEnvironmentVariableDisposition.DEPRECATE_REMOVE: 6,
+    }
+
+
+def test_new_public_omni_names_use_project_prefix():
+    """Grandfather legacy names without allowing more prefix exceptions."""
+    legacy_public_names = {
+        "DIFFUSION_ATTENTION_BACKEND",
+        "DIFFUSION_CACHE_ADAPTER",
+        "DIFFUSION_CACHE_BACKEND",
+        "OMNI_DIFFUSION_PROMPT_EMBED_CACHE",
+        "OMNI_DIFFUSION_PROMPT_EMBED_CACHE_SIZE",
+        "OMNI_DIFFUSION_SESSION_STATE_MANAGER",
+        "OMNI_DIFFUSION_SESSION_STATE_MANAGER_MAX_SESSIONS",
+        "SPEAKER_MAX_UPLOADED",
+        "SPEAKER_SAMPLES_DIR",
+        "VLLM_VIDEO_ASYNC_CHUNK",
+        "VLLM_VIDEO_AUDIO_DELTA_MODE",
+    }
+    public_names = {
+        name
+        for name, item in ENVIRONMENT_VARIABLE_INVENTORY.items()
+        if item.category is EnvironmentVariableCategory.PUBLIC_OMNI
+    }
+
+    assert {name for name in public_names if not name.startswith("VLLM_OMNI_")} == legacy_public_names
+
+
+def test_statically_resolvable_environment_accesses_are_classified():
+    discovered: set[str] = set()
+    for path in _PACKAGE_ROOT.rglob("*.py"):
+        discovered.update(_environment_accesses(path))
+
+    assert discovered - ENVIRONMENT_VARIABLE_INVENTORY.keys() == set()
+
+
+def test_model_and_benchmark_inventory_entries_are_still_referenced():
+    """Reject transitional rows after their implementation has disappeared."""
+    package_source = "\n".join(
+        path.read_text(encoding="utf-8") for path in _PACKAGE_ROOT.rglob("*.py") if path != _INVENTORY_MODULE
+    )
+    transitional_categories = {
+        EnvironmentVariableCategory.MODEL_SPECIFIC,
+        EnvironmentVariableCategory.BENCHMARK_TRANSITIONAL,
+    }
+
+    stale_names = {
+        name
+        for name, item in ENVIRONMENT_VARIABLE_INVENTORY.items()
+        if item.category in transitional_categories and name not in package_source
+    }
+
+    assert stale_names == set()
+
+
+def test_environment_scanner_covers_indirection_aliases_membership_and_casing():
+    expected_by_path = {
+        "metrics/definitions.py": {
+            "VLLM_OMNI_BENCH_AUDIO_CHANNELS",
+            "VLLM_OMNI_BENCH_AUDIO_SAMPLE_RATE",
+        },
+        "entrypoints/openai/video_stream_envs.py": {
+            "VLLM_VIDEO_ASYNC_CHUNK",
+            "VLLM_VIDEO_AUDIO_DELTA_MODE",
+        },
+        "model_executor/models/mimo_audio/mimo_audio.py": {"model_stage"},
+        "model_executor/models/moss_tts/modeling_moss_tts_local_depth.py": {"MOSS_TTS_DEBUG_STOP"},
+        "distributed/ray_utils/utils.py": {"RAY_RAYLET_PID"},
+    }
+    for relative_path, expected in expected_by_path.items():
+        assert expected <= _environment_accesses(_PACKAGE_ROOT / relative_path)
+
+
+def test_generated_server_storage_environment_names_are_classified():
+    from vllm_omni.config.server_settings import FileBackend
+
+    generated_names = {
+        f"VLLM_OMNI_SERVER_STORAGE__{field_name.upper()}"
+        for field_name in FileBackend.model_fields
+        if field_name != "type"
+    }
+    assert generated_names <= ENVIRONMENT_VARIABLE_INVENTORY.keys()
+
+
+def test_public_omni_variables_are_in_the_reference_page():
+    reference = _REFERENCE_PAGE.read_text(encoding="utf-8")
+    missing = {
+        name
+        for name, item in ENVIRONMENT_VARIABLE_INVENTORY.items()
+        if item.is_public_omni and f"`{name}`" not in reference
+    }
+    assert missing == set()
+
+
+def test_secret_values_are_marked_for_redaction():
+    assert {name for name, item in ENVIRONMENT_VARIABLE_INVENTORY.items() if item.redact_value} == {
+        "HF_TOKEN",
+        "HUGGINGFACE_HUB_TOKEN",
+        "OPENAI_API_KEY",
+    }
+
+
+def test_collect_env_reports_only_safe_public_omni_values(monkeypatch):
+    from collect_env import get_env_vars
+
+    monkeypatch.setenv("DIFFUSION_CACHE_BACKEND", "tea_cache")
+    monkeypatch.setenv("HF_TOKEN", "hf-secret")
+    monkeypatch.setenv("OPENAI_API_KEY", "openai-secret")
+    monkeypatch.setenv("VLLM_OMNI_REPLICA_ID", "internal-replica")
+
+    report = get_env_vars()
+
+    assert "DIFFUSION_CACHE_BACKEND=tea_cache" in report.splitlines()
+    assert "hf-secret" not in report
+    assert "openai-secret" not in report
+    assert "VLLM_OMNI_REPLICA_ID" not in report
+
+
+@pytest.mark.parametrize("error_type", [ImportError, OSError, AssertionError, RuntimeError])
+def test_collect_env_survives_inventory_import_error(monkeypatch, error_type):
+    from collect_env import get_env_vars
+
+    real_import = builtins.__import__
+
+    def fail_inventory_import(name, *args, **kwargs):
+        if name == "vllm_omni.config.environment_variable_inventory":
+            raise error_type("broken vllm-omni installation")
+        return real_import(name, *args, **kwargs)
+
+    monkeypatch.setattr(builtins, "__import__", fail_inventory_import)
+    monkeypatch.setenv("CUDA_VISIBLE_DEVICES", "0")
+
+    assert "CUDA_VISIBLE_DEVICES=0" in get_env_vars().splitlines()

tests/config/test_omni_config.py

@@ -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
 """Tests for the additive structured Omni config."""
 
 from __future__ import annotations
@@ -244,6 +244,36 @@ def test_from_pipeline_config_rejects_unowned_deploy_engine_extras(engine_extras
         VllmOmniConfig.from_pipeline_config(pipeline, user_deploy_config=deploy)
 
 
+@pytest.mark.parametrize(
+    ("engine_extras", "cli_overrides"),
+    [
+        ({"requires_full_payload_input": False}, {}),
+        ({}, {"stage_1_requires_full_payload_input": False}),
+    ],
+    ids=["deploy", "stage-cli"],
+)
+def test_from_pipeline_config_rejects_full_payload_input_capability_overrides(
+    engine_extras,
+    cli_overrides,
+):
+    pipeline = _resolve_pipeline_or_skip("qwen3_tts")
+    deploy = DeployConfig(
+        stages=[
+            StageDeployConfig(
+                stage_id=1,
+                engine_extras=engine_extras,
+            )
+        ]
+    )
+
+    with pytest.raises(ValueError, match=r"no structured config owner: requires_full_payload_input"):
+        VllmOmniConfig.from_pipeline_config(
+            pipeline,
+            user_deploy_config=deploy,
+            cli_overrides=cli_overrides,
+        )
+
+
 @pytest.mark.parametrize(
     ("cli_overrides", "stage_id"),
     [
@@ -443,10 +473,12 @@ def test_from_pipeline_config_dispatches_async_chunk_processors_without_mutating
     async_config = _from_pipeline_key("qwen3_tts")
     assert async_config.stage_by_id(0).custom_process_next_stage_input_func.endswith("talker2code2wav_async_chunk")
     assert async_config.stage_by_id(1).custom_process_input_func is None
+    assert async_config.stage_by_id(1).model_config.requires_full_payload_input is True
 
     sync_config = _from_pipeline_key("qwen3_tts", cli_overrides={"async_chunk": False})
     assert sync_config.stage_by_id(0).custom_process_next_stage_input_func.endswith("talker2code2wav_full_payload")
     assert sync_config.stage_by_id(1).custom_process_input_func.endswith("talker2code2wav_token_only")
+    assert sync_config.stage_by_id(1).model_config.requires_full_payload_input is True
 
     assert pipeline.get_stage(0).custom_process_next_stage_input_func.endswith("talker2code2wav_full_payload")
     assert pipeline.get_stage(1).custom_process_input_func is None
@@ -567,6 +599,7 @@ def test_sub_config_fields_match_structured_scopes():
         # legacy engine-args path.
         "model_subdir",
         "tokenizer_subdir",
+        "requires_full_payload_input",
     }
     vllm_load_fields = {f.name for f in fields(VllmLoadConfig)}
     assert issubclass(OmniStageLoadConfig, VllmLoadConfig)
@@ -1152,6 +1185,35 @@ def test_from_pipeline_config_normalizes_diffusion_config_aliases_from_engine_ar
     assert stage.diffusion_config.diffusion_kv_max_rows_per_request == 2
 
 
+def test_from_pipeline_config_forwards_fastvideo_vsa_topk(tmp_path, monkeypatch):
+    from vllm_omni.platforms import current_omni_platform
+
+    monkeypatch.setattr(current_omni_platform, "is_cuda", lambda: True)
+    deploy_path = tmp_path / "dreamzero_fastvideo_vsa.yaml"
+    deploy_path.write_text(
+        "\n".join(
+            [
+                "pipeline: dreamzero",
+                "async_chunk: false",
+                "stages:",
+                "  - stage_id: 0",
+                "    diffusion_attention_backend: FASTVIDEO_VSA",
+                "    fastvideo_vsa_topk: 96",
+            ]
+        )
+    )
+
+    stage = _from_pipeline_key(
+        "dreamzero",
+        deploy_config_path=str(deploy_path),
+    ).stage_by_id(0)
+
+    attention_config = stage.diffusion_config.diffusion_attention_config
+    assert attention_config.default is not None
+    assert attention_config.default.backend == "FASTVIDEO_VSA"
+    assert attention_config.default.fastvideo_vsa_topk == 96
+
+
 def test_from_pipeline_config_rejects_reserved_diffusion_kv_mode(tmp_path):
     deploy_path = tmp_path / "dreamzero_reserved_diffusion_kv.yaml"
     deploy_path.write_text(

tests/conftest.py

@@ -1,7 +1,10 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 Root pytest entrypoint for the vLLM-Omni test suite.
 
-- `tests/conftest.py` stays thin: plugin registration + compatibility re-exports.
+- `tests/conftest.py` stays thin: plugin registration only.
 - Importable utilities live under `tests/helpers/`.
 - Helper unit tests live under `tests/helpers/tests/`.
 - Fixtures live under `tests/helpers/fixtures/` and are loaded via `pytest_plugins`.
@@ -19,12 +22,11 @@ bootstrap_vllm_layer_custom_op_modules()
 
 pytest_plugins = (
     "tests.helpers.fixtures.config",
-    "tests.helpers.fixtures.env",
+    "tests.helpers.fixtures.clean",
     "tests.helpers.fixtures.log",
     "tests.helpers.fixtures.media",
     "tests.helpers.fixtures.run_args",
     "tests.helpers.fixtures.runtime",
-    "tests.helpers.fixtures.speaker_cache",
     "tests.helpers.fixtures.xdist",
 )
 
@@ -32,46 +34,3 @@ pytest_plugins = (
 def pytest_terminal_summary(terminalreporter, exitstatus, config):
     # Marker for Buildkite log folding before pytest summary lines.
     terminalreporter.write_line("--- Running Summary")
-
-
-# Backward-compatible lazy re-exports.
-# (Many tests still import from `tests.conftest`; migrate these imports to `tests.helpers.*` over time.)
-# Keep these lazy so conftest import does not trigger heavy helper dependencies.
-_ASSERTION_EXPORT_NAMES = (
-    "assert_audio_speech_response",
-    "assert_diffusion_response",
-    "assert_http_error",
-    "assert_image_diffusion_response",
-    "assert_image_valid",
-    "assert_omni_response",
-    "assert_video_diffusion_response",
-    "assert_video_valid",
-)
-_MEDIA_EXPORT_NAMES = (
-    "convert_audio_bytes_to_text",
-    "convert_audio_file_to_text",
-    "cosine_similarity_text",
-    "decode_b64_image",
-    "generate_synthetic_audio",
-    "generate_synthetic_image",
-    "generate_synthetic_video",
-)
-_STAGE_CONFIG_EXPORT_NAMES = ("modify_stage_config",)
-_RUNTIME_EXPORT_NAMES = (
-    "DiffusionResponse",
-    "HttpResponse",
-    "OmniResponse",
-    "OmniRunner",
-    "OmniRunnerHandler",
-    "OmniServer",
-    "OmniServerParams",
-    "OmniServerStageCli",
-    "OpenAIClientHandler",
-    "dummy_messages_from_mix_data",
-)
-_LAZY_EXPORT_MODULES = {
-    **{name: "tests.helpers.assertions" for name in _ASSERTION_EXPORT_NAMES},
-    **{name: "tests.helpers.media" for name in _MEDIA_EXPORT_NAMES},
-    **{name: "tests.helpers.stage_config" for name in _STAGE_CONFIG_EXPORT_NAMES},
-    **{name: "tests.helpers.runtime" for name in _RUNTIME_EXPORT_NAMES},
-}

tests/core/sched/test_omni_ar_scheduler_streaming.py

@@ -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 Omni AR streaming-session async placeholder handling."""
 
@@ -64,6 +64,9 @@ def _run_resumable_segment_stop(
     session: Request,
     *,
     session_finished: bool = False,
+    handle_stopped=None,
+    chunk_transfer_adapter=None,
+    inter_stage_output=None,
 ):
     sched = MagicMock()
     sched.requests = {session.request_id: session}
@@ -75,11 +78,14 @@ def _run_resumable_segment_stop(
         return [42], True
 
     sched._update_request_with_output.side_effect = stop_request
-    sched._handle_stopped_request.return_value = session_finished
     # vLLM 0.26 returns (kv_xfer_params, ec_xfer_params); an unconfigured
     # MagicMock iterates empty and fails to unpack at the call site.
     sched._free_request.return_value = (None, None)
-    sched.chunk_transfer_adapter = None
+    if handle_stopped is None:
+        sched._handle_stopped_request.return_value = session_finished
+    else:
+        sched._handle_stopped_request.side_effect = handle_stopped
+    sched.chunk_transfer_adapter = chunk_transfer_adapter
     sched.running = [session]
     sched.waiting_for_transfer_free = set()
     sched.transfer_triggered_requests = set()
@@ -105,6 +111,7 @@ def _run_resumable_segment_stop(
     model_runner_output.cudagraph_stats = None
     model_runner_output.req_id_to_index = {session.request_id: 0}
     model_runner_output.routed_experts = None
+    model_runner_output.inter_stage_outputs = [inter_stage_output] if inter_stage_output is not None else None
 
     return OmniARScheduler.update_from_output(sched, scheduler_output, model_runner_output)
 
@@ -160,6 +167,41 @@ def test_update_from_output_settles_in_flight_tokens() -> None:
     assert session.num_in_flight_tokens == 0
 
 
+def test_resumable_segment_boundary_keeps_pre_transition_send_watermark() -> None:
+    """The old segment's transfer must not observe the next segment's reset.
+
+    A queued streaming update can replace the same mutable Request while the
+    stop output is still being handled. The connector needs the confirmed
+    token count from before that replacement to avoid dropping the boundary.
+    """
+    session = _make_request()
+    session.status = RequestStatus.RUNNING
+    session.resumable = True
+    session.num_computed_tokens = 26
+    adapter = MagicMock()
+    adapter._confirmed_num_computed_tokens.return_value = 26
+    inter_stage_output = {"codes": {"audio": [7]}}
+
+    def replace_with_next_segment(request: Request) -> bool:
+        request.num_computed_tokens = 0
+        request.status = RequestStatus.WAITING
+        return False
+
+    _run_resumable_segment_stop(
+        session,
+        handle_stopped=replace_with_next_segment,
+        chunk_transfer_adapter=adapter,
+        inter_stage_output=inter_stage_output,
+    )
+
+    adapter.save_async.assert_called_once_with(
+        inter_stage_output,
+        session,
+        True,
+        confirmed_num_computed_tokens=26,
+    )
+
+
 def test_running_decode_step_without_inter_stage_payload_does_not_raise() -> None:
     """A decode step that neither stops nor carries an inter-stage payload.
 
@@ -210,6 +252,37 @@ def test_running_decode_step_without_inter_stage_payload_does_not_raise() -> Non
     sched.chunk_transfer_adapter.save_async.assert_not_called()
 
 
+def test_queued_streaming_update_on_async_stop_fences_in_flight_once() -> None:
+    """Native duplex + async: the next unit is usually already queued when the
+    current one stops, so _handle_stopped_request applies the update and the
+    stop site below it fences the same in-flight decode a second time. An
+    accumulated residue outlives the drain and swallows the next unit's
+    listen/speak, which the client sees as silence until its timeout.
+    """
+    session = _make_request()
+    session.status = RequestStatus.RUNNING
+    session.resumable = True
+    session.append_output_token_ids([7])
+    session.num_computed_tokens = 4
+    session.num_output_placeholders = 1
+    # This step's frame (settled to 0 by update_from_output) plus one extra
+    # async decode still in flight.
+    session.num_in_flight_tokens = 2
+
+    sched = _make_scheduler(stage_id=0)
+    sched._enqueue_waiting_request = MagicMock()
+
+    def handle_stopped(request: Request) -> bool:
+        sched._update_request_as_session(request, _make_update([10, 20]))
+        return False
+
+    _run_resumable_segment_stop(session, handle_stopped=handle_stopped)
+
+    # Exactly the one unreported decode, so the drain reaches zero before the
+    # new segment's first frame arrives.
+    assert session.num_stale_output_tokens == session.num_in_flight_tokens == 1
+
+
 def test_stage0_streaming_update_discards_outstanding_async_placeholder_token() -> None:
     sched = _make_scheduler(stage_id=0)
     session = _make_request()

tests/core/sched/test_omni_scheduler_mixin_shared.py

@@ -4,10 +4,8 @@ from types import SimpleNamespace
 import pytest
 from vllm.v1.engine import FinishReason
 
-from vllm_omni.core.sched.omni_scheduler_mixin import (
-    DEFAULT_INPUT_WAIT_TIMEOUT_S,
-    OmniSchedulerMixin,
-)
+from vllm_omni.core.sched import omni_scheduler_mixin
+from vllm_omni.core.sched.omni_scheduler_mixin import OmniSchedulerMixin
 from vllm_omni.core.sched.output import OmniChunkRecvHandle
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -17,6 +15,77 @@ class _Scheduler(OmniSchedulerMixin):
     pass
 
 
+def test_async_chunk_adapter_initializes_for_stage_zero_sender_and_stage_one_receiver(monkeypatch):
+    created_adapters = []
+
+    def adapter_factory(config):
+        adapter = SimpleNamespace(vllm_config=config)
+        created_adapters.append(adapter)
+        return adapter
+
+    monkeypatch.setattr(omni_scheduler_mixin, "OmniChunkTransferAdapter", adapter_factory)
+
+    def init_scheduler(**model_config):
+        scheduler = _Scheduler()
+        scheduler.vllm_config = SimpleNamespace(model_config=SimpleNamespace(**model_config))
+        scheduler._init_omni_io_scheduling_state()
+        return scheduler
+
+    producer = init_scheduler(
+        stage_id=0,
+        async_chunk=True,
+        requires_full_payload_input=False,
+        custom_process_next_stage_input_func="test.pipeline.produce_async_chunk",
+    )
+    receiver = init_scheduler(
+        stage_id=1,
+        async_chunk=True,
+        requires_full_payload_input=True,
+        custom_process_next_stage_input_func=None,
+    )
+    terminal_stage = init_scheduler(
+        stage_id=2,
+        async_chunk=True,
+        requires_full_payload_input=False,
+        custom_process_next_stage_input_func=None,
+    )
+
+    assert producer.chunk_transfer_adapter is not None
+    assert receiver.chunk_transfer_adapter is not None
+    assert terminal_stage.chunk_transfer_adapter is not None
+    assert receiver.input_coordinator is None
+    assert len(created_adapters) == 3
+
+
+@pytest.mark.parametrize(
+    ("stage_id", "async_chunk", "required", "enabled"),
+    [
+        (0, False, True, False),
+        (1, True, True, False),
+        (1, False, False, False),
+        (1, False, True, True),
+    ],
+)
+def test_full_payload_coordinator_matches_legacy_gate(monkeypatch, stage_id, async_chunk, required, enabled):
+    monkeypatch.setattr(
+        omni_scheduler_mixin,
+        "OmniChunkTransferAdapter",
+        lambda _config: SimpleNamespace(),
+    )
+    scheduler = _Scheduler()
+    scheduler.vllm_config = SimpleNamespace(
+        model_config=SimpleNamespace(
+            stage_id=stage_id,
+            async_chunk=async_chunk,
+            requires_full_payload_input=required,
+        )
+    )
+
+    scheduler._init_omni_io_scheduling_state()
+
+    assert (scheduler.input_coordinator is not None) is enabled
+
+
 def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
     calls = []
     scheduler = _Scheduler()
@@ -58,7 +127,7 @@ def test_schedule_lifecycle_helpers_process_and_restore_both_input_paths():
         ("process", scheduler.waiting, scheduler.running, scheduler.requests),
         # The chunk deadline runs after chunks are applied, so a chunk that
         # arrived this cycle resets the clock before it is measured (R1.1).
-        ("chunk-timeouts", DEFAULT_INPUT_WAIT_TIMEOUT_S),
+        ("chunk-timeouts", omni_scheduler_mixin.DEFAULT_INPUT_WAIT_TIMEOUT_S),
         ("failed-sends",),
         ("restore-chunks", scheduler.waiting, scheduler.running, scheduler.requests),
         ("restore-full", scheduler.waiting),

tests/core/sched/test_omni_scheduling_coordinator.py

@@ -20,7 +20,6 @@ import torch
 import vllm_omni.core.sched.omni_scheduling_coordinator as coord_mod
 from vllm_omni.core.sched.omni_scheduling_coordinator import (
     OmniSchedulingCoordinator,
-    uses_full_payload_input_coordinator,
 )
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -96,109 +95,6 @@ class MockQueue:
 # ------------------------------------------------------------------ #
 
 
-class TestFullPayloadCoordinatorSelection(unittest.TestCase):
-    """Tests for the (model_arch, model_stage) whitelist gate.
-
-    The init_omni_connectors arch allowlist is keyed by ``model_arch`` and
-    is a superset of the stages registered here -- consumer-wait stages
-    must be registered explicitly in ``_FULL_PAYLOAD_INPUT_STAGES``, while
-    the init allowlist covers both producer- and consumer-side runners.
-    These tests pin which ``(arch, stage)`` pairs the gate fires for today.
-    """
-
-    # Expected whitelist (model_arch, model_stage).  Hardcoded to avoid the
-    # tautology of importing _FULL_PAYLOAD_INPUT_STAGES and asserting it
-    # against itself; any drift between this matrix and the whitelist will
-    # fail loudly here.
-    EXPECTED_FULL_PAYLOAD_INPUT_STAGES: frozenset[tuple[str, str]] = frozenset(
-        {
-            ("Qwen3OmniMoeForConditionalGeneration", "talker"),
-            ("Qwen3OmniMoeForConditionalGeneration", "code2wav"),
-            ("Qwen2_5OmniForConditionalGeneration", "talker"),
-            ("Qwen2_5OmniForConditionalGeneration", "code2wav"),
-            ("CovoAudioForConditionalGeneration", "code2wav"),
-            ("MiMoAudioModel", "code2wav"),
-            ("Qwen3TTSCode2Wav", "code2wav"),
-            ("MiniCPMO45Code2Wav", "code2wav"),
-            # nemotron_voicechat: sync-path RVQ-VAE decode waits for the
-            # talker's full [frames, 31] code-stack payload.
-            ("NemotronVoiceChatCode2Wav", "code2wav"),
-            ("CosyVoice3Model", "cosyvoice3_code2wav"),
-            # audex: sync-path TTS decoder (async_chunk: false deploys) and
-            # the TTA XCodec1 stage (always sync full-payload).
-            ("AudexCode2Wav", "audex_code2wav"),
-            ("AudexXCodec1", "audex_xcodec"),
-            ("IndexTTS2S2MelDecoder", "indextts2_s2mel_decoder"),
-            ("IndexTTS25S2MelDecoder", "indextts2_5_s2mel_decoder"),
-            ("DyninOmniForConditionalGeneration", "token2image"),
-            ("DyninOmniForConditionalGeneration", "token2audio"),
-        }
-    )
-
-    def test_whitelist_matches_expected_matrix(self):
-        """_FULL_PAYLOAD_INPUT_STAGES must equal the hardcoded expected matrix.
-
-        Catches both accidental additions (which would silently enable the
-        consumer-wait gate for a new arch) and accidental removals (which
-        would silently disable an enabled arch).
-        """
-        from vllm_omni.core.sched.omni_scheduling_coordinator import _FULL_PAYLOAD_INPUT_STAGES
-
-        self.assertEqual(
-            frozenset(_FULL_PAYLOAD_INPUT_STAGES),
-            self.EXPECTED_FULL_PAYLOAD_INPUT_STAGES,
-            msg="_FULL_PAYLOAD_INPUT_STAGES drifted from the expected matrix; "
-            "update EXPECTED_FULL_PAYLOAD_INPUT_STAGES if intentional.",
-        )
-
-    def test_all_whitelisted_arch_stage_pairs_fire_gate(self):
-        """Every (arch, stage) pair in the expected matrix must fire
-        the gate when stage_id > 0 and async_chunk=False.
-        """
-        for arch, stage in self.EXPECTED_FULL_PAYLOAD_INPUT_STAGES:
-            model_config = SimpleNamespace(
-                stage_id=1,
-                async_chunk=False,
-                model_arch=arch,
-                model_stage=stage,
-            )
-            self.assertTrue(
-                uses_full_payload_input_coordinator(model_config),
-                msg=f"expected gate to fire for {arch}/{stage}",
-            )
-
-    def test_other_arch_or_stage_or_mode_does_not_fire(self):
-        cases = [
-            SimpleNamespace(
-                stage_id=1, async_chunk=True, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage="talker"
-            ),
-            SimpleNamespace(
-                stage_id=0, async_chunk=False, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage="thinker"
-            ),
-            SimpleNamespace(
-                stage_id=1,
-                async_chunk=False,
-                model_arch="Qwen3OmniMoeForConditionalGeneration",
-                model_stage="some_future_stage",
-            ),
-            SimpleNamespace(
-                stage_id=1, async_chunk=False, model_arch="Qwen3TTSForConditionalGeneration", model_stage="code2wav"
-            ),
-            SimpleNamespace(
-                stage_id=1, async_chunk=False, model_arch="MingFlashOmniForConditionalGeneration", model_stage="talker"
-            ),
-            SimpleNamespace(stage_id=1, async_chunk=False, model_arch=None, model_stage="talker"),
-            SimpleNamespace(
-                stage_id=1, async_chunk=False, model_arch="Qwen3OmniMoeForConditionalGeneration", model_stage=None
-            ),
-        ]
-        for model_config in cases:
-            self.assertFalse(
-                uses_full_payload_input_coordinator(model_config),
-                msg=f"expected gate OFF for {model_config}",
-            )
-
-
 class TestCoordinatorUpdateRequestMetadata(unittest.TestCase):
     """Test update_request_metadata applies scheduling metadata to requests."""
 

tests/dfx/conftest.py

@@ -105,7 +105,7 @@ def create_unique_server_pytest_params(
     configs: list[dict[str, Any]],
     stage_configs_dir: Path,
 ) -> list[Any]:
-    """Like :func:`create_unique_server_params`, but wrap each row in ``pytest.param`` with JSON marks."""
+    """Like :func:`_create_unique_server_params`, but wrap each row in ``pytest.param`` with JSON marks."""
     marks_by_name = _marks_by_test_name(configs)
     return [
         pytest.param(
@@ -113,7 +113,7 @@ def create_unique_server_pytest_params(
             marks=marks_by_name.get(row[0], []),
             id=row[0],
         )
-        for row in create_unique_server_params(configs, stage_configs_dir)
+        for row in _create_unique_server_params(configs, stage_configs_dir)
     ]
 
 
@@ -150,7 +150,7 @@ def create_paired_omni_benchmark_pytest_params(
     """Paired params for ``run_benchmark.py`` (omni/tts)."""
     mapping = create_test_parameter_mapping(configs)
     marks_by_name = _marks_by_test_name(configs)
-    server_entries = [(row, row[0]) for row in create_unique_server_params(configs, stage_configs_dir)]
+    server_entries = [(row, row[0]) for row in _create_unique_server_params(configs, stage_configs_dir)]
     params_by_test_name = {
         test_name: get_benchmark_params_for_server(test_name, mapping) for _, test_name in server_entries
     }
@@ -232,7 +232,7 @@ def _build_serve_args(serve_args: Any) -> list[str]:
     return args
 
 
-def create_unique_server_params(
+def _create_unique_server_params(
     configs: list[dict[str, Any]],
     stage_configs_dir: Path,
 ) -> list[tuple[str, str, str | None, str | None, tuple[str, ...], bool]]:
@@ -311,7 +311,7 @@ def create_reliability_omni_server_params(
     configs: list[dict[str, Any]], stage_configs_dir: Path
 ) -> list[OmniServerParams]:
     adjusted_configs = configs_with_platform_stage_configs(configs)
-    unique_params = create_unique_server_params(adjusted_configs, stage_configs_dir)
+    unique_params = _create_unique_server_params(adjusted_configs, stage_configs_dir)
     server_args_by_name = extract_server_args_by_test_name(adjusted_configs)
     return [
         OmniServerParams(

tests/dfx/perf/tests/test_minicpmo_4_5_duplex_seed_tts.json

@@ -18,7 +18,7 @@
       "model": "openbmb/MiniCPM-o-4_5",
       "extra_cli_args": [
         "--deploy-config",
-        "vllm_omni/deploy/minicpmo_4_5_duplex.yaml",
+        "vllm_omni/deploy/minicpmo_4_5.yaml",
         "--trust-remote-code"
       ]
     },

tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json

@@ -21,47 +21,6 @@
       ]
     },
     "benchmark_params": [
-      {
-        "dataset_name": "random",
-        "backend": "openai-chat-omni",
-        "endpoint": "/v1/chat/completions",
-        "num_prompts": [
-          4,
-          16,
-          32,
-          64,
-          128
-        ],
-        "max_concurrency": [
-          1,
-          4,
-          8,
-          16,
-          32
-        ],
-        "random_input_len": 2500,
-        "random_output_len": 900,
-        "ignore_eos": true,
-        "percentile-metrics": "ttft,tpot,itl,e2el,audio_rtf,audio_ttfp,audio_duration",
-        "baseline": {
-          "H100": {
-            "mean_e2el_ms": [
-              11701.3818,
-              27510.3956,
-              31907.9613,
-              42173.9832,
-              63970.0751
-            ],
-            "mean_audio_rtf": [
-              0.1175,
-              0.1647,
-              0.2003,
-              0.2726,
-              0.4195
-            ]
-          }
-        }
-      },
       {
         "dataset_name": "random-mm",
         "backend": "openai-chat-omni",
@@ -170,45 +129,6 @@
           }
         }
       },
-      {
-        "dataset_name": "random",
-        "backend": "openai-chat-omni",
-        "endpoint": "/v1/chat/completions",
-        "num_prompts": [
-          4,
-          16,
-          32,
-          64,
-          128
-        ],
-        "max_concurrency": [
-          1,
-          4,
-          8,
-          16,
-          32
-        ],
-        "random_input_len": 2500,
-        "random_output_len": 900,
-        "ignore_eos": true,
-        "extra_body": {
-          "modalities": [
-            "text"
-          ]
-        },
-        "percentile-metrics": "ttft,tpot,itl,e2el",
-        "baseline": {
-          "H100": {
-            "mean_e2el_ms": [
-              4354.8945,
-              7672.4248,
-              9861.1687,
-              13086.8865,
-              18716.709
-            ]
-          }
-        }
-      },
       {
         "dataset_name": "random-mm",
         "backend": "openai-chat-omni",
@@ -324,5 +244,61 @@
         }
       }
     ]
+  },
+  {
+    "test_name": "test_qwen3_omni_chunk_cuda",
+    "mark": [
+      {
+        "hardware_marks": {
+          "res": {
+            "cuda": "H100"
+          },
+          "num_cards": 2
+        }
+      },
+      "omni",
+      "slow"
+    ],
+    "server_params": {
+      "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
+      "extra_cli_args": [
+        "--async-chunk"
+      ]
+    },
+    "benchmark_params": [
+      {
+        "dataset_name": "random",
+        "backend": "openai-chat-omni",
+        "endpoint": "/v1/chat/completions",
+        "num_prompts": [
+          4,
+          32,
+          128
+        ],
+        "max_concurrency": [
+          1,
+          8,
+          32
+        ],
+        "random_input_len": 100,
+        "random_output_len": 100,
+        "ignore_eos": true,
+        "extra_body": {
+          "modalities": [
+            "text"
+          ]
+        },
+        "percentile-metrics": "ttft,tpot,itl,e2el",
+        "baseline": {
+          "H100": {
+            "mean_e2el_ms": [
+              4354.8945,
+              9861.1687,
+              18716.709
+            ]
+          }
+        }
+      }
+    ]
   }
 ]

tests/dfx/perf/tests/test_qwen3_omni_vllm_text.json

@@ -29,16 +29,12 @@
         "endpoint": "/v1/completions",
         "num_prompts": [
           4,
-          16,
           32,
-          64,
           128
         ],
         "max_concurrency": [
           1,
-          4,
           8,
-          16,
           32
         ],
         "random_input_len": 2500,
@@ -54,9 +50,7 @@
           "H100": {
             "mean_e2el_ms": [
               4619.0068,
-              7868.3523,
               9760.6175,
-              13422.3491,
               18007.6918
             ]
           }
@@ -99,45 +93,6 @@
           }
         }
       },
-      {
-        "dataset_name": "random-mm",
-        "backend": "vllm",
-        "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
-        "endpoint": "/v1/completions",
-        "num_prompts": [
-          40
-        ],
-        "request_rate": [
-          0.5
-        ],
-        "random_input_len": 100,
-        "random_output_len": 100,
-        "random_range_ratio": 0.0,
-        "ignore_eos": true,
-        "extra_body": {
-          "modalities": [
-            "text"
-          ]
-        },
-        "random_mm_base_items_per_request": 2,
-        "random_mm_num_mm_items_range_ratio": 0.5,
-        "random_mm_limit_mm_per_prompt": {
-          "image": 1,
-          "video": 1
-        },
-        "random_mm_bucket_config": {
-          "(256, 256, 1)": 0.5,
-          "(720, 1280, 2)": 0.5
-        },
-        "percentile-metrics": "ttft,tpot,itl,e2el",
-        "baseline": {
-          "H100": {
-            "mean_e2el_ms": [
-              524.5321
-            ]
-          }
-        }
-      },
       {
         "dataset_name": "random-mm",
         "backend": "vllm",

tests/dfx/perf/tests/test_qwen_image_vllm_omni.json

@@ -74,13 +74,14 @@
       "full_model",
       "diffusion"
     ],
-    "description": "Single-device baseline (no parallelism) with step execution",
+    "description": "Single-device step execution: sequential 512/1536 plus 512 concurrency sweep. One server (max-num-seqs 8).",
     "server_type": "vllm-omni",
     "server_params": {
       "model": "Qwen/Qwen-Image",
       "serve_args": {
         "enable-diffusion-pipeline-profiler": true,
-        "step-execution": true
+        "step-execution": true,
+        "max-num-seqs": 8
       }
     },
     "benchmark_params": [
@@ -119,6 +120,27 @@
             "peak_memory_mb_mean": 64500.0
           }
         }
+      },
+      {
+        "name": "512x512_steps20_high_concurrency",
+        "dataset": "random",
+        "task": "t2i",
+        "width": 512,
+        "height": 512,
+        "num-inference-steps": 20,
+        "num-prompts": [20, 40, 80, 160],
+        "max-concurrency": [1, 2, 4, 8],
+        "warmup-requests": 8,
+        "warmup-concurrency": 8,
+        "warmup-num-inference-steps": 20,
+        "enable-negative-prompt": true,
+        "baseline": {
+          "H100": {
+            "throughput_qps": [0.3937, 0.6606, 0.6188, 0.7752],
+            "latency_mean": [2.5466, 3.0235, 6.3948, 10.1944],
+            "peak_memory_mb_mean": [55149.0, 55150.0, 55153.1458, 55156.325]
+          }
+        }
       }
     ]
   },
@@ -243,53 +265,5 @@
         }
       }
     ]
-  },
-  {
-    "test_name": "test_qwen_image_single_device_step_execution_high_concurrency",
-    "mark": [
-      {
-        "hardware_marks": {
-          "res": {
-            "cuda": "H100"
-          },
-          "num_cards": 1
-        }
-      },
-      "full_model",
-      "diffusion"
-    ],
-    "description": "Single-device Qwen-Image step-level execution benchmark at 512x512, 20 steps, concurrency sweep",
-    "server_type": "vllm-omni",
-    "server_params": {
-      "model": "Qwen/Qwen-Image",
-      "serve_args": {
-        "enable-diffusion-pipeline-profiler": true,
-        "step-execution": true,
-        "max-num-seqs": 8
-      }
-    },
-    "benchmark_params": [
-      {
-        "name": "512x512_steps20_high_concurrency",
-        "dataset": "random",
-        "task": "t2i",
-        "width": 512,
-        "height": 512,
-        "num-inference-steps": 20,
-        "num-prompts": [20, 40, 80, 160],
-        "max-concurrency": [1, 2, 4, 8],
-        "warmup-requests": 8,
-        "warmup-concurrency": 8,
-        "warmup-num-inference-steps": 20,
-        "enable-negative-prompt": true,
-        "baseline": {
-            "H100": {
-                "throughput_qps": [0.3937, 0.6606, 0.6188, 0.7752],
-                "latency_mean": [2.5466, 3.0235, 6.3948, 10.1944],
-                "peak_memory_mb_mean": [55149.0, 55150.0, 55153.1458, 55156.325]
-            }
-        }
-      }
-    ]
   }
 ]

tests/dfx/reliability/invalid_param_test/test_invalid_audio_diffusion.py

@@ -1,13 +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
 """``POST /v1/audio/generate`` validation (live Stable Audio Open server)."""
 
 from __future__ import annotations
 
+from typing import Any
+
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
 pytestmark = [pytest.mark.slow, pytest.mark.diffusion]
 
@@ -95,20 +97,20 @@ _PARAMS = [
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_audio_generate_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     overrides: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
-    body = {"model": omni_server.model, "input": "ambient electronic pad"}
+    body: dict[str, Any] = {"model": omni_server.model, "input": "ambient electronic pad"}
     body.update(overrides)
-    openai_client.send_audio_generate_http_request(
+    online_client.send_audio_generate_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
 
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
-def test_audio_generate_missing_input(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_audio_generate_http_request(
+def test_audio_generate_missing_input(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_audio_generate_http_request(
         {
             "json": {"model": omni_server.model},
             "timeout": 120,

tests/dfx/reliability/invalid_param_test/test_invalid_audio_speech.py

@@ -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
 """HTTP + WebSocket validation for Qwen3-TTS Base, Higgs Audio V3 and Audex TTA ``/v1/audio/speech`` (and related routes)."""
 
 from __future__ import annotations
@@ -13,8 +13,8 @@ from typing import Any
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.media import load_test_audio_data_url
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.media import get_asset_path
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 from tests.helpers.stage_config import get_deploy_config_path
 
 # Speech / voice-upload numeric caps for these tests only (not exported from vllm_omni).
@@ -26,7 +26,7 @@ _SPEECH_API_MAX_NEW_TOKENS = 4096
 _VOICE_UPLOAD_MAX_CONSENT_LEN = 1024
 _VOICE_UPLOAD_MAX_REF_TEXT_CHARS = 8192
 _VOICE_UPLOAD_MAX_SPEAKER_DESCRIPTION_CHARS = 2048
-REF_AUDIO_URL = load_test_audio_data_url("qwen3_tts/clone_2.wav")
+REF_AUDIO_URL = get_asset_path("qwen3_tts/clone_2.wav", as_data_url=True)
 REF_TEXT = "Okay. Yeah. I resent you. I love you. I respect you. But you know what? You blew it! And thanks to you."
 # Qwen3-TTS 0.6B expects 1024-dim speaker embeddings (see serving_speech / talker hidden_size).
 _QWEN3_TTS_06B_SPEAKER_EMBEDDING_DIM = 1024
@@ -57,6 +57,7 @@ _QWEN3_TTS_SPEECH = [
     ),
 ]
 
+
 _HIGGS_AUDIO_V3_SPEECH = [
     pytest.param(
         OmniServerParams(
@@ -84,7 +85,7 @@ _AUDEX_TTA_SPEECH = [
 ]
 
 
-def _apply_batch_overrides(body: dict[str, object], *, loc: str, overrides: dict[str, object]) -> None:
+def _apply_batch_overrides(body: dict[str, Any], *, loc: str, overrides: dict[str, object]) -> None:
     if loc == "batch":
         body.update(overrides)
         return
@@ -112,8 +113,8 @@ def _pcm_wav_mono_bytes(*, duration_s: float = 1.1, sample_rate: int = 44100) ->
 
 
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
-def test_speech_malformed_json(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_audio_speech_http_request(
+def test_speech_malformed_json(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_audio_speech_http_request(
         {
             "raw_body": "{",
             "timeout": 120,
@@ -124,8 +125,8 @@ def test_speech_malformed_json(omni_server: OmniServer, openai_client: OpenAICli
 
 
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
-def test_speech_missing_required_fields(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_audio_speech_http_request(
+def test_speech_missing_required_fields(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_audio_speech_http_request(
         {
             "json": {"model": omni_server.model},
             "timeout": 120,
@@ -156,8 +157,8 @@ def test_speech_missing_required_fields(omni_server: OmniServer, openai_client:
                 "ref_audio": None,
                 "ref_text": None,
             },
-            ("CustomVoice", "does not support"),
-            id="customvoice_task_not_supported",
+            ("Invalid voice", "vivian"),
+            id="customvoice_invalid_voice",
         ),
         pytest.param(
             {"task_type": "InvalidEnum"}, ("task_type", "literal_error", "CustomVoice"), id="task_type_invalid"
@@ -180,13 +181,17 @@ def test_speech_missing_required_fields(omni_server: OmniServer, openai_client:
             ("mutually exclusive", "speaker_embedding", "ref_audio"),
             id="speaker_embedding_with_ref_audio_mutually_exclusive",
         ),
-        pytest.param({"max_new_tokens": 0}, ("max_new_tokens", "least 1"), id="max_new_tokens_below_min"),
+        pytest.param({"max_new_tokens": 0}, ("max_new_tokens", "greater_than_equal"), id="max_new_tokens_below_min"),
         pytest.param(
             {"max_new_tokens": _SPEECH_API_MAX_NEW_TOKENS + 1},
             ("max_new_tokens", "exceed 4096"),
             id="max_new_tokens_above_max",
         ),
-        pytest.param({"seed": -1}, ("seed", "greater_than_equal", "0"), id="seed_negative"),
+        pytest.param(
+            {"seed": -(2**63) - 1},
+            ("seed", "greater_than_equal", "-9223372036854775808"),
+            id="seed_below_min",
+        ),
         pytest.param({"seed": 2**63}, ("seed", "less_than_equal", "9223372036854775807"), id="seed_above_max"),
         pytest.param(
             {"initial_codec_chunk_frames": -1},
@@ -242,13 +247,13 @@ def test_speech_missing_required_fields(omni_server: OmniServer, openai_client:
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
 def test_speech_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     overrides: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     ov = dict(overrides)
     embedding_only = bool(ov.pop(_SPEECH_INVALID_EMBEDDING_ONLY_BODY, False))
-    body = {
+    body: dict[str, Any] = {
         "model": omni_server.model,
         "input": "Hello.",
         "voice": "clone",
@@ -261,7 +266,7 @@ def test_speech_invalid_field_values(
         body.pop("ref_audio", None)
         body.pop("ref_text", None)
     body.update(ov)
-    openai_client.send_audio_speech_http_request(
+    online_client.send_audio_speech_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -274,7 +279,9 @@ def test_speech_invalid_field_values(
         pytest.param({"input": "", "response_format": "wav"}, ("input", "empty"), id="input_empty"),
         pytest.param({"input": "   ", "response_format": "wav"}, ("input", "empty"), id="input_whitespace_only"),
         pytest.param(
-            {"input": "Hello world.", "max_new_tokens": 0}, ("max_new_tokens", "least 1"), id="max_new_tokens_below_min"
+            {"input": "Hello world.", "max_new_tokens": 0},
+            ("max_new_tokens", "greater_than_equal"),
+            id="max_new_tokens_below_min",
         ),
         pytest.param(
             {
@@ -328,12 +335,12 @@ def test_speech_invalid_field_values(
 @pytest.mark.parametrize("omni_server", _HIGGS_AUDIO_V3_SPEECH, indirect=True)
 def test_speech_higgs_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     extra_json: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     body = {"model": omni_server.model, **extra_json}
-    openai_client.send_audio_speech_http_request(
+    online_client.send_audio_speech_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -359,12 +366,12 @@ def test_speech_higgs_invalid_field_values(
 @pytest.mark.parametrize("omni_server", _AUDEX_TTA_SPEECH, indirect=True)
 def test_speech_audex_tta_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     overrides: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     body = {"model": omni_server.model, "input": _AUDEX_TTA_CAPTION, **overrides}
-    openai_client.send_audio_speech_http_request(
+    online_client.send_audio_speech_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -375,15 +382,15 @@ def test_speech_audex_tta_invalid_field_values(
 
 
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
-def test_speech_batch_empty_items(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_audio_speech_batch_http_request(
+def test_speech_batch_empty_items(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_audio_speech_batch_http_request(
         {"json": {"items": []}, "timeout": 120, "err_code": (400, 422), "err_message": ("items", "least 1 item")}
     )
 
 
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
-def test_speech_batch_missing_items(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_audio_speech_batch_http_request(
+def test_speech_batch_missing_items(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_audio_speech_batch_http_request(
         {
             "json": {"model": omni_server.model},
             "timeout": 120,
@@ -553,14 +560,14 @@ def test_speech_batch_missing_items(omni_server: OmniServer, openai_client: Open
         pytest.param(
             "batch",
             {"max_new_tokens": 0},
-            ("max_new_tokens", "least 1"),
+            ("max_new_tokens", "greater_than_equal"),
             id="batch_max_new_tokens_below_min",
             marks=_SKIP_ISSUE_3649,
         ),
         pytest.param(
             "item",
             {"max_new_tokens": 0},
-            ("max_new_tokens", "least 1"),
+            ("max_new_tokens", "greater_than_equal"),
             id="item_max_new_tokens_below_min",
             marks=_SKIP_ISSUE_3649,
         ),
@@ -618,14 +625,14 @@ def test_speech_batch_missing_items(omni_server: OmniServer, openai_client: Open
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
 def test_speech_batch_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     loc: str,
     overrides: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
     ov = dict(overrides)
     strip_clone = bool(ov.pop(_SPEECH_BATCH_NO_CLONE_FIELDS, False))
-    body = {
+    body: dict[str, Any] = {
         "model": omni_server.model,
         "voice": "clone",
         "ref_audio": REF_AUDIO_URL,
@@ -638,7 +645,7 @@ def test_speech_batch_invalid_field_values(
         body.pop("ref_audio", None)
         body.pop("ref_text", None)
     _apply_batch_overrides(body, loc=loc, overrides=ov)
-    openai_client.send_audio_speech_batch_http_request(
+    online_client.send_audio_speech_batch_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -670,7 +677,7 @@ _SPEECH_STREAM_WS_SESSION_STREAM_AUDIO_WAV_FRAMES = object()
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
 def test_speech_stream_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     send_frames_spec: Any,
     err_message: str | tuple[str, ...],
 ) -> None:
@@ -687,7 +694,7 @@ def test_speech_stream_invalid_requests(
     else:
         send_frames = send_frames_spec
     assert isinstance(send_frames, str)
-    openai_client.send_audio_speech_stream_ws_request(
+    online_client.send_audio_speech_stream_ws_request(
         {
             "send_frames": send_frames,
             "timeout": 120,
@@ -893,7 +900,7 @@ def _voices_upload_multipart_files(kind: str | None) -> dict[str, Any] | None:
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
 def test_voices_create_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     multipart_kind: str | None,
     form_template: dict[str, Any],
     err_message: str | tuple[str, ...],
@@ -909,7 +916,7 @@ def test_voices_create_invalid_requests(
     files = _voices_upload_multipart_files(multipart_kind)
     if files is not None:
         cfg["files"] = files
-    openai_client.send_audio_voices_create_http_request(cfg)
+    online_client.send_audio_voices_create_http_request(cfg)
 
 
 @pytest.mark.parametrize(
@@ -962,13 +969,13 @@ def test_voices_create_invalid_requests(
 @pytest.mark.parametrize("omni_server", _QWEN3_TTS_SPEECH, indirect=True)
 def test_voices_delete_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     voice_name: str,
     err_code: int | tuple[int, ...],
     err_message: str | tuple[str, ...],
 ) -> None:
     """DELETE ``/v1/audio/voices/{name}``: missing voice (404), odd segments (often 404), ``.``paths (405)."""
-    openai_client.send_audio_voices_delete_http_request(
+    online_client.send_audio_voices_delete_http_request(
         {
             "name": voice_name,
             "timeout": 120,
@@ -1021,7 +1028,7 @@ _MINIMAX_MUSIC3_MAX_FRAMES = 9000
 @pytest.mark.parametrize("omni_server", _MINIMAX_MUSIC3_SPEECH, indirect=True)
 def test_minimax_music3_speech_invalid_field_values(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     overrides: dict[str, object],
     err_message: str | tuple[str, ...],
 ) -> None:
@@ -1038,6 +1045,6 @@ def test_minimax_music3_speech_invalid_field_values(
             body.pop(key, None)
         else:
             body[key] = value
-    openai_client.send_audio_speech_http_request(
+    online_client.send_audio_speech_http_request(
         {"json": body, "timeout": 120, "err_code": 400, "err_message": err_message}
     )

tests/dfx/reliability/invalid_param_test/test_invalid_image_editing.py

@@ -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
 """``POST /v1/images/edits`` multipart validation (live ``Qwen/Qwen-Image-Edit``)."""
 
 from __future__ import annotations
@@ -9,7 +9,7 @@ from typing import Any
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
 pytestmark = [pytest.mark.slow, pytest.mark.diffusion]
 
@@ -72,7 +72,7 @@ def _finalize_edits_form_data(template: dict[str, Any], server_model: str) -> di
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_images_edits_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     tiny_png_bytes: bytes,
     include_image: bool,
     data_template: dict[str, Any],
@@ -87,4 +87,4 @@ def test_images_edits_invalid_requests(
     }
     if include_image:
         cfg["files"] = {"image": ("x.png", tiny_png_bytes, "image/png")}
-    openai_client.send_images_edits_http_request(cfg)
+    online_client.send_images_edits_http_request(cfg)

tests/dfx/reliability/invalid_param_test/test_invalid_image_generation.py

@@ -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
 """``POST /v1/images/generations`` validation (live ``Qwen/Qwen-Image``)."""
 
 from __future__ import annotations
@@ -9,7 +9,7 @@ from typing import Any
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
 pytestmark = [pytest.mark.slow, pytest.mark.diffusion]
 
@@ -126,7 +126,7 @@ def _minimal_images_gen_json(omni_server: OmniServer) -> dict[str, object]:
 @pytest.mark.parametrize("omni_server", _PARAMS, indirect=True)
 def test_images_generations_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     body_spec: Any,
     err_message: str | tuple[str, ...],
 ) -> None:
@@ -137,6 +137,6 @@ def test_images_generations_invalid_requests(
         assert isinstance(body_spec, dict)
         body = _minimal_images_gen_json(omni_server)
         body.update(body_spec)
-    openai_client.send_images_generations_http_request(
+    online_client.send_images_generations_http_request(
         {"json": body, "timeout": 300, "err_code": 400, "err_message": err_message}
     )

tests/dfx/reliability/invalid_param_test/test_invalid_omni_chat.py

@@ -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
 """Invalid inputs on Qwen3-Omni: ``POST /v1/chat/completions``, ``WS /v1/video/chat/stream``, ``WS /v1/realtime``."""
 
 from __future__ import annotations
@@ -10,7 +10,7 @@ from typing import Any
 import pytest
 
 from tests.helpers.mark import hardware_test
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 from tests.helpers.stage_config import get_deploy_config_path
 
 pytestmark = [pytest.mark.slow, pytest.mark.omni]
@@ -72,6 +72,9 @@ def _chat_completions_request_without_expectations(omni_server: OmniServer, case
         body["top_logprobs"] = 5
     elif case_id == "speaker_unknown":
         body["speaker"] = "zz_invalid_qwen3_omni_chat_speaker_xyz"
+    elif case_id == "audio_format_unsupported":
+        body["modalities"] = ["text", "audio"]
+        body["audio"] = {"voice": "alloy", "format": "aac"}
     else:
         raise AssertionError(f"unknown chat completions invalid case_id {case_id!r}")
     return {"json": body, "timeout": 120}
@@ -119,7 +122,7 @@ def _chat_completions_request_without_expectations(omni_server: OmniServer, case
         pytest.param(
             "response_format_json_schema_incomplete",
             400,
-            ("response_format", "value_error", "json_schema"),
+            ("response_format", "BadRequestError", "json_schema"),
             id="invalid_response_format_json_schema",
         ),
         pytest.param("logprobs_wrong_type", 400, "logprobs", id="logprobs_wrong_type", marks=_SKIP_ISSUE_3649),
@@ -135,12 +138,18 @@ def _chat_completions_request_without_expectations(omni_server: OmniServer, case
             ("Invalid speaker", "Supported"),
             id="speaker_unknown_preset",
         ),
+        pytest.param(
+            "audio_format_unsupported",
+            400,
+            ("Invalid audio format", "aac", "Supported formats"),
+            id="unsupported_audio_format",
+        ),
     ],
 )
 @pytest.mark.parametrize("omni_server", _QWEN3_OMNI_SERVER, indirect=True)
 def test_chat_completions_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     case_id: str,
     err_code: int | tuple[int, ...],
     err_message: str | tuple[str, ...],
@@ -148,7 +157,7 @@ def test_chat_completions_invalid_requests(
     cfg = _chat_completions_request_without_expectations(omni_server, case_id)
     cfg["err_code"] = err_code
     cfg["err_message"] = err_message
-    openai_client.send_chat_completions_http_request(cfg)[0]
+    online_client.send_chat_completions_http_request(cfg)[0]
 
 
 # ─────────────────────────────────────────────────────────────────────────────
@@ -174,7 +183,7 @@ _VIDEO_CHAT_WS_SESSION_MODALITIES_SCALAR = object()
 @pytest.mark.parametrize("omni_server", _QWEN3_OMNI_SERVER, indirect=True)
 def test_video_chat_stream_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     send_frames_spec: Any,
     err_message: str,
 ) -> None:
@@ -183,7 +192,7 @@ def test_video_chat_stream_invalid_requests(
     else:
         send_frames = send_frames_spec
     assert isinstance(send_frames, str)
-    openai_client.send_video_chat_stream_ws_request(
+    online_client.send_video_chat_stream_ws_request(
         {
             "send_frames": send_frames,
             "timeout": 120,
@@ -243,11 +252,12 @@ _REALTIME_WS_INVALID_AUDIO_APPEND = object()
 @pytest.mark.parametrize("omni_server", _QWEN3_OMNI_SERVER, indirect=True)
 def test_realtime_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     send_frames_spec: Any,
     ws_error_code: str,
     err_message: str | tuple[str, ...],
 ) -> None:
+    send_frames: Any
     if send_frames_spec is _REALTIME_WS_MODEL_MISMATCH:
         send_frames = json.dumps(
             {
@@ -262,7 +272,7 @@ def test_realtime_invalid_requests(
         ]
     else:
         send_frames = send_frames_spec
-    openai_client.send_realtime_ws_request(
+    online_client.send_realtime_ws_request(
         {
             "send_frames": send_frames,
             "timeout": 120,

tests/dfx/reliability/invalid_param_test/test_invalid_server_control.py

@@ -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
 """Omni sleep/wakeup JSON validation and level-2 wake contract (live ``Qwen/Qwen-Image``)."""
 
 from __future__ import annotations
@@ -9,7 +9,7 @@ from typing import Any, Literal
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
 pytestmark = [pytest.mark.slow, pytest.mark.diffusion, pytest.mark.omni]
 
@@ -86,7 +86,7 @@ _QWEN_IMAGE = [
 @pytest.mark.parametrize("omni_server", _QWEN_IMAGE, indirect=True)
 def test_omni_sleep_wakeup_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     endpoint: Literal["sleep", "wakeup"],
     json_body: dict[str, Any],
     err_message: str | tuple[str, ...],
@@ -98,9 +98,9 @@ def test_omni_sleep_wakeup_invalid_requests(
         "err_message": err_message,
     }
     if endpoint == "sleep":
-        openai_client.send_omni_sleep_http_request(cfg)
+        online_client.send_omni_sleep_http_request(cfg)
     else:
-        openai_client.send_omni_wakeup_http_request(cfg)
+        online_client.send_omni_wakeup_http_request(cfg)
 
 
 # Keep this after the invalid-JSON cases: sleep(level=2) discards weights and the
@@ -108,7 +108,7 @@ def test_omni_sleep_wakeup_invalid_requests(
 @pytest.mark.parametrize("omni_server", _QWEN_IMAGE, indirect=True)
 def test_wakeup_after_level2_sleep_fails(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
 ) -> None:
     """Regression for #4473 Repro A via HTTP: wake after sleep(level=2) → 501.
 
@@ -116,7 +116,7 @@ def test_wakeup_after_level2_sleep_fails(
     Live serve surfaces ``NotImplementedError`` as OpenAI-style JSON with code 501.
     """
     assert omni_server is not None
-    sleep_resps = openai_client.send_omni_sleep_http_request(
+    sleep_resps = online_client.send_omni_sleep_http_request(
         {
             "json": {"stage_ids": [0], "level": 2},
             "timeout": 180,
@@ -124,10 +124,11 @@ def test_wakeup_after_level2_sleep_fails(
     )
     sleep_resp = sleep_resps[0]
     assert sleep_resp.status_code == 200, sleep_resp
-    assert sleep_resp.json_body is not None
-    assert sleep_resp.json_body.get("status") == "SUCCESS", sleep_resp.json_body
+    json_body = sleep_resp.json_body
+    assert isinstance(json_body, dict)
+    assert json_body.get("status") == "SUCCESS", json_body
 
-    openai_client.send_omni_wakeup_http_request(
+    online_client.send_omni_wakeup_http_request(
         {
             "json": {"stage_ids": [0]},
             "timeout": 120,

tests/dfx/reliability/invalid_param_test/test_invalid_video_generation.py

@@ -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
 """HTTP video job routes (live ``Wan-AI/Wan2.2-T2V-A14B-Diffusers``).
 
 WebSocket ``/v1/video/chat/stream`` and ``/v1/realtime`` invalid-input checks live in
@@ -13,7 +13,7 @@ from typing import Any
 import pytest
 
 from tests.helpers.mark import hardware_marks
-from tests.helpers.runtime import OmniServer, OmniServerParams, OpenAIClientHandler
+from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
 pytestmark = [pytest.mark.slow, pytest.mark.diffusion]
 
@@ -54,12 +54,12 @@ def _minimal_video_form_data() -> dict[str, str]:
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
 def test_post_videos_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     form_extra: dict[str, str] | None,
     err_message: str | tuple[str, ...],
 ) -> None:
     data: dict[str, str] = {} if form_extra is None else {**_minimal_video_form_data(), **form_extra}
-    openai_client.send_videos_create_http_request(
+    online_client.send_videos_create_http_request(
         {"data": data, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -85,12 +85,12 @@ def test_post_videos_invalid_requests(
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
 def test_post_videos_sync_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     form_extra: dict[str, str] | None,
     err_message: str | tuple[str, ...],
 ) -> None:
     data: dict[str, str] = {} if form_extra is None else {**_minimal_video_form_data(), **form_extra}
-    openai_client.send_videos_sync_http_request(
+    online_client.send_videos_sync_http_request(
         {"data": data, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -111,11 +111,11 @@ def test_post_videos_sync_invalid_requests(
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
 def test_get_videos_list_invalid_requests(
     omni_server: OmniServer,
-    openai_client: OpenAIClientHandler,
+    online_client: OnlineOmniClient,
     params: dict[str, Any],
     err_message: str | tuple[str, ...],
 ) -> None:
-    openai_client.send_videos_list_http_request(
+    online_client.send_videos_list_http_request(
         {"params": params, "timeout": 120, "err_code": 400, "err_message": err_message}
     )
 
@@ -126,8 +126,8 @@ def test_get_videos_list_invalid_requests(
 
 
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
-def test_get_video_invalid_requests(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_video_retrieve_http_request(
+def test_get_video_invalid_requests(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_video_retrieve_http_request(
         {
             "video_id": "does-not-exist",
             "timeout": 120,
@@ -143,8 +143,8 @@ def test_get_video_invalid_requests(omni_server: OmniServer, openai_client: Open
 
 
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
-def test_delete_video_invalid_requests(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_video_delete_http_request(
+def test_delete_video_invalid_requests(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_video_delete_http_request(
         {
             "video_id": "does-not-exist",
             "timeout": 120,
@@ -160,8 +160,8 @@ def test_delete_video_invalid_requests(omni_server: OmniServer, openai_client: O
 
 
 @pytest.mark.parametrize("omni_server", _WAN_T2V, indirect=True)
-def test_get_video_content_invalid_requests(omni_server: OmniServer, openai_client: OpenAIClientHandler) -> None:
-    openai_client.send_video_content_http_request(
+def test_get_video_content_invalid_requests(omni_server: OmniServer, online_client: OnlineOmniClient) -> None:
+    online_client.send_video_content_http_request(
         {
             "video_id": "does-not-exist",
             "timeout": 120,

tests/dfx/reliability/test_reliability_hunyuan_image.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 HunyuanImage-3.0-Instruct reliability integration tests (DiT-only deploy).
 
@@ -35,7 +38,7 @@ from tests.dfx.reliability.helpers import (
     worker_residual_timeout_after_kill_signal,
 )
 from tests.helpers.mark import hardware_test
-from tests.helpers.runtime import OpenAIClientHandler
+from tests.helpers.runtime import OnlineOmniClient
 
 RELIABILITY_SCENARIOS: list[dict[str, Any]] = [
     {
@@ -133,8 +136,8 @@ def _image_request_config(omni_server: Any) -> dict[str, Any]:
     }
 
 
-def _submit_image_generation(openai_client: OpenAIClientHandler, omni_server: Any) -> None:
-    responses = openai_client.send_images_generations_http_request(_image_request_config(omni_server))
+def _submit_image_generation(online_client: OnlineOmniClient, omni_server: Any) -> None:
+    responses = online_client.send_images_generations_http_request(_image_request_config(omni_server))
     resp = responses[0]
     if resp.status_code >= 400:
         raise RuntimeError(
@@ -189,7 +192,7 @@ def _assert_post_fault_health_terminal(host: str, port: int, *, scenario: str) -
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_gpu_oom_image_large_request_failure(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
 ) -> None:
     stage_config_path = getattr(omni_server_function, "stage_config_path", None)
     device_spec = resolve_oom_device_spec(OOM_INJECTION_CONFIG, stage_config_path)
@@ -202,7 +205,7 @@ def test_reliability_fault_gpu_oom_image_large_request_failure(
     )
     try:
         try:
-            _submit_image_generation(openai_client_function, omni_server_function)
+            _submit_image_generation(online_client_function, omni_server_function)
         except Exception as exc:
             assert_fault_exception(exc, FAULT_ERROR_KEYWORDS)
         else:
@@ -222,10 +225,10 @@ def test_reliability_fault_gpu_oom_image_large_request_failure(
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_request_failure(
     omni_server_after_fault_function,
-    openai_client_function,
+    online_client_function,
 ) -> None:
     try:
-        _submit_image_generation(openai_client_function, omni_server_after_fault_function)
+        _submit_image_generation(online_client_function, omni_server_after_fault_function)
     except Exception as exc:
         assert_fault_exception(exc, PROCESS_KILL_ERROR_KEYWORDS)
     else:
@@ -244,7 +247,7 @@ def test_reliability_fault_process_kill_request_failure(
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
     omni_server_after_fault_function,
-    openai_client_function,
+    online_client_function,
 ) -> None:
     host = omni_server_after_fault_function.host
     port = omni_server_after_fault_function.port
@@ -280,7 +283,7 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
     start = time.monotonic()
     with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
         futures = [
-            executor.submit(_submit_image_generation, openai_client_function, omni_server_after_fault_function)
+            executor.submit(_submit_image_generation, online_client_function, omni_server_after_fault_function)
             for _ in range(3)
         ]
         done, pending = concurrent.futures.wait(
@@ -335,12 +338,12 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_worker_with_load_request_failure(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     fault_injector: FaultInjector,
 ) -> None:
     scenario = "kill_worker_with_load"
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: _submit_image_generation(openai_client_function, omni_server_function),
+        submit_request=lambda: _submit_image_generation(online_client_function, omni_server_function),
         inject_fault=lambda: fault_injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -363,13 +366,13 @@ def test_reliability_fault_process_kill_worker_with_load_request_failure(
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_serve_root_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     scenario = f"kill_serve_root_with_load_{signal_name.lower()}"
     injector = make_server_root_kill_fault_injector(signal_name=signal_name, post_kill_wait_seconds=2.0)
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: _submit_image_generation(openai_client_function, omni_server_function),
+        submit_request=lambda: _submit_image_generation(online_client_function, omni_server_function),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -448,7 +451,7 @@ def test_reliability_fault_process_kill_tree_no_load_fast_fail_and_cleanup(
 @pytest.mark.parametrize("omni_server_function", HUNYUAN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     scenario = f"kill_serve_tree_with_load_{signal_name.lower()}"
@@ -458,7 +461,7 @@ def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
         inter_kill_wait_seconds=0.1,
     )
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: _submit_image_generation(openai_client_function, omni_server_function),
+        submit_request=lambda: _submit_image_generation(online_client_function, omni_server_function),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,

tests/dfx/reliability/test_reliability_qwen3_omni.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 Qwen3-Omni reliability integration tests.
 """
@@ -63,7 +66,7 @@ DEPLOY_CONFIGS_DIR = Path(__file__).resolve().parent.parent.parent.parent / "vll
 
 def _default_oom_device_spec() -> str:
     """Use currently visible CUDA ordinals to avoid invalid device index in sidecar."""
-    count = torch.accelerator.device_count()
+    count = torch.accelerator.device_count() or 0
     if count <= 0:
         return "0"
     return ",".join(str(i) for i in range(count))
@@ -414,7 +417,7 @@ def test_reliability_fault_gpu_oom_error_contract_consistent_chat_speech(
 @pytest.mark.slow
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
-def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_function, openai_client_function) -> None:
+def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_function, online_client_function) -> None:
     device_spec = resolve_oom_device_spec(
         OOM_INJECTION_CONFIG,
         _deploy_config_path_from_omni_server(omni_server_function),
@@ -444,7 +447,7 @@ def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_functi
             "key_words": {"audio": ["test"]},
         }
         try:
-            openai_client_function.send_omni_request(request_config, request_num=1)
+            online_client_function.send_omni_request(request_config, request_num=1)
         except Exception as exc:
             assert_fault_exception(exc, FAULT_ERROR_KEYWORDS)
         else:
@@ -456,7 +459,7 @@ def test_reliability_fault_gpu_oom_chat_large_payload_failure(omni_server_functi
 @pytest.mark.slow
 @hardware_test(res={"cuda": "H100"}, num_cards=2)
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
-def test_reliability_fault_gpu_oom_concurrent_pressure_failure(omni_server_function, openai_client_function) -> None:
+def test_reliability_fault_gpu_oom_concurrent_pressure_failure(omni_server_function, online_client_function) -> None:
     device_spec = resolve_oom_device_spec(
         OOM_INJECTION_CONFIG,
         _deploy_config_path_from_omni_server(omni_server_function),
@@ -487,7 +490,7 @@ def test_reliability_fault_gpu_oom_concurrent_pressure_failure(omni_server_funct
             "key_words": {"audio": ["test"]},
         }
         try:
-            openai_client_function.send_omni_request(request_config, request_num=4)
+            online_client_function.send_omni_request(request_config, request_num=4)
         except Exception as exc:
             assert_fault_exception(exc, FAULT_ERROR_KEYWORDS)
         else:
@@ -505,7 +508,7 @@ def test_reliability_fault_gpu_oom_concurrent_pressure_failure(omni_server_funct
 )
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_request_failure(
-    omni_server_after_fault_function, openai_client_function
+    omni_server_after_fault_function, online_client_function
 ) -> None:
     messages = dummy_messages_from_mix_data(
         system_prompt=_get_system_prompt(),
@@ -519,7 +522,7 @@ def test_reliability_fault_process_kill_request_failure(
         "key_words": {"text": ["beijing"]},
     }
     try:
-        openai_client_function.send_omni_request(request_config, request_num=1)
+        online_client_function.send_omni_request(request_config, request_num=1)
     except Exception as exc:
         assert_fault_exception(exc, PROCESS_KILL_ERROR_KEYWORDS)
     else:
@@ -567,7 +570,7 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
         assert elapsed <= _PROCESS_KILL_WORKER_CHAT_FF_MAX_ELAPSED_SEC + 1.0, (
             f"[process_kill fast_fail] request did not fail fast after fault: {elapsed:.2f}s"
         )
-        assert ff_status >= 500, (
+        assert ff_status is not None and ff_status >= 500, (
             f"[process_kill fast_fail] expected server-side failure after fault, "
             f"got status={ff_status}, body={ff_body[:200]!r}"
         )
@@ -652,13 +655,13 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_worker_with_load_request_failure(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     fault_injector: FaultInjector,
 ) -> None:
     request_config = _chat_request_config(omni_server_function)
     scenario = "kill_worker_with_load"
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_omni_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_omni_request(request_config, request_num=1),
         inject_fault=lambda: fault_injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -704,14 +707,14 @@ def test_reliability_fault_process_kill_serve_root_no_load_fast_fail_and_cleanup
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_serve_root_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     request_config = _chat_request_config(omni_server_function)
     scenario = f"kill_serve_root_with_load_{signal_name.lower()}"
     injector = make_server_root_kill_fault_injector(signal_name=signal_name, post_kill_wait_seconds=2.0)
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_omni_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_omni_request(request_config, request_num=1),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -764,7 +767,7 @@ def test_reliability_fault_process_kill_tree_no_load_fast_fail_and_cleanup(
 @pytest.mark.parametrize("omni_server_function", QWEN_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     request_config = _chat_request_config(omni_server_function)
@@ -775,7 +778,7 @@ def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
         inter_kill_wait_seconds=0.1,
     )
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_omni_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_omni_request(request_config, request_num=1),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,

tests/dfx/reliability/test_reliability_wan22.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """
 Wan2.2 reliability integration tests.
 """
@@ -183,7 +186,7 @@ def _assert_post_fault_health_terminal(host: str, port: int, *, scenario: str) -
 @pytest.mark.slow
 @hardware_test(res={"cuda": "H100"}, num_cards=1)
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
-def test_reliability_fault_gpu_oom_video_large_request_failure(omni_server_function, openai_client_function) -> None:
+def test_reliability_fault_gpu_oom_video_large_request_failure(omni_server_function, online_client_function) -> None:
     stage_config_path = getattr(omni_server_function, "stage_config_path", None)
     device_spec = resolve_oom_device_spec(OOM_INJECTION_CONFIG, stage_config_path)
     handle = inject_gpu_oom(
@@ -211,7 +214,7 @@ def test_reliability_fault_gpu_oom_video_large_request_failure(omni_server_funct
             "stream": False,
         }
         try:
-            openai_client_function.send_video_diffusion_request(request_config, request_num=1)
+            online_client_function.send_video_diffusion_request(request_config, request_num=1)
         except Exception as exc:
             assert_fault_exception(exc, FAULT_ERROR_KEYWORDS)
         else:
@@ -230,11 +233,11 @@ def test_reliability_fault_gpu_oom_video_large_request_failure(omni_server_funct
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_request_failure(
     omni_server_after_fault_function,
-    openai_client_function,
+    online_client_function,
 ) -> None:
     request_config = _video_request_config()
     try:
-        openai_client_function.send_video_diffusion_request(request_config, request_num=1)
+        online_client_function.send_video_diffusion_request(request_config, request_num=1)
     except Exception as exc:
         assert_fault_exception(exc, PROCESS_KILL_ERROR_KEYWORDS)
     else:
@@ -251,7 +254,7 @@ def test_reliability_fault_process_kill_request_failure(
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
     omni_server_after_fault_function,
-    openai_client_function,
+    online_client_function,
 ) -> None:
     """Black-box: after process kill, /v1/videos fails fast and concurrent calls don't hang.
 
@@ -308,7 +311,7 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
     start = time.monotonic()
     with concurrent.futures.ThreadPoolExecutor(max_workers=3) as executor:
         futures = [
-            executor.submit(openai_client_function.send_video_diffusion_request, request_config, 1) for _ in range(3)
+            executor.submit(online_client_function.send_video_diffusion_request, request_config, 1) for _ in range(3)
         ]
         done, pending = concurrent.futures.wait(
             futures,
@@ -367,13 +370,13 @@ def test_reliability_fault_process_kill_health_fast_fail_and_concurrent(
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_worker_with_load_request_failure(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     fault_injector: FaultInjector,
 ) -> None:
     request_config = _video_request_config()
     scenario = "kill_worker_with_load"
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_video_diffusion_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_video_diffusion_request(request_config, request_num=1),
         inject_fault=lambda: fault_injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -516,7 +519,7 @@ def test_reliability_video_oom_recovers_after_fault_removed(
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_serve_root_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     """Black-box: during in-flight video requests, kill serve root and verify post-fault behavior/cleanup."""
@@ -524,7 +527,7 @@ def test_reliability_fault_process_kill_serve_root_with_load_fast_fail_and_clean
     scenario = f"kill_serve_root_with_load_{signal_name.lower()}"
     injector = make_server_root_kill_fault_injector(signal_name=signal_name, post_kill_wait_seconds=2.0)
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_video_diffusion_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_video_diffusion_request(request_config, request_num=1),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,
@@ -601,7 +604,7 @@ def test_reliability_fault_process_kill_tree_no_load_fast_fail_and_cleanup(
 @pytest.mark.parametrize("omni_server_function", DIFFUSION_VIDEO_PARAMS, indirect=True)
 def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
     omni_server_function,
-    openai_client_function,
+    online_client_function,
     signal_name: str,
 ) -> None:
     """Black-box: during in-flight video requests, kill server process tree and verify post-fault behavior/cleanup."""
@@ -613,7 +616,7 @@ def test_reliability_fault_process_kill_tree_with_load_fast_fail_and_cleanup(
         inter_kill_wait_seconds=0.1,
     )
     load_result = run_fault_injection_with_rate_load(
-        submit_request=lambda: openai_client_function.send_video_diffusion_request(request_config, request_num=1),
+        submit_request=lambda: online_client_function.send_video_diffusion_request(request_config, request_num=1),
         inject_fault=lambda: injector(omni_server_function),
         num_requests=INFLIGHT_INJECTION_REQUEST_COUNT,
         request_rate=INFLIGHT_INJECTION_REQUEST_RATE,

tests/dfx/stability/conftest.py

@@ -37,7 +37,7 @@ def omni_server(request: pytest.FixtureRequest):
     with _omni_server_lock:
         # Same tuple and CLI composition as ``tests/dfx/perf/scripts/run_benchmark.py``;
         # ``serve_args`` from JSON are folded into ``extra_cli_args`` inside
-        # ``create_unique_server_params``.
+        # ``create_unique_server_pytest_params`` (JSON ``mark`` → pytest marks).
         test_name, model, deploy_path, stage_overrides, extra_cli_args, use_omni = request.param
 
         print(f"Starting OmniServer with test: {test_name}, model: {model}")

tests/dfx/stability/scripts/test_stability_hunyuan_image.py

@@ -13,7 +13,7 @@ import pytest
 from tests.dfx.conftest import (
     create_benchmark_indices,
     create_test_parameter_mapping,
-    create_unique_server_params,
+    create_unique_server_pytest_params,
     load_configs,
 )
 from tests.dfx.stability.helpers import _run_one_diffusion_batch, run_stability_benchmark_loop
@@ -29,7 +29,7 @@ try:
 except FileNotFoundError:
     BENCHMARK_CONFIGS = []
 
-test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+test_params = create_unique_server_pytest_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
 server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
 benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
 

tests/dfx/stability/scripts/test_stability_qwen3_omni.py

@@ -13,7 +13,7 @@ import pytest
 from tests.dfx.conftest import (
     create_benchmark_indices,
     create_test_parameter_mapping,
-    create_unique_server_params,
+    create_unique_server_pytest_params,
     load_configs,
 )
 from tests.dfx.stability.helpers import _run_one_vllm_bench_batch, run_stability_benchmark_loop
@@ -29,7 +29,7 @@ try:
 except FileNotFoundError:
     BENCHMARK_CONFIGS = []
 
-test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+test_params = create_unique_server_pytest_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
 server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
 benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
 

tests/dfx/stability/scripts/test_stability_qwen3_tts.py

@@ -13,7 +13,7 @@ import pytest
 from tests.dfx.conftest import (
     create_benchmark_indices,
     create_test_parameter_mapping,
-    create_unique_server_params,
+    create_unique_server_pytest_params,
     load_configs,
 )
 from tests.dfx.stability.helpers import _run_one_vllm_bench_batch, run_stability_benchmark_loop
@@ -29,7 +29,7 @@ try:
 except FileNotFoundError:
     BENCHMARK_CONFIGS = []
 
-test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+test_params = create_unique_server_pytest_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
 server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
 benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
 

tests/dfx/stability/scripts/test_stability_qwen_image.py

@@ -13,7 +13,7 @@ import pytest
 from tests.dfx.conftest import (
     create_benchmark_indices,
     create_test_parameter_mapping,
-    create_unique_server_params,
+    create_unique_server_pytest_params,
     load_configs,
 )
 from tests.dfx.stability.helpers import _run_one_diffusion_batch, run_stability_benchmark_loop
@@ -29,7 +29,7 @@ try:
 except FileNotFoundError:
     BENCHMARK_CONFIGS = []
 
-test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+test_params = create_unique_server_pytest_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
 server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
 benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
 

tests/dfx/stability/scripts/test_stability_wan22.py

@@ -13,7 +13,7 @@ import pytest
 from tests.dfx.conftest import (
     create_benchmark_indices,
     create_test_parameter_mapping,
-    create_unique_server_params,
+    create_unique_server_pytest_params,
     load_configs,
 )
 from tests.dfx.stability.helpers import _run_one_diffusion_batch, run_stability_benchmark_loop
@@ -29,7 +29,7 @@ try:
 except FileNotFoundError:
     BENCHMARK_CONFIGS = []
 
-test_params = create_unique_server_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
+test_params = create_unique_server_pytest_params(BENCHMARK_CONFIGS, DEPLOY_CONFIGS_DIR) if BENCHMARK_CONFIGS else []
 server_to_benchmark_mapping = create_test_parameter_mapping(BENCHMARK_CONFIGS) if BENCHMARK_CONFIGS else {}
 benchmark_indices = create_benchmark_indices(BENCHMARK_CONFIGS, server_to_benchmark_mapping)
 

tests/dfx/stability/tests/test_hunyuan_image.json

@@ -1,6 +1,18 @@
 [
     {
         "test_name": "test_hunyuan_image_stability_tp2_fp8_sp2",
+        "mark": [
+            {
+                "hardware_marks": {
+                    "res": {
+                        "cuda": "H800"
+                    },
+                    "num_cards": 2
+                }
+            },
+            "local_model",
+            "diffusion"
+        ],
         "server_params": {
             "model": "tencent/HunyuanImage-3.0-Instruct",
             "serve_args": {

tests/dfx/stability/tests/test_qwen3_omni.json

@@ -1,6 +1,18 @@
 [
     {
         "test_name": "test_qwen3_omni_stability_async_chunk",
+        "mark": [
+            {
+                "hardware_marks": {
+                    "res": {
+                        "cuda": "H800"
+                    },
+                    "num_cards": 2
+                }
+            },
+            "local_model",
+            "omni"
+        ],
         "server_params": {
             "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
             "extra_cli_args": ["--async-chunk"]

tests/dfx/stability/tests/test_qwen3_tts.json

@@ -1,6 +1,19 @@
 [
     {
         "test_name": "test_qwen3_tts_stability",
+        "mark": [
+            {
+                "hardware_marks": {
+                    "res": {
+                        "cuda": "H800",
+                        "npu": "A3"
+                    },
+                    "num_cards": 2
+                }
+            },
+            "local_model",
+            "tts"
+        ],
         "server_params": {
             "model": "Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice"
         },