Qwen3vl qwen3omni

docs/models/supported_models.md

@@ -562,7 +562,10 @@ Specified using `--task generate`.
 | `Qwen2AudioForConditionalGeneration`         | Qwen2-Audio                                                              | T + A<sup>+</sup>                                                     | `Qwen/Qwen2-Audio-7B-Instruct`                                                                                                                          |                       | ✅︎                          | ✅︎                      |
 | `Qwen2VLForConditionalGeneration`            | QVQ, Qwen2-VL                                                            | T + I<sup>E+</sup> + V<sup>E+</sup>                                   | `Qwen/QVQ-72B-Preview`, `Qwen/Qwen2-VL-7B-Instruct`, `Qwen/Qwen2-VL-72B-Instruct`, etc.                                                                 | ✅︎                     | ✅︎                          | ✅︎                    |
 | `Qwen2_5_VLForConditionalGeneration`         | Qwen2.5-VL                                                               | T + I<sup>E+</sup> + V<sup>E+</sup>                                   | `Qwen/Qwen2.5-VL-3B-Instruct`, `Qwen/Qwen2.5-VL-72B-Instruct`, etc.                                                                                     | ✅︎                     | ✅︎                          | ✅︎                    |
-| `Qwen2_5OmniThinkerForConditionalGeneration` | Qwen2.5-Omni                                                             | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup>                   | `Qwen/Qwen2.5-Omni-7B`                                                                                                                                  |                      | ✅︎                         | ✅︎\*                     |
+| `Qwen2_5OmniThinkerForConditionalGeneration` | Qwen2.5-Omni                                                             | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup>                   | `Qwen/Qwen2.5-Omni-7B`
+| `Qwen3VLForConditionalGeneration` | Qwen3-VL | T + I<sup>E+</sup> + V<sup>E+</sup> | `Qwen/Qwen3-VL-4B-Instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |
+| `Qwen3VLMoeForConditionalGeneration` | Qwen3-VL-MOE | T + I<sup>E+</sup> + V<sup>E+</sup> | `Qwen/Qwen3-VL-30B-A3B-Instruct`, etc. | ✅︎ | ✅︎ | ✅︎ |                                                                                                                                 |                      | ✅︎                         | ✅︎\*                     |
+| `Qwen3OmniMoeThinkerForConditionalGeneration` | Qwen3-Omni | T + I<sup>E+</sup> + V<sup>E+</sup> + A<sup>+</sup> | `Qwen/Qwen3-Omni-30B-A3B-Instruct`, `Qwen/Qwen3-Omni-30B-A3B-Thinking` | ✅︎ | ✅︎ | ✅︎ |
 | `SkyworkR1VChatModel`                        | Skywork-R1V-38B                                                          | T + I                                                                 | `Skywork/Skywork-R1V-38B`                                                                                                                               |                       | ✅︎                          | ✅︎                      |
 | `SmolVLMForConditionalGeneration`            | SmolVLM2                                                                 | T + I                                                                 | `SmolVLM2-2.2B-Instruct`                                                                                                                                | ✅︎                     |                           | ✅︎                      |
 | `TarsierForConditionalGeneration`                | Tarsier                                                                  | T + I<sup>E+</sup>                                                                                                     | `omni-search/Tarsier-7b`,`omni-search/Tarsier-34b`                                                                                                           |                      | ✅︎                      | ✅︎                   |

examples/offline_inference/vision_language.py

@@ -1076,6 +1076,116 @@ def run_qwen2_5_omni(questions: list[str], modality: str):
     )
 
 
+# Qwen3-VL-Dense
+def run_qwen3_vl(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "Qwen/Qwen3-VL-4B-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        max_num_seqs=5,
+        mm_processor_kwargs={
+            "min_pixels": 28 * 28,
+            "max_pixels": 1280 * 28 * 28,
+            "fps": 1,
+        },
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    if modality == "image":
+        placeholder = "<|image_pad|>"
+    elif modality == "video":
+        placeholder = "<|video_pad|>"
+
+    prompts = [
+        (
+            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
+            f"{question}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
+# Qwen3-VL-MOE
+def run_qwen3_vl_moe(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "Qwen/Qwen3-VL-30B-A3B-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        max_num_seqs=5,
+        mm_processor_kwargs={
+            "min_pixels": 28 * 28,
+            "max_pixels": 1280 * 28 * 28,
+            "fps": 1,
+        },
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    if modality == "image":
+        placeholder = "<|image_pad|>"
+    elif modality == "video":
+        placeholder = "<|video_pad|>"
+
+    prompts = [
+        (
+            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
+            f"{question}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
+def run_qwen3_omni_moe(questions: list[str], modality: str) -> ModelRequestData:
+    model_name = "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=4096,
+        max_num_seqs=5,
+        mm_processor_kwargs={
+            "min_pixels": 28 * 28,
+            "max_pixels": 1280 * 28 * 28,
+            "fps": 1,
+        },
+        limit_mm_per_prompt={modality: 1},
+    )
+
+    if modality == "image":
+        placeholder = "<|image_pad|>"
+    elif modality == "video":
+        placeholder = "<|video_pad|>"
+
+    prompts = [
+        (
+            "<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
+            f"<|im_start|>user\n<|vision_start|>{placeholder}<|vision_end|>"
+            f"{question}<|im_end|>\n"
+            "<|im_start|>assistant\n"
+        )
+        for question in questions
+    ]
+
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
+
+
 # SkyworkR1V
 def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
     assert modality == "image"
@@ -1146,11 +1256,20 @@ def run_skyworkr1v(questions: list[str], modality: str) -> ModelRequestData:
     "qwen2_vl": run_qwen2_vl,
     "qwen2_5_vl": run_qwen2_5_vl,
     "qwen2_5_omni": run_qwen2_5_omni,
+    "qwen3_vl": run_qwen3_vl,
+    "qwen3_vl_moe": run_qwen3_vl_moe,
+    "qwen3_omni_moe": run_qwen3_omni_moe,
     "skywork_chat": run_skyworkr1v,
     "smolvlm": run_smolvlm,
     "tarsier": run_tarsier,
 }
 
+MODELS_NEED_VIDEO_METADATA = [
+    "glm4_1v",
+    "qwen3_vl",
+    "qwen3_vl_moe",
+]
+
 
 def get_multi_modal_input(args):
     """
@@ -1176,12 +1295,13 @@ def get_multi_modal_input(args):
 
     if args.modality == "video":
         # Input video and question
+        needs_metadata = args.model_type in MODELS_NEED_VIDEO_METADATA
         video = VideoAsset(name="baby_reading", num_frames=args.num_frames).np_ndarrays
         metadata = VideoAsset(name="baby_reading", num_frames=args.num_frames).metadata
         vid_questions = ["Why is this video funny?"]
 
         return {
-            "data": [(video, metadata)] if args.model_type == "glm4_1v" else video,
+            "data": [(video, metadata)] if needs_metadata else video,
             "questions": vid_questions,
         }
 

tests/models/multimodal/processing/test_common.py

@@ -30,6 +30,7 @@ def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
     """
     # Ensure video metadata is included
     if "video" in mm_data:
+        # GLM4.1V doesn't support multiple videos
         video = mm_data["video"]
         mm_data["video"] = (video, {
             "total_num_frames": len(video),
@@ -39,6 +40,33 @@ def glm4_1v_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
         })
     return mm_data
 
+def qwen3_vl_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
+    """
+    Patch the multimodal data for Qwen3-VL model.
+    """
+
+    def create_metadata(frames: np.ndarray):
+        num_frames = len(frames)
+        return {
+            "total_num_frames": num_frames,
+            "fps": 2.0,
+            "duration": num_frames / 2.0,
+            "video_backend": "opencv",
+            "frames_indices": list(range(num_frames)),
+            "do_sample_frames": True,
+        }
+
+    # Ensure video metadata is included
+    if "video" in mm_data:
+        video = mm_data["video"]
+        if isinstance(video, list):
+            # multiple videos
+            mm_data["video"] = [(vid, create_metadata(vid)) for vid in video]
+        else:
+            # single video
+            mm_data["video"] = (video, create_metadata(video))
+    return mm_data
+
 
 def _test_processing_correctness(
     model_id: str,
@@ -171,8 +199,10 @@ def _test_processing_correctness(
 }
 
 MM_DATA_PATCHES = {
-    # GLM4.1V requires video metadata to be included in the input
+    # GLM4.1V and Qwen3-VL requires video metadata to be included in the input
     "glm4v": glm4_1v_patch_mm_data,
+    "qwen3_vl": qwen3_vl_patch_mm_data,
+    "qwen3_vl_moe": qwen3_vl_patch_mm_data,
 }
 
 
@@ -304,6 +334,8 @@ def _test_processing_correctness_one(
     "Qwen/Qwen2.5-VL-3B-Instruct",
     "Qwen/Qwen2-Audio-7B-Instruct",
     "Qwen/Qwen2.5-Omni-3B",
+    "Qwen/Qwen3-VL-4B-Instruct",
+    "Qwen/Qwen3-VL-30B-A3B-Instruct",
     "Skywork/Skywork-R1V-38B",
     "fixie-ai/ultravox-v0_5-llama-3_2-1b",
     "openai/whisper-large-v3",

tests/models/registry.py

@@ -404,6 +404,15 @@ def check_available_online(
     "Qwen2_5_VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-VL-3B-Instruct"),  # noqa: E501
     "Qwen2_5OmniModel": _HfExamplesInfo("Qwen/Qwen2.5-Omni-3B"),
     "Qwen2_5OmniForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen2.5-Omni-7B-AWQ"),  # noqa: E501
+    "Qwen3VLForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-4B-Instruct", # noqa: E501
+                                                        max_model_len=4096,
+                                                        min_transformers_version="4.57"),  # noqa: E501
+    "Qwen3VLMoeForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-VL-30B-A3B-Instruct", # noqa: E501
+                                                        max_model_len=4096,
+                                                        min_transformers_version="4.57"),
+    "Qwen3OmniMoeThinkerForConditionalGeneration": _HfExamplesInfo("Qwen/Qwen3-Omni-30B-A3B-Instruct", # noqa: E501
+                                                          max_model_len=4096,
+                                                          min_transformers_version="4.57"),
     "SkyworkR1VChatModel": _HfExamplesInfo("Skywork/Skywork-R1V-38B"),
     "SmolVLMForConditionalGeneration": _HfExamplesInfo("HuggingFaceTB/SmolVLM2-2.2B-Instruct"),  # noqa: E501
     "UltravoxModel": _HfExamplesInfo("fixie-ai/ultravox-v0_5-llama-3_2-1b",  # noqa: E501

vllm/assets/video.py

@@ -77,7 +77,7 @@ def video_to_pil_images_list(path: str,
     ]
 
 
-def video_get_metadata(path: str) -> dict[str, Any]:
+def video_get_metadata(path: str, num_frames: int = -1) -> dict[str, Any]:
     cap = cv2.VideoCapture(path)
     if not cap.isOpened():
         raise ValueError(f"Could not open video file {path}")
@@ -86,11 +86,18 @@ def video_get_metadata(path: str) -> dict[str, Any]:
     fps = cap.get(cv2.CAP_PROP_FPS)
     duration = total_frames / fps if fps > 0 else 0
 
+    if num_frames == -1 or num_frames > total_frames:
+        num_frames = total_frames
+
     metadata = {
         "total_num_frames": total_frames,
         "fps": fps,
         "duration": duration,
-        "video_backend": "opencv"
+        "video_backend": "opencv",
+        "frames_indices": list(range(num_frames)),
+        # extra field used to control hf processor's video
+        # sampling behavior
+        "do_sample_frames": num_frames == total_frames,
     }
     return metadata
 
@@ -126,7 +133,7 @@ def np_ndarrays(self) -> npt.NDArray:
     @property
     def metadata(self) -> dict[str, Any]:
         video_path = download_video_asset(self.filename)
-        ret = video_get_metadata(video_path)
+        ret = video_get_metadata(video_path, self.num_frames)
         return ret
 
     def get_audio(self, sampling_rate: Optional[float] = None) -> npt.NDArray:

vllm/attention/layer.py

@@ -24,6 +24,14 @@
 from vllm.v1.attention.backends.utils import validate_kv_sharing_target
 
 
+def check_upstream_fa_availability(dtype: torch.dtype):
+    if dtype in (torch.float16, torch.bfloat16) and current_platform.is_cuda(
+    ) and current_platform.has_device_capability(80):
+        from transformers.utils import is_flash_attn_2_available
+        return is_flash_attn_2_available()
+    return False
+
+
 class Attention(nn.Module):
     """Attention layer.
 

vllm/entrypoints/chat_utils.py

@@ -537,6 +537,8 @@ def _placeholder_str(self, modality: ModalityStr,
                 return "<|vision_start|><|image_pad|><|vision_end|>"
             if model_type == "qwen2_5_omni":
                 return "<|vision_start|><|IMAGE|><|vision_end|>"
+            if model_type in ("qwen3_omni_moe", "qwen3_vl_moe", "qwen3_vl"):
+                return "<|vision_start|><|image_pad|><|vision_end|>"
             if model_type == "molmo":
                 return ""
             if model_type == "aria":
@@ -555,6 +557,8 @@ def _placeholder_str(self, modality: ModalityStr,
             if model_type in ("qwen2_audio", "qwen2_5_omni"):
                 return (f"Audio {current_count}: "
                         f"<|audio_bos|><|AUDIO|><|audio_eos|>")
+            if model_type == "qwen3_omni_moe":
+                return "<|audio_start|><|audio_pad|><|audio_end|>"
             if model_type == "minicpmo":
                 return "(<audio>./</audio>)"
             raise TypeError(f"Unknown model type: {model_type}")
@@ -567,6 +571,8 @@ def _placeholder_str(self, modality: ModalityStr,
                 return "<|vision_start|><|video_pad|><|vision_end|>"
             if model_type == "qwen2_5_omni":
                 return "<|vision_start|><|VIDEO|><|vision_end|>"
+            if model_type in ("qwen3_omni_moe", "qwen3_vl_moe", "qwen3_vl"):
+                return "<|vision_start|><|video_pad|><|vision_end|>"
             if model_type in ("minicpmo", "minicpmv"):
                 return "(<video>./</video>)"
             if model_type.startswith("llava"):