update release

Signed-off-by: hsliuustc0106 <liuhongsheng4@huawei.com>

.buildkite/test-nightly.yml

@@ -355,6 +355,48 @@ steps:
                   path: /mnt/hf-cache
                   type: DirectoryOrCreate
 
+  - label: ":full_moon: Wan22 I2V Accuracy Test with H100"
+    key: nightly-wan22-i2v-accuracy
+    timeout_in_minutes: 180
+    depends_on: upload-nightly-pipeline
+    if: build.env("NIGHTLY") == "1" || build.pull_request.labels includes "nightly-test"
+    commands:
+      - export VLLM_WORKER_MULTIPROC_METHOD=spawn
+      - pytest -s -v tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py --run-level advanced_model
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:c392ce21e9cf9ea65c52b866447793db10e0261c
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+                  - name: HF_TOKEN
+                    valueFrom:
+                      secretKeyRef:
+                        name: hf-token-secret
+                        key: token
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
+
   - label: ":full_moon: Qwen-Image Diffusion Perf Test with H100"
     key: nightly-qwen-image-performance
     timeout_in_minutes: 300

.buildkite/test-ready.yml

@@ -216,44 +216,42 @@ steps:
           volumes:
             - "/fsx/hf_cache:/fsx/hf_cache"
 
-  # - label: "Omni Model Test with H100"
-  #   depends_on: upload-ready-pipeline
-  #   commands:
-  #     - |
-  #       timeout 20m bash -c '
-  #         export VLLM_WORKER_MULTIPROC_METHOD=spawn
-  #         export VLLM_TEST_CLEAN_GPU_MEMORY="1"
-  #         # - pytest -s -v tests/e2e/offline_inference/test_qwen3_omni.py
-  #         pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
-  #       '
-  #   agents:
-  #     queue: "mithril-h100-pool"
-  #   plugins:
-  #     - kubernetes:
-  #         podSpec:
-  #           containers:
-  #             - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
-  #               resources:
-  #                 limits:
-  #                   nvidia.com/gpu: 2
-  #               volumeMounts:
-  #                 - name: devshm
-  #                   mountPath: /dev/shm
-  #                 - name: hf-cache
-  #                   mountPath: /root/.cache/huggingface
-  #               env:
-  #                 - name: HF_HOME
-  #                   value: /root/.cache/huggingface
-  #           nodeSelector:
-  #             node.kubernetes.io/instance-type: gpu-h100-sxm
-  #           volumes:
-  #             - name: devshm
-  #               emptyDir:
-  #                 medium: Memory
-  #             - name: hf-cache
-  #               hostPath:
-  #                 path: /mnt/hf-cache
-  #                 type: DirectoryOrCreate
+  - label: "Omni Model Test with H100"
+    depends_on: upload-ready-pipeline
+    commands:
+      - |
+        timeout 20m bash -c '
+          export VLLM_WORKER_MULTIPROC_METHOD=spawn
+          pytest -s -v tests/e2e/online_serving/test_qwen3_omni.py -m "core_model" --run-level "core_model"
+        '
+    agents:
+      queue: "mithril-h100-pool"
+    plugins:
+      - kubernetes:
+          podSpec:
+            containers:
+              - image: 936637512419.dkr.ecr.us-west-2.amazonaws.com/vllm-ci-pull-through-cache/q9t5s3a7/vllm-ci-test-repo:$BUILDKITE_COMMIT
+                resources:
+                  limits:
+                    nvidia.com/gpu: 2
+                volumeMounts:
+                  - name: devshm
+                    mountPath: /dev/shm
+                  - name: hf-cache
+                    mountPath: /root/.cache/huggingface
+                env:
+                  - name: HF_HOME
+                    value: /root/.cache/huggingface
+            nodeSelector:
+              node.kubernetes.io/instance-type: gpu-h100-sxm
+            volumes:
+              - name: devshm
+                emptyDir:
+                  medium: Memory
+              - name: hf-cache
+                hostPath:
+                  path: /mnt/hf-cache
+                  type: DirectoryOrCreate
 
   - label: "MiMo-Audio E2E Test with H100"
     depends_on: upload-ready-pipeline

README.md

@@ -16,7 +16,7 @@ Easy, fast, and cheap omni-modality model serving for everyone
 ---
 
 *Latest News* 🔥
-
+- [2026/03] We released [0.18.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.18.0) - strengthens the core runtime through a large entrypoint refactor and scheduler/runtime cleanups, expands unified quantization and diffusion execution, broadens multimodal model coverage, and improves production readiness across audio, omni, image, video, RL, and multi-platform deployments.
 - [2026/03] Check out our first public [project deepdive](https://youtu.be/sgwNfsNnR9I) at the vLLM Hong Kong Meetup!
 - [2026/03] **[vllm-omni-skills](https://github.com/hsliuustc0106/vllm-omni-skills)** is a community-driven collection of AI assistant skills that help developers work with vLLM-Omni more effectively. These skills can be used with popular agentic AI coding assistants like **Cursor IDE**, **Claude**, **Codex**, and more.
 - [2026/02] We released [0.16.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.16.0) - A major alignment + capability release that rebases onto **upstream vLLM v0.16.0** and significantly expands performance, distributed execution, and production readiness across **Qwen3-Omni / Qwen3-TTS**, **Bagel**, **MiMo-Audio**, **GLM-Image** and the **Diffusion (DiT) image/video stack**—while also improving platform coverage (CUDA / ROCm / NPU / XPU), CI quality, and documentation.

examples/online_serving/qwen3_tts/gradio_demo.py

@@ -125,7 +125,7 @@ PLAYER_HTML = """
 def _build_player_js(sample_rate: int) -> str:
     """Build the JavaScript that powers the AudioWorklet player."""
     return f"""
-() => {{
+    <script>
     const SR = {sample_rate};
     const WC = {json.dumps(WORKLET_JS)};
     let ctx = null, node = null, abort = null, gen = false, st = {{}};
@@ -284,7 +284,7 @@ def _build_player_js(sample_rate: int) -> str:
             }}
         }}
     }};
-}}
+    </script>
 """
 
 
@@ -435,10 +435,7 @@ def create_app(api_base: str):
     )
 
     with gr.Blocks(
-        css=css,
         title="Qwen3-TTS Demo",
-        js=_build_player_js(PCM_SAMPLE_RATE),
-        theme=theme,
     ) as demo:
         gr.HTML(f"""
         <div style="display:flex; align-items:center; gap:16px; margin-bottom:8px;">
@@ -773,7 +770,14 @@ def create_app(api_base: str):
 
         demo.queue()
 
-    return gr.mount_gradio_app(fastapi_app, demo, path="/")
+    return gr.mount_gradio_app(
+        fastapi_app,
+        demo,
+        path="/",
+        css=css,
+        theme=theme,
+        head=_build_player_js(PCM_SAMPLE_RATE),
+    )
 
 
 def parse_args():

examples/online_serving/qwen3_tts/tts_common.py

@@ -103,7 +103,7 @@ def build_payload(
 
     if task_type == "CustomVoice":
         if voice:
-            payload["speaker"] = voice
+            payload["voice"] = voice
         if instructions and instructions.strip():
             payload["instructions"] = instructions.strip()
 

tests/conftest.py

@@ -688,8 +688,112 @@ def generate_synthetic_audio(
     return result
 
 
-def generate_synthetic_video(width: int, height: int, num_frames: int, save_to_file: bool = False) -> dict[str, Any]:
-    """Generate synthetic video with bouncing balls and return base64 string."""
+def _mux_mp4_bytes_with_synthetic_audio(
+    video_mp4_bytes: bytes,
+    *,
+    num_frames: int,
+    fps: float = 30.0,
+    sample_rate: int = 48000,
+) -> bytes:
+    """
+    Mux a video-only MP4 with mono TTS audio from :func:`generate_synthetic_audio` (AAC).
+
+    Audio length is at least the video duration in whole seconds (rounded up); ffmpeg
+    ``-shortest`` trims to the video when the WAV is longer.
+
+    Uses ffmpeg from ``imageio_ffmpeg`` when available, else ``ffmpeg`` on PATH.
+    If TTS or mux fails, returns ``video_mp4_bytes`` unchanged.
+
+    Mux subprocess does **not** use ``capture_output=True``: ffmpeg can block writing
+    to a full stderr pipe while :func:`subprocess.run` waits for exit (classic deadlock).
+    """
+    duration_sec = num_frames / fps if fps > 0 else 0.0
+    # generate_synthetic_audio(duration=int) uses at least 1s of buffer internally
+    duration_int = max(1, int(math.ceil(duration_sec)))
+
+    try:
+        audio_result = generate_synthetic_audio(
+            duration=duration_int,
+            num_channels=1,
+            sample_rate=sample_rate,
+            save_to_file=False,
+        )
+        audio_pcm = audio_result["np_array"]
+    except Exception as e:
+        logger.warning("Synthetic video: generate_synthetic_audio failed (%s); using video-only MP4.", e)
+        return video_mp4_bytes
+
+    try:
+        import imageio_ffmpeg
+
+        ffmpeg_exe = imageio_ffmpeg.get_ffmpeg_exe()
+    except Exception:
+        ffmpeg_exe = "ffmpeg"
+
+    import tempfile
+
+    try:
+        with tempfile.TemporaryDirectory(prefix="syn_vid_mux_") as tmp:
+            vid_path = os.path.join(tmp, "video.mp4")
+            wav_path = os.path.join(tmp, "audio.wav")
+            out_path = os.path.join(tmp, "out.mp4")
+            with open(vid_path, "wb") as f:
+                f.write(video_mp4_bytes)
+            sf.write(wav_path, audio_pcm, sample_rate, format="WAV", subtype="PCM_16")
+            cmd = [
+                ffmpeg_exe,
+                "-y",
+                "-nostdin",
+                "-hide_banner",
+                "-loglevel",
+                "error",
+                "-i",
+                vid_path,
+                "-i",
+                wav_path,
+                "-c:v",
+                "copy",
+                "-c:a",
+                "aac",
+                "-b:a",
+                "128k",
+                "-shortest",
+                "-movflags",
+                "+faststart",
+                out_path,
+            ]
+            subprocess.run(
+                cmd,
+                check=True,
+                stdin=subprocess.DEVNULL,
+                timeout=300,
+            )
+            with open(out_path, "rb") as f:
+                return f.read()
+    except (
+        FileNotFoundError,
+        subprocess.CalledProcessError,
+        subprocess.TimeoutExpired,
+        OSError,
+    ) as e:
+        logger.warning("Synthetic video: audio mux failed (%s); using video-only MP4.", e)
+        return video_mp4_bytes
+
+
+def generate_synthetic_video(
+    width: int,
+    height: int,
+    num_frames: int,
+    save_to_file: bool = False,
+    *,
+    embed_audio: bool = False,
+) -> dict[str, Any]:
+    """Generate synthetic video with bouncing balls and base64 MP4.
+
+    When ``embed_audio`` is True, muxes mono AAC from :func:`generate_synthetic_audio`
+    (TTS + ffmpeg) into the MP4; otherwise returns video-only MP4 (faster when tests do
+    not need an audio track).
+    """
 
     import cv2
     import imageio
@@ -762,13 +866,13 @@ def generate_synthetic_video(width: int, height: int, num_frames: int, save_to_f
     result = {
         "np_array": video_array,
     }
-    video_bytes = None
     saved_file_path = None
 
+    fps = 30
     buffer = io.BytesIO()
     writer_kwargs = {
         "format": "mp4",
-        "fps": 30,
+        "fps": fps,
         "codec": "libx264",
         "quality": 7,
         "pixelformat": "yuv420p",
@@ -787,32 +891,31 @@ def generate_synthetic_video(width: int, height: int, num_frames: int, save_to_f
         ],
     }
 
-    if save_to_file:
-        import datetime
-
-        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
-        output_path = f"video_{width}x{height}_{timestamp}.mp4"
-        try:
-            with imageio.get_writer(output_path, **writer_kwargs) as writer:
-                for frame in video_frames:
-                    writer.append_data(frame)
-
-            saved_file_path = output_path
-            print(f"Video saved to: {saved_file_path}")
-            with open(output_path, "rb") as f:
-                video_bytes = f.read()
-
-        except Exception as e:
-            print(f"Warning: Failed to save video to file {output_path}: {e}")
-            save_to_file = False
-
-    if not save_to_file or video_bytes is None:
+    try:
         with imageio.get_writer(buffer, **writer_kwargs) as writer:
             for frame in video_frames:
                 writer.append_data(frame)
-
         buffer.seek(0)
-        video_bytes = buffer.read()
+        video_only_bytes = buffer.read()
+    except Exception as e:
+        print(f"Warning: Failed to encode synthetic video: {e}")
+        raise
+
+    if embed_audio:
+        video_bytes = _mux_mp4_bytes_with_synthetic_audio(video_only_bytes, num_frames=num_frames, fps=float(fps))
+    else:
+        video_bytes = video_only_bytes
+
+    if save_to_file:
+        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+        output_path = f"video_{width}x{height}_{timestamp}.mp4"
+        try:
+            with open(output_path, "wb") as f:
+                f.write(video_bytes)
+            saved_file_path = output_path
+            print(f"Video saved to: {saved_file_path}")
+        except Exception as e:
+            print(f"Warning: Failed to save video to file {output_path}: {e}")
 
     base64_video = base64.b64encode(video_bytes).decode("utf-8")
 
@@ -1262,10 +1365,12 @@ def modify_stage_config(
                 # Direct top-level key
                 config[key] = value
 
-    # Save to new file with timestamp
-    timestamp = int(time.time())
+    # Unique suffix: multiple modify_stage_config calls in one process often run
+    # within the same second (e.g. test_qwen3_omni_expansion imports both
+    # get_chunk_config and get_batch_token_config). int(time.time()) would collide
+    # and the later write would overwrite the earlier YAML on disk.
     base_name = yaml_path.rsplit(".", 1)[0] if "." in yaml_path else yaml_path
-    output_path = f"{base_name}_{timestamp}.yaml"
+    output_path = f"{base_name}_{time.time_ns()}.yaml"
 
     with open(output_path, "w", encoding="utf-8") as f:
         yaml.dump(config, f, default_flow_style=None, sort_keys=False, allow_unicode=True, indent=2)
@@ -1733,9 +1838,11 @@ def assert_omni_response(response: OmniResponse, request_config: dict[str, Any],
                         "The output does not contain any of the keywords."
                     )
 
-        # Verify similarity
+        # Verify similarity (Whisper transcript vs streamed/detokenized text)
         if "text" in modalities and "audio" in modalities:
-            assert response.similarity > 0.9, "The audio content is not same as the text"
+            assert response.similarity is not None and response.similarity > 0.9, (
+                "The audio content is not same as the text"
+            )
             print(f"similarity is: {response.similarity}")
 
 
@@ -2135,7 +2242,11 @@ class OpenAIClientHandler:
         Send OpenAI requests.
 
         Args:
-            request_config: Request configuration dictionary containing parameters like model, messages, stream
+            request_config: Request configuration dictionary containing parameters like model, messages, stream.
+                Optional ``use_audio_in_video`` (bool): when true, sets
+                ``extra_body["mm_processor_kwargs"] = {"use_audio_in_video": True}`` for Qwen-Omni video+audio
+                extraction (merged with any existing ``extra_body`` / ``mm_processor_kwargs``).
+                Optional ``extra_body`` (dict): passed through to ``chat.completions.create`` after merge.
             request_num: Number of requests, defaults to 1 (single request)
 
         Returns:
@@ -2146,14 +2257,28 @@ class OpenAIClientHandler:
         stream = request_config.get("stream", False)
         modalities = request_config.get("modalities", ["text", "audio"])
 
+        extra_body: dict[str, Any] = {}
+        raw_extra = request_config.get("extra_body")
+        if raw_extra:
+            extra_body.update(raw_extra)
+        if request_config.get("use_audio_in_video"):
+            mm = dict(extra_body.get("mm_processor_kwargs") or {})
+            mm["use_audio_in_video"] = True
+            extra_body["mm_processor_kwargs"] = mm
+        extra_body_arg: dict[str, Any] | None = extra_body if extra_body else None
+
+        create_kwargs: dict[str, Any] = {
+            "model": request_config.get("model"),
+            "messages": request_config.get("messages"),
+            "stream": stream,
+            "modalities": modalities,
+        }
+        if extra_body_arg is not None:
+            create_kwargs["extra_body"] = extra_body_arg
+
         if request_num == 1:
             # Send single request
-            chat_completion = self.client.chat.completions.create(
-                model=request_config.get("model"),
-                messages=request_config.get("messages"),
-                stream=stream,
-                modalities=modalities,
-            )
+            chat_completion = self.client.chat.completions.create(**create_kwargs)
 
             if stream:
                 response = self._process_stream_omni_response(chat_completion)

tests/e2e/accuracy/conftest.py

@@ -47,6 +47,19 @@ def pytest_addoption(parser):
         help="Balanced sample count per GEdit task group",
     )
     group.addoption("--accuracy-workers", action="store", type=int, default=1, help="Worker count for accuracy benches")
+    group.addoption(
+        "--wan22-i2v-image-source",
+        action="store",
+        default=None,
+        help="Image source for Wan2.2 I2V accuracy tests. Can be local path or remote URL.",
+    )
+    group.addoption(
+        "--wan22-i2v-online-timeout-seconds",
+        action="store",
+        type=int,
+        default=1200,
+        help="Online serving timeout in seconds for Wan2.2 I2V accuracy tests.",
+    )
 
 
 def _hf_cache_root() -> Path:
@@ -142,6 +155,17 @@ def accuracy_workers(request: pytest.FixtureRequest) -> int:
     return int(request.config.getoption("accuracy_workers"))
 
 
+@pytest.fixture(scope="session")
+def wan22_i2v_image_source(request: pytest.FixtureRequest) -> str | None:
+    value = request.config.getoption("wan22_i2v_image_source")
+    return str(value) if value else None
+
+
+@pytest.fixture(scope="session")
+def wan22_i2v_online_timeout_seconds(request: pytest.FixtureRequest) -> int:
+    return int(request.config.getoption("wan22_i2v_online_timeout_seconds"))
+
+
 @pytest.fixture(scope="session")
 def gebench_samples_per_type(request: pytest.FixtureRequest) -> int:
     return int(request.config.getoption("gebench_samples_per_type"))

tests/e2e/accuracy/wan22_i2v/run_wan22_i2v_diffusers_cp.py

@@ -0,0 +1,165 @@
+from __future__ import annotations
+
+import argparse
+import base64
+import json
+from io import BytesIO
+from pathlib import Path
+
+import requests
+import torch
+from diffusers import UniPCMultistepScheduler, WanImageToVideoPipeline
+from diffusers.pipelines.wan import pipeline_wan_i2v as wan_i2v_module
+from diffusers.utils import export_to_video, load_image
+from PIL import Image
+
+from tests.e2e.accuracy.wan22_i2v.wan22_i2v_video_similarity_common import BOUNDARY_RATIO
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Run Wan2.2 I2V diffusers offline generation.")
+    parser.add_argument("--model", required=True)
+    parser.add_argument("--image-source", required=True)
+    parser.add_argument("--prompt", required=True)
+    parser.add_argument("--negative-prompt", required=True)
+    parser.add_argument("--size", required=True)
+    parser.add_argument("--fps", type=int, required=True)
+    parser.add_argument("--num-frames", type=int, required=True)
+    parser.add_argument("--guidance-scale", type=float, required=True)
+    parser.add_argument("--guidance-scale-2", type=float, required=True)
+    parser.add_argument("--flow-shift", type=float, required=True)
+    parser.add_argument("--num-inference-steps", type=int, required=True)
+    parser.add_argument("--seed", type=int, required=True)
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--metadata-output", required=True)
+    return parser.parse_args()
+
+
+def _parse_size(size: str) -> tuple[int, int]:
+    width_str, height_str = size.lower().split("x", 1)
+    return int(width_str), int(height_str)
+
+
+class _IdentityFtfy:
+    @staticmethod
+    def fix_text(text: str) -> str:
+        return text
+
+
+def _ensure_wan_ftfy_fallback() -> None:
+    if not hasattr(wan_i2v_module, "ftfy"):
+        wan_i2v_module.ftfy = _IdentityFtfy()
+
+
+def _offline_cuda_device() -> torch.device:
+    return torch.device("cuda:0")
+
+
+def _load_input_image(source: str) -> Image.Image:
+    if source.startswith("data:image"):
+        _, encoded = source.split(",", 1)
+        image = Image.open(BytesIO(base64.b64decode(encoded)))
+        image.load()
+        return image.convert("RGB")
+
+    source_path = Path(source)
+    if source_path.exists():
+        image = Image.open(source_path)
+        image.load()
+        return image.convert("RGB")
+
+    image = load_image(source)
+    if isinstance(image, Image.Image):
+        image.load()
+        return image.convert("RGB")
+
+    response = requests.get(source, timeout=60)
+    response.raise_for_status()
+    image = Image.open(BytesIO(response.content))
+    image.load()
+    return image.convert("RGB")
+
+
+def _resize_to_target(image: Image.Image, *, width: int, height: int) -> Image.Image:
+    return image.resize((width, height), Image.Resampling.LANCZOS)
+
+
+def _configure_scheduler(pipe: WanImageToVideoPipeline, *, flow_shift: float) -> None:
+    pipe.scheduler = UniPCMultistepScheduler.from_config(
+        pipe.scheduler.config,
+        flow_shift=flow_shift,
+    )
+
+
+def _write_metadata(
+    path: Path,
+    *,
+    args: argparse.Namespace,
+    width: int,
+    height: int,
+    frame_count: int,
+) -> None:
+    payload = {
+        "model": args.model,
+        "image_source": args.image_source,
+        "size": args.size,
+        "width": width,
+        "height": height,
+        "fps": args.fps,
+        "num_frames": args.num_frames,
+        "actual_frame_count": frame_count,
+        "guidance_scale": args.guidance_scale,
+        "guidance_scale_2": args.guidance_scale_2,
+        "boundary_ratio": BOUNDARY_RATIO,
+        "flow_shift": args.flow_shift,
+        "num_inference_steps": args.num_inference_steps,
+        "seed": args.seed,
+        "world_size": 1,
+    }
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+
+def main() -> int:
+    args = _parse_args()
+    device = _offline_cuda_device()
+    torch.cuda.set_device(device)
+    _ensure_wan_ftfy_fallback()
+
+    pipe = WanImageToVideoPipeline.from_pretrained(args.model, torch_dtype=torch.bfloat16)
+    pipe.register_to_config(boundary_ratio=BOUNDARY_RATIO)
+    _configure_scheduler(pipe, flow_shift=args.flow_shift)
+    pipe.to(device)
+    pipe.set_progress_bar_config(disable=False)
+
+    input_image = _load_input_image(args.image_source)
+    width, height = _parse_size(args.size)
+    resized_image = _resize_to_target(input_image, width=width, height=height)
+
+    generator = torch.Generator(device=device.type).manual_seed(args.seed)
+    frames = pipe(
+        image=resized_image,
+        prompt=args.prompt,
+        negative_prompt=args.negative_prompt,
+        height=height,
+        width=width,
+        num_frames=args.num_frames,
+        guidance_scale=args.guidance_scale,
+        guidance_scale_2=args.guidance_scale_2,
+        num_inference_steps=args.num_inference_steps,
+        generator=generator,
+    ).frames[0]
+
+    output_path = Path(args.output)
+    metadata_path = Path(args.metadata_output)
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    export_to_video(frames, str(output_path), fps=args.fps)
+    _write_metadata(metadata_path, args=args, width=width, height=height, frame_count=len(frames))
+
+    if hasattr(pipe, "maybe_free_model_hooks"):
+        pipe.maybe_free_model_hooks()
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py

@@ -0,0 +1,672 @@
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+import subprocess
+import sys
+import time
+from base64 import b64decode, b64encode
+from fractions import Fraction
+from hashlib import sha1
+from io import BytesIO
+from mimetypes import guess_type
+from pathlib import Path
+from typing import Any
+from urllib.parse import urlparse
+
+import pytest
+import requests
+import torch
+from diffusers import UniPCMultistepScheduler
+from PIL import Image
+
+from tests.conftest import OmniServerParams
+from tests.e2e.accuracy.wan22_i2v.run_wan22_i2v_diffusers_cp import (
+    _configure_scheduler,
+    _ensure_wan_ftfy_fallback,
+    _IdentityFtfy,
+    _offline_cuda_device,
+    _resize_to_target,
+)
+from tests.e2e.accuracy.wan22_i2v.wan22_i2v_video_similarity_common import (
+    FLOW_SHIFT,
+    FPS,
+    GUIDANCE_SCALE,
+    GUIDANCE_SCALE_2,
+    HEIGHT,
+    MODEL_NAME,
+    NEGATIVE_PROMPT,
+    NUM_FRAMES,
+    NUM_INFERENCE_STEPS,
+    PROMPT,
+    PSNR_THRESHOLD,
+    RABBIT_IMAGE_URL,
+    SEED,
+    SIZE,
+    SSIM_THRESHOLD,
+    WIDTH,
+)
+from tests.utils import hardware_test
+
+
+def test_parse_video_metadata_extracts_dimensions_and_fps() -> None:
+    payload = {
+        "streams": [
+            {
+                "width": 832,
+                "height": 480,
+                "avg_frame_rate": "16/1",
+                "nb_read_frames": "93",
+            }
+        ]
+    }
+
+    metadata = _parse_video_metadata(payload)
+
+    assert metadata["width"] == 832
+    assert metadata["height"] == 480
+    assert metadata["fps"] == 16.0
+    assert metadata["frame_count"] == 93
+
+
+def test_parse_ssim_summary_extracts_all_score() -> None:
+    output = """
+    [Parsed_ssim_0 @ 000001] SSIM Y:0.971903 (15.512007) U:0.965077 (14.569044) V:0.962414 (14.252637) All:0.968311 (15.035654)
+    """
+
+    assert _parse_ssim_score(output) == 0.968311
+
+
+def test_parse_psnr_summary_extracts_average_score() -> None:
+    output = """
+    [Parsed_psnr_0 @ 000001] PSNR y:32.670157 u:31.844621 v:31.513839 average:32.148004 min:31.563744 max:33.201457
+    """
+
+    assert _parse_psnr_score(output) == 32.148004
+
+
+def test_build_diffusers_command_uses_python_runner_path(tmp_path: Path) -> None:
+    runner_path = tmp_path / "run_wan22_i2v_diffusers_cp.py"
+    command = _build_diffusers_command(
+        runner_path=runner_path,
+        image_source=RABBIT_IMAGE_URL,
+        output_path=tmp_path / "offline.mp4",
+        metadata_path=tmp_path / "offline.json",
+    )
+
+    assert command[:2] == [
+        sys.executable,
+        str(runner_path),
+    ]
+    assert "--output" in command
+    assert "--metadata-output" in command
+
+
+def test_resolve_image_source_prefers_existing_local_path(tmp_path: Path, monkeypatch: pytest.MonkeyPatch) -> None:
+    image_path = tmp_path / "rabbit.png"
+    Image.new("RGB", (8, 8), color=(255, 128, 64)).save(image_path)
+
+    assert _resolve_image_source(str(image_path)) == str(image_path.resolve())
+
+
+def test_build_online_image_reference_uses_data_url_for_local_path(tmp_path: Path) -> None:
+    image_path = tmp_path / "rabbit.png"
+    Image.new("RGB", (4, 2), color=(10, 20, 30)).save(image_path)
+
+    reference = _build_online_image_reference(str(image_path))
+
+    assert reference.startswith("data:image/png;base64,")
+    encoded = reference.split(",", 1)[1]
+    assert len(b64decode(encoded)) > 0
+
+
+def test_send_video_request_with_timeout_uses_requested_timeout(monkeypatch: pytest.MonkeyPatch) -> None:
+    class DummyResponse:
+        def raise_for_status(self) -> None:
+            return None
+
+        def json(self) -> dict[str, str]:
+            return {"id": "video_123"}
+
+    class DummyClient:
+        def __init__(self) -> None:
+            self.run_level = "core_model"
+            self.wait_args: tuple[str, int, int] | None = None
+
+        def _build_url(self, path: str) -> str:
+            return f"http://localhost:8000{path}"
+
+        def _wait_until_video_completed(
+            self,
+            video_id: str,
+            poll_interval_seconds: int = 2,
+            timeout_seconds: int = 300,
+        ) -> None:
+            self.wait_args = (video_id, poll_interval_seconds, timeout_seconds)
+
+        def _download_video_content(self, video_id: str) -> bytes:
+            return b"video-bytes"
+
+    monkeypatch.setattr(requests, "post", lambda *args, **kwargs: DummyResponse())
+
+    client = DummyClient()
+    result = _send_video_request_with_timeout(
+        client,
+        {
+            "form_data": {"prompt": "test"},
+            "image_reference": RABBIT_IMAGE_URL,
+        },
+        timeout_seconds=1200,
+    )
+
+    assert result == b"video-bytes"
+    assert client.wait_args == ("video_123", 2, 1200)
+
+
+def test_online_timeout_defaults_to_1200() -> None:
+    assert _online_timeout_seconds(None) == 1200
+
+
+def test_artifact_dir_is_under_repo_result_folder(tmp_path: Path) -> None:
+    artifact_dir = _artifact_dir(str(tmp_path / "rabbit.png"))
+
+    assert artifact_dir.parent == Path(__file__).parent / "result"
+    assert artifact_dir.name.startswith("rabbit-")
+
+
+def test_offline_cuda_device_uses_indexed_cuda_device() -> None:
+    assert _offline_cuda_device() == torch.device("cuda:0")
+
+
+def test_ensure_wan_ftfy_fallback_sets_identity(monkeypatch: pytest.MonkeyPatch) -> None:
+    from diffusers.pipelines.wan import pipeline_wan_i2v as wan_i2v_module
+
+    monkeypatch.delattr(wan_i2v_module, "ftfy", raising=False)
+    _ensure_wan_ftfy_fallback()
+
+    assert hasattr(wan_i2v_module, "ftfy")
+    assert isinstance(wan_i2v_module.ftfy, _IdentityFtfy)
+    assert wan_i2v_module.ftfy.fix_text("abc") == "abc"
+
+
+def test_resize_to_target_matches_requested_dimensions() -> None:
+    image = Image.new("RGB", (100, 50), color=(1, 2, 3))
+
+    resized = _resize_to_target(image, width=832, height=480)
+
+    assert resized.size == (832, 480)
+
+
+def test_configure_scheduler_switches_to_unipc() -> None:
+    class DummyPipe:
+        def __init__(self) -> None:
+            self.scheduler = UniPCMultistepScheduler(num_train_timesteps=1000, flow_shift=1.0)
+
+    pipe = DummyPipe()
+    _configure_scheduler(pipe, flow_shift=5.0)
+
+    assert isinstance(pipe.scheduler, UniPCMultistepScheduler)
+    assert pipe.scheduler.config.flow_shift == 5.0
+
+
+REPO_ROOT = Path(__file__).resolve().parents[4]
+WORKSPACE_ROOT = REPO_ROOT.parent
+RUNNER_PATH = Path(__file__).with_name("run_wan22_i2v_diffusers_cp.py")
+RESULT_ROOT = Path(__file__).parent / "result"
+VIDEO_TIMEOUT_SECONDS = 60 * 60
+_SSIM_RE = re.compile(r"All:(?P<score>[0-9.]+)")
+_PSNR_RE = re.compile(r"average:(?P<score>[0-9.]+)")
+SERVER_CASES = [
+    pytest.param(
+        OmniServerParams(
+            model=MODEL_NAME,
+            server_args=[
+                "--usp",
+                "2",
+                "--use-hsdp",
+                "--hsdp-shard-size",
+                "2",
+            ],
+            use_omni=True,
+        ),
+        id="wan22_i2v_usp2_hsdp2",
+    )
+]
+
+
+def _parse_video_metadata(payload: dict[str, Any]) -> dict[str, int | float]:
+    streams = payload.get("streams")
+    if not isinstance(streams, list) or not streams:
+        raise ValueError(f"ffprobe payload did not include video streams: {payload}")
+
+    stream = streams[0]
+    width = int(stream["width"])
+    height = int(stream["height"])
+    fps = float(Fraction(stream["avg_frame_rate"]))
+    frame_count_value = stream.get("nb_read_frames") or stream.get("nb_frames")
+    if frame_count_value is None:
+        raise ValueError(f"ffprobe payload did not include frame count: {payload}")
+    frame_count = int(frame_count_value)
+    return {
+        "width": width,
+        "height": height,
+        "fps": fps,
+        "frame_count": frame_count,
+    }
+
+
+def _parse_ssim_score(output: str) -> float:
+    match = _SSIM_RE.search(output)
+    if match is None:
+        raise ValueError(f"Could not parse SSIM score from ffmpeg output:\n{output}")
+    return float(match.group("score"))
+
+
+def _parse_psnr_score(output: str) -> float:
+    match = _PSNR_RE.search(output)
+    if match is None:
+        raise ValueError(f"Could not parse PSNR score from ffmpeg output:\n{output}")
+    return float(match.group("score"))
+
+
+def _probe_binary(binary: str) -> str:
+    resolved = shutil.which(binary)
+    if resolved is None:
+        pytest.skip(f"{binary} is required for Wan2.2 video similarity e2e test.")
+    return resolved
+
+
+def _build_diffusers_command(
+    *,
+    runner_path: Path,
+    image_source: str,
+    output_path: Path,
+    metadata_path: Path,
+) -> list[str]:
+    return [
+        sys.executable,
+        str(runner_path),
+        "--model",
+        MODEL_NAME,
+        "--image-source",
+        image_source,
+        "--prompt",
+        PROMPT,
+        "--negative-prompt",
+        NEGATIVE_PROMPT,
+        "--size",
+        SIZE,
+        "--fps",
+        str(FPS),
+        "--num-frames",
+        str(NUM_FRAMES),
+        "--guidance-scale",
+        str(GUIDANCE_SCALE),
+        "--guidance-scale-2",
+        str(GUIDANCE_SCALE_2),
+        "--flow-shift",
+        str(FLOW_SHIFT),
+        "--num-inference-steps",
+        str(NUM_INFERENCE_STEPS),
+        "--seed",
+        str(SEED),
+        "--output",
+        str(output_path),
+        "--metadata-output",
+        str(metadata_path),
+    ]
+
+
+def _runner_env() -> dict[str, str]:
+    env = os.environ.copy()
+    pythonpath_parts = [
+        str(REPO_ROOT),
+        str(WORKSPACE_ROOT / "diffusers" / "src"),
+    ]
+    existing_pythonpath = env.get("PYTHONPATH")
+    if existing_pythonpath:
+        pythonpath_parts.append(existing_pythonpath)
+    env["PYTHONPATH"] = os.pathsep.join(pythonpath_parts)
+    env.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
+    return env
+
+
+def _resolve_image_source(configured: str | None) -> str:
+    configured = configured or RABBIT_IMAGE_URL
+    candidate = Path(configured)
+    if candidate.exists():
+        return str(candidate.resolve())
+    return configured
+
+
+def _is_remote_image_source(source: str) -> bool:
+    return source.startswith("http://") or source.startswith("https://")
+
+
+def _validate_image_source(source: str) -> None:
+    if _is_remote_image_source(source):
+        response = requests.get(source, timeout=60)
+        response.raise_for_status()
+        return
+
+    image_path = Path(source)
+    if not image_path.exists():
+        raise FileNotFoundError(f"Local image source does not exist: {image_path}")
+
+
+def _build_online_image_reference(source: str) -> str:
+    if _is_remote_image_source(source) or source.startswith("data:image"):
+        return source
+
+    image_path = Path(source)
+    mime_type = guess_type(image_path.name)[0] or "image/png"
+    encoded = b64encode(image_path.read_bytes()).decode("ascii")
+    return f"data:{mime_type};base64,{encoded}"
+
+
+def _probe_video(path: Path) -> dict[str, int | float]:
+    _probe_binary("ffprobe")
+    result = subprocess.run(
+        [
+            "ffprobe",
+            "-v",
+            "error",
+            "-select_streams",
+            "v:0",
+            "-count_frames",
+            "-show_entries",
+            "stream=width,height,avg_frame_rate,nb_read_frames",
+            "-of",
+            "json",
+            str(path),
+        ],
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    return _parse_video_metadata(json.loads(result.stdout))
+
+
+def _run_ffmpeg_similarity(filter_name: str, first: Path, second: Path) -> str:
+    _probe_binary("ffmpeg")
+    result = subprocess.run(
+        [
+            "ffmpeg",
+            "-hide_banner",
+            "-i",
+            str(first),
+            "-i",
+            str(second),
+            "-lavfi",
+            f"[0:v][1:v]{filter_name}",
+            "-f",
+            "null",
+            "-",
+        ],
+        capture_output=True,
+        text=True,
+        check=True,
+    )
+    return result.stderr
+
+
+def _online_timeout_seconds(configured: int | None) -> int:
+    return int(configured) if configured is not None else 1200
+
+
+def _artifact_dir(image_source: str) -> Path:
+    if _is_remote_image_source(image_source):
+        source_name = Path(urlparse(image_source).path).stem or "remote"
+    else:
+        source_name = Path(image_source).stem or "local"
+
+    safe_name = re.sub(r"[^A-Za-z0-9_.-]+", "_", source_name)
+    digest = sha1(image_source.encode("utf-8")).hexdigest()[:8]
+    return RESULT_ROOT / f"{safe_name}-{digest}"
+
+
+def _artifact_paths(image_source: str) -> tuple[Path, Path, Path]:
+    artifact_dir = _artifact_dir(image_source)
+    artifact_dir.mkdir(parents=True, exist_ok=True)
+    return (
+        artifact_dir / "online.mp4",
+        artifact_dir / "offline.mp4",
+        artifact_dir / "offline_metadata.json",
+    )
+
+
+def _send_video_request_with_timeout(
+    openai_client,
+    request_config: dict[str, Any],
+    *,
+    timeout_seconds: int,
+) -> bytes:
+    form_data = request_config.get("form_data")
+    if not isinstance(form_data, dict):
+        raise ValueError("Video request_config must contain 'form_data'")
+
+    if not form_data.get("prompt"):
+        raise ValueError("Video request_config['form_data'] must contain 'prompt'")
+
+    normalized_form_data = {key: str(value) for key, value in form_data.items() if value is not None}
+
+    files: dict[str, tuple[str, BytesIO, str]] = {}
+    image_reference = request_config.get("image_reference")
+    if image_reference:
+        if image_reference.startswith("data:image"):
+            header, encoded = image_reference.split(",", 1)
+            content_type = header.split(";")[0].removeprefix("data:")
+            extension = content_type.split("/")[-1]
+            file_data = b64decode(encoded)
+            files["input_reference"] = (
+                f"reference.{extension}",
+                BytesIO(file_data),
+                content_type,
+            )
+        else:
+            normalized_form_data["image_reference"] = json.dumps({"image_url": image_reference})
+
+    start_time = time.perf_counter()
+    create_url = openai_client._build_url("/v1/videos")
+    response = requests.post(
+        create_url,
+        data=normalized_form_data,
+        files=files,
+        headers={"Accept": "application/json"},
+        timeout=60,
+    )
+    response.raise_for_status()
+    job_data = response.json()
+    video_id = job_data["id"]
+    openai_client._wait_until_video_completed(video_id, timeout_seconds=timeout_seconds)
+    video_content = openai_client._download_video_content(video_id)
+    print(f"online_video_e2e_latency_s={time.perf_counter() - start_time:.3f}")
+    return video_content
+
+
+def _generate_online_video(
+    *,
+    omni_server,
+    openai_client,
+    image_source: str,
+    online_timeout_seconds: int,
+) -> Path:
+    online_path, _, _ = _artifact_paths(image_source)
+    request_config = {
+        "model": omni_server.model,
+        "form_data": {
+            "prompt": PROMPT,
+            "negative_prompt": NEGATIVE_PROMPT,
+            "size": SIZE,
+            "fps": FPS,
+            "num_frames": NUM_FRAMES,
+            "guidance_scale": GUIDANCE_SCALE,
+            "guidance_scale_2": GUIDANCE_SCALE_2,
+            "flow_shift": FLOW_SHIFT,
+            "num_inference_steps": NUM_INFERENCE_STEPS,
+            "seed": SEED,
+        },
+        "image_reference": _build_online_image_reference(image_source),
+    }
+    online_video_bytes = _send_video_request_with_timeout(
+        openai_client,
+        request_config,
+        timeout_seconds=_online_timeout_seconds(online_timeout_seconds),
+    )
+    online_path.write_bytes(online_video_bytes)
+    return online_path
+
+
+def _generate_offline_video(*, image_source: str) -> tuple[Path, Path]:
+    _, offline_path, offline_metadata_path = _artifact_paths(image_source)
+    command = _build_diffusers_command(
+        runner_path=RUNNER_PATH,
+        image_source=image_source,
+        output_path=offline_path,
+        metadata_path=offline_metadata_path,
+    )
+    subprocess.run(
+        command,
+        cwd=REPO_ROOT,
+        env=_runner_env(),
+        check=True,
+        timeout=VIDEO_TIMEOUT_SECONDS,
+    )
+    return offline_path, offline_metadata_path
+
+
+@pytest.mark.advanced_model
+@pytest.mark.benchmark
+@pytest.mark.diffusion
+@hardware_test(res={"cuda": "H100"}, num_cards=1)
+def test_wan22_i2v_diffusers_offline_generates_video(
+    wan22_i2v_image_source: str | None,
+) -> None:
+    if not torch.cuda.is_available():
+        pytest.skip("Wan2.2 I2V diffusers offline test requires CUDA.")
+
+    _probe_binary("ffprobe")
+    if not RUNNER_PATH.exists():
+        raise AssertionError(f"Offline diffusers runner does not exist: {RUNNER_PATH}")
+
+    image_source = _resolve_image_source(wan22_i2v_image_source)
+    _validate_image_source(image_source)
+    offline_path, offline_metadata_path = _generate_offline_video(image_source=image_source)
+    assert offline_path.exists(), f"Expected offline video artifact at {offline_path}"
+    assert offline_metadata_path.exists(), f"Expected offline metadata artifact at {offline_metadata_path}"
+    offline_metadata = _probe_video(offline_path)
+    assert offline_metadata["width"] == WIDTH
+    assert offline_metadata["height"] == HEIGHT
+    assert offline_metadata["fps"] == float(FPS)
+    assert offline_metadata["frame_count"] == NUM_FRAMES
+
+
+@pytest.mark.advanced_model
+@pytest.mark.benchmark
+@pytest.mark.diffusion
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", SERVER_CASES, indirect=True)
+def test_wan22_i2v_online_serving_generates_video(
+    omni_server,
+    openai_client,
+    wan22_i2v_image_source: str | None,
+    wan22_i2v_online_timeout_seconds: int,
+) -> None:
+    if not torch.cuda.is_available() or torch.cuda.device_count() < 2:
+        pytest.skip("Wan2.2 I2V similarity e2e test requires >= 2 CUDA GPUs.")
+
+    _probe_binary("ffprobe")
+    image_source = _resolve_image_source(wan22_i2v_image_source)
+    _validate_image_source(image_source)
+    online_path = _generate_online_video(
+        omni_server=omni_server,
+        openai_client=openai_client,
+        image_source=image_source,
+        online_timeout_seconds=wan22_i2v_online_timeout_seconds,
+    )
+    assert online_path.exists(), f"Expected online video artifact at {online_path}"
+    online_metadata = _probe_video(online_path)
+    assert online_metadata["width"] == WIDTH
+    assert online_metadata["height"] == HEIGHT
+    assert online_metadata["fps"] == float(FPS)
+    assert online_metadata["frame_count"] == NUM_FRAMES
+
+
+@pytest.mark.advanced_model
+@pytest.mark.benchmark
+@pytest.mark.diffusion
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+def test_wan22_i2v_serving_matches_diffusers_video_similarity(
+    wan22_i2v_image_source: str | None,
+) -> None:
+    if not torch.cuda.is_available() or torch.cuda.device_count() < 2:
+        pytest.skip("Wan2.2 I2V similarity e2e test requires >= 2 CUDA GPUs.")
+
+    _probe_binary("ffmpeg")
+    _probe_binary("ffprobe")
+    if not RUNNER_PATH.exists():
+        raise AssertionError(f"Offline diffusers runner does not exist: {RUNNER_PATH}")
+
+    image_source = _resolve_image_source(wan22_i2v_image_source)
+    _validate_image_source(image_source)
+    online_path, offline_path, offline_metadata_path = _artifact_paths(image_source)
+
+    if not online_path.exists():
+        pytest.skip(f"Missing online artifact from prerequisite test: {online_path}")
+    if not offline_path.exists() or not offline_metadata_path.exists():
+        pytest.skip(f"Missing offline artifacts from prerequisite test: {offline_path}, {offline_metadata_path}")
+
+    assert online_path.exists(), f"Expected online video artifact at {online_path}"
+    assert offline_path.exists(), f"Expected offline video artifact at {offline_path}"
+    assert offline_metadata_path.exists(), f"Expected offline metadata artifact at {offline_metadata_path}"
+
+    online_metadata = _probe_video(online_path)
+    offline_metadata = _probe_video(offline_path)
+    assert online_metadata == offline_metadata, (
+        f"Video metadata mismatch:\n"
+        f"online={online_metadata}\n"
+        f"offline={offline_metadata}\n"
+        f"online_path={online_path}\n"
+        f"offline_path={offline_path}"
+    )
+    assert online_metadata["width"] == WIDTH
+    assert online_metadata["height"] == HEIGHT
+    assert online_metadata["fps"] == float(FPS)
+    assert online_metadata["frame_count"] == NUM_FRAMES
+
+    ssim_output = _run_ffmpeg_similarity("ssim", online_path, offline_path)
+    psnr_output = _run_ffmpeg_similarity("psnr", online_path, offline_path)
+    ssim_score = _parse_ssim_score(ssim_output)
+    psnr_score = _parse_psnr_score(psnr_output)
+
+    print("wan22_i2v_similarity metrics:")
+    print(
+        "  SSIM:"
+        f" value={ssim_score:.6f},"
+        f" threshold>={SSIM_THRESHOLD:.6f},"
+        " range=[0, 1],"
+        " higher_is_better=True,"
+        " interpretation=structural_similarity"
+    )
+    print(
+        "  PSNR:"
+        f" value={psnr_score:.6f} dB,"
+        f" threshold>={PSNR_THRESHOLD:.6f} dB,"
+        " range=[0, +inf),"
+        " higher_is_better=True,"
+        " interpretation=pixel_error_in_decibels"
+    )
+    print(f"online_video={online_path}")
+    print(f"offline_video={offline_path}")
+    print(f"offline_metadata={offline_metadata_path}")
+
+    assert ssim_score >= SSIM_THRESHOLD, (
+        f"SSIM below threshold: got {ssim_score:.6f}, expected >= {SSIM_THRESHOLD:.6f}. "
+        f"online={online_path} offline={offline_path}"
+    )
+    assert psnr_score >= PSNR_THRESHOLD, (
+        f"PSNR below threshold: got {psnr_score:.6f}, expected >= {PSNR_THRESHOLD:.6f}. "
+        f"online={online_path} offline={offline_path}"
+    )

tests/e2e/accuracy/wan22_i2v/wan22_i2v_video_similarity_common.py

@@ -0,0 +1,24 @@
+from __future__ import annotations
+
+RABBIT_IMAGE_URL = "https://vllm-public-assets.s3.us-west-2.amazonaws.com/omni-assets/rabbit.png"
+MODEL_NAME = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
+SIZE = "832x480"
+WIDTH = 832
+HEIGHT = 480
+FPS = 16
+NUM_FRAMES = 93
+GUIDANCE_SCALE = 3.5
+GUIDANCE_SCALE_2 = 3.5
+FLOW_SHIFT = 5.0
+NUM_INFERENCE_STEPS = 50
+SEED = 42
+BOUNDARY_RATIO = 0.875
+SSIM_THRESHOLD = 0.94
+PSNR_THRESHOLD = 28.0
+
+PROMPT = """一只棕色野兔的正面特写镜头，采用低角度仰拍视角，营造亲密而庄严的视觉冲击。兔子一双圆润漆黑的大眼睛直视镜头深处，眼神中交织着野生动物的警觉与一丝难以言喻的温柔好奇，仿佛在与观者建立跨越物种的静默对话。它毛色呈现层次丰富的棕褐渐变，从浅奶油色腹部过渡到深棕背部，每根毛发纹理清晰可辨，在侧光下泛着丝绸般的光泽。细长洁白的胡须共有三对，随呼吸节奏微微颤动，偶尔因捕捉气流信息而轻轻摇摆。
+两只标志性的长耳完全竖立，耳廓外侧覆盖短密棕毛，内侧则露出粉嫩的血管网络，薄如蝉翼的皮肤下血液流动隐约可见，耳朵以细微幅度不时转动，精准定位声源方向。背景是一片澄澈的蔚蓝天空，形态蓬松的白色积云以缓慢速度横向漂移，云影在兔子头顶交替变化，光线随之明暗流转。晴朗天气的明媚阳光从画面左上方45度角倾泻而下，在兔脸右侧形成柔和的伦勃朗式阴影，强化了面部立体感和皮毛质感。
+兔子湿润的黑鼻子持续进行每秒三至四次的快速抽动，这是它们感知化学信号的本能动作，粉色三瓣嘴随之轻启，露出正在反刍的洁白门齿，下颌以稳定节奏左右研磨。摄影采用大光圈浅景深，焦点牢牢锁定在兔子双眼连线所在的焦平面，背景天空和远景绿色植被虚化成圆润的彩色光斑，前景几根嫩绿草叶闯入画面边缘，以缓慢弧线随风摇曳，暗示着和煦的春日微风。
+整个场景弥漫着宁静致远的田园诗意，色彩温暖饱和，充满生命力。兔子在持续五秒的对视后，以典型 lagomorph 特征完成一次完整的瞬膜眨眼——第三眼睑从内侧横向滑过眼球，继而缓缓歪头向右十五度，这个行为在动物行为学中代表认知加工和好奇心表达，耳朵随之向同一方向倾斜，最终恢复正视姿态，胡须舒展，完成这段短暂而珍贵的自然纪录。"""
+
+NEGATIVE_PROMPT = """色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"""

tests/e2e/online_serving/test_qwen3_omni_expansion.py

@@ -23,7 +23,7 @@ from tests.conftest import (
 )
 from tests.utils import hardware_test
 
-models = ["Qwen/Qwen3-Omni-30B-A3B-Instruct"]
+model = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
 
 AUDIO_KEY = ["test"]
 IMAGE_KEY = ["square", "quadrate"]
@@ -49,16 +49,32 @@ def get_chunk_config(default_path):
     return path
 
 
+def get_batch_token_config(default_path):
+    path = modify_stage_config(
+        default_path,
+        updates={
+            "stage_args": {1: {"engine_args.max_num_batched_tokens": 64}},
+        },
+    )
+    return path
+
+
 # CI stage config for 2*H100-80G GPUs
 default_path = str(Path(__file__).parent.parent / "stage_configs" / "qwen3_omni_ci.yaml")
-stage_configs = [default_path, get_chunk_config(default_path)]
 
 if current_omni_platform.is_xpu():
-    stage_configs = [str(Path(__file__).parent.parent / "stage_configs" / "xpu" / "qwen3_omni_ci.yaml")]
+    default_path = str(Path(__file__).parent.parent / "stage_configs" / "xpu" / "qwen3_omni_ci.yaml")
 
 # Create parameter combinations for model and stage config
 test_params = [
-    OmniServerParams(model=model, stage_config_path=stage_config) for model in models for stage_config in stage_configs
+    pytest.param(OmniServerParams(model=model, stage_config_path=default_path), id="default"),
+    pytest.param(OmniServerParams(model=model, stage_config_path=get_chunk_config(default_path)), id="async_chunk"),
+]
+
+test_token_params = [
+    pytest.param(
+        OmniServerParams(model=model, stage_config_path=get_batch_token_config(default_path)), id="batch_token_64"
+    )
 ]
 
 
@@ -85,6 +101,8 @@ def get_prompt(prompt_type="text_only"):
         "text_video": "What is in this video? ",
         "text_image": "What is in this image? ",
         "text_audio": "What is in this audio? ",
+        "text_audio_video": "First, what is in this audio? Then, what is in this video? ",
+        "one_word": "What is the capital of UK? Answer in one word",
     }
     return prompts.get(prompt_type, prompts["text_only"])
 
@@ -121,7 +139,7 @@ def test_text_to_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("omni_server", test_params + test_token_params, indirect=True)
 def test_text_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Input Modal: text
@@ -288,7 +306,7 @@ def test_video_to_text_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("omni_server", test_params + test_token_params, indirect=True)
 def test_text_audio_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Input Modal: text, audio
@@ -313,7 +331,7 @@ def test_text_audio_to_text_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("omni_server", test_params + test_token_params, indirect=True)
 def test_text_image_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Input Modal: text, image
@@ -339,7 +357,7 @@ def test_text_image_to_text_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("omni_server", test_params + test_token_params, indirect=True)
 def test_text_video_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Input Modal: text, video
@@ -367,7 +385,7 @@ def test_text_video_to_text_audio_001(omni_server, openai_client) -> None:
 @pytest.mark.advanced_model
 @pytest.mark.omni
 @hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
-@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+@pytest.mark.parametrize("omni_server", test_params + test_token_params, indirect=True)
 def test_mix_to_text_audio_001(omni_server, openai_client) -> None:
     """
     Input Modal: text, audio, image, video
@@ -393,3 +411,106 @@ def test_mix_to_text_audio_001(omni_server, openai_client) -> None:
         "key_words": {"audio": AUDIO_KEY, "image": IMAGE_KEY, "video": VIDEO_KEY},
     }
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_audio_in_video_001(omni_server, openai_client) -> None:
+    """
+    Input Modal: text + video (synthetic MP4 with embedded audio; ``use_audio_in_video`` uses audio from the video).
+    Output Modal: text, audio
+    Input Setting: stream=False
+    Datasets: single request
+    """
+    video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(224, 224, 300, embed_audio=True)['base64']}"
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        video_data_url=video_data_url,
+        content_text=get_prompt("text_audio_video"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": False,
+        "use_audio_in_video": True,
+        "key_words": {"video": VIDEO_KEY, "audio": AUDIO_KEY + ["beep", "electronic"]},
+    }
+    openai_client.send_omni_request(request_config)
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_audio_in_video_002(omni_server, openai_client) -> None:
+    """
+    Input Modal: text + video (synthetic MP4 with embedded audio; ``use_audio_in_video`` uses audio from the video).
+    Output Modal: text, audio
+    Input Setting: stream=True
+    Datasets: few requests
+    """
+    video_data_url = f"data:video/mp4;base64,{generate_synthetic_video(224, 224, 300, embed_audio=True)['base64']}"
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        video_data_url=video_data_url,
+        content_text=get_prompt("text_audio_video"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "use_audio_in_video": True,
+        "key_words": {"video": VIDEO_KEY, "audio": AUDIO_KEY + ["beep", "electronic"]},
+    }
+
+    # Retry when assert_omni_response fails on key_words (see tests/conftest.py).
+    _keyword_assert_msg = "The output does not contain any of the keywords."
+    _max_retries = 3
+    for attempt in range(_max_retries):
+        try:
+            openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
+            break
+        except AssertionError as e:
+            if _keyword_assert_msg not in str(e) or attempt == _max_retries - 1:
+                raise
+            print(f"Keyword assertion failed, retrying {attempt + 2}/{_max_retries}: {e!r}")
+
+
+@pytest.mark.advanced_model
+@pytest.mark.omni
+@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", test_params, indirect=True)
+def test_one_word_prompt_001(omni_server, openai_client) -> None:
+    """
+    Input Modal: text only (one-word answer constraint).
+    Output Modal: text, audio (default ``modalities``); ``key_words`` only assert on text.
+    Input Setting: stream=True
+    Datasets: single request
+    """
+    messages = dummy_messages_from_mix_data(
+        system_prompt=get_system_prompt(),
+        content_text=get_prompt("one_word"),
+    )
+
+    request_config = {
+        "model": omni_server.model,
+        "messages": messages,
+        "stream": True,
+        "key_words": {"text": ["london"]},
+    }
+
+    # Retry only when assert_omni_response fails on text/audio cosine similarity (see tests/conftest.py).
+    _similarity_assert_msg = "The audio content is not same as the text"
+    _max_retries = 3
+    for attempt in range(_max_retries):
+        try:
+            openai_client.send_omni_request(request_config, request_num=get_max_batch_size())
+            break
+        except AssertionError as e:
+            if _similarity_assert_msg not in str(e) or attempt == _max_retries - 1:
+                raise
+            print(f"Similarity assertion failed, retrying {attempt + 2}/{_max_retries}: {e!r}")

tests/e2e/online_serving/test_qwen_image_layered_expansion.py

@@ -7,7 +7,7 @@ and are supported by Qwen-Image-Layered model.
 Kept cases (maximum feature coverage, no cpu-offload):
   sp_001              : cache_dit + Ulysses-SP 2          (2×H100)
   cfg_parallel_002    : cache_dit + CFG-Parallel 2        (2×H100)
-  layers_guard_001    : layerwise offload + layers=2      (1×H100, issue #1969 guard)
+  layers_guard_001    : layerwise offload + layers=3      (1×H100, issue #1969 guard)
 
 Total distinct features covered: cache_dit, Ulysses-SP, CFG-Parallel, layerwise-offload.
 """
@@ -102,8 +102,8 @@ LAYERS_GUARD_CASES = [
             model=MODEL,
             server_args=["--enable-layerwise-offload"],
         ),
-        2,
-        id="layers_guard_001_layers2",
+        3,
+        id="layers_guard_001_layers3",
         marks=SINGLE_CARD_FEATURE_MARKS,
     ),
 ]

tests/e2e/stage_configs/qwen3_omni_ci.yaml

@@ -33,7 +33,7 @@ stage_args:
     temperature: 0.4
     top_p: 0.9
     top_k: 1
-    max_tokens: 100
+    max_tokens: 150
     seed: 42
     ignore_eos: False
     detokenize: True
@@ -74,7 +74,7 @@ stage_args:
     devices: "1"
   engine_args:
     model_stage: code2wav
-    max_num_seqs: 1
+    max_num_seqs: 5
     model_arch: Qwen3OmniMoeForConditionalGeneration
     worker_type: generation
     scheduler_cls: vllm_omni.core.sched.omni_generation_scheduler.OmniGenerationScheduler

tests/model_executor/stage_input_processors/test_qwen3_tts_async_chunk.py

@@ -16,6 +16,8 @@ from vllm_omni.model_executor.stage_input_processors.qwen3_tts import (
     talker2code2wav_async_chunk,
 )
 
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
 _FRAME = [1, 2, 3, 4]
 _Q = len(_FRAME)
 
@@ -90,7 +92,7 @@ def test_flush_on_finish():
         is_finished=True,
     )
     assert p is not None
-    assert p["finished"].item() is True
+    assert p["finished"] is True
     assert len(p["code_predictor_codes"]) == _Q * 24
 
 
@@ -158,24 +160,29 @@ def test_dynamic_ic_adapts_to_load():
 
 
 def test_ic_load_change_mid_request():
-    """IC stateless: load spike mid-request shifts initial_coverage."""
+    """IC is cached per request; a load spike only affects new requests."""
     tm = _tm(chunk_frames=25, left_context=25, max_num_seqs=8)
 
-    # Low load -> IC=2 -> emit at frame 2
+    # Low load -> IC=2 (cached for "r"), emit at frame 2
     p1 = _call(tm, "r", n_frames=2)
     assert p1 is not None
 
-    # Spike load: 6 others -> IC=16 -> initial_coverage=16
+    # Spike load: 6 others running
     for i in range(6):
         tm.code_prompt_token_ids[f"other-{i}"] = [[0]] * 10
 
-    # adjusted=25-16=9, 9%25!=0 -> hold
+    # IC for "r" is still cached as 2.
+    # initial_coverage = ((25-1)//2)*2 = 24, first normal emit at 24+25=49
     assert _call(tm, "r", n_frames=25) is None
-
-    # First normal emit at 16+25=41
-    p3 = _call(tm, "r", n_frames=41)
+    assert _call(tm, "r", n_frames=27) is None
+    p3 = _call(tm, "r", n_frames=49)
     assert p3 is not None
-    assert p3["left_context_size"] == 16
+
+    # A *new* request under high load gets IC=16 (not IC=2).
+    # Frame 2 would emit under IC=2 but must hold under IC=16.
+    assert _call(tm, "new_req", n_frames=2) is None
+    p4 = _call(tm, "new_req", n_frames=16)
+    assert p4 is not None
 
 
 @pytest.mark.parametrize(
@@ -295,8 +302,13 @@ def test_non_async_processor_prepends_ref_code_and_sets_trim_context():
         ],
         dtype=torch.long,
     )
-    output = SimpleNamespace(multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code})
-    stage = SimpleNamespace(engine_outputs=[SimpleNamespace(outputs=[output])])
+    output = SimpleNamespace(
+        multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code},
+        token_ids=list(range(3)),
+    )
+    stage = SimpleNamespace(
+        engine_outputs=[SimpleNamespace(outputs=[output], finished=True)],
+    )
 
     prompts = talker2code2wav(stage_list=[stage], engine_input_source=[0])
 
@@ -335,8 +347,13 @@ def test_non_async_processor_filters_out_of_range_codec_values():
         ],
         dtype=torch.long,
     )
-    output = SimpleNamespace(multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code})
-    stage = SimpleNamespace(engine_outputs=[SimpleNamespace(outputs=[output])])
+    output = SimpleNamespace(
+        multimodal_output={"audio_codes": audio_codes, "ref_code": ref_code},
+        token_ids=list(range(4)),
+    )
+    stage = SimpleNamespace(
+        engine_outputs=[SimpleNamespace(outputs=[output], finished=True)],
+    )
 
     prompts = talker2code2wav(stage_list=[stage], engine_input_source=[0])
 

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py

@@ -1122,7 +1122,7 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
         )
         # Prefer GPU for encoder if available; otherwise keep CPU.
         dev = next(self.parameters()).device
-        if getattr(dev, "type", None) == "cuda":
+        if dev.type != "cpu":
             try:
                 tok.model.to(dev)
                 tok.device = dev
@@ -1450,10 +1450,10 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
             )
             if not speaker:
                 raise ValueError("CustomVoice requires additional_information.speaker.")
-            spk_id_map = getattr(self.talker_config, "spk_id", None) or {}
+            spk_id_map = {k.lower(): v for k, v in (getattr(self.talker_config, "spk_id", None) or {}).items()}
             if speaker not in spk_id_map:
                 raise ValueError(f"Unsupported speaker: {speaker}")
-            spk_id = spk_id_map[speaker.lower()]
+            spk_id = spk_id_map[speaker]
             # Keep it at least 1D; embedding on a 0-d tensor can return 1D.
             spk_tensor = torch.tensor([spk_id], device=input_ids.device, dtype=torch.long)
             spk_embed = self.embed_input_ids(spk_tensor)