[Model][VLM] Add LLaVA-Onevision model support (#8486)

Co-authored-by: litianjian <litianjian@bytedance.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Roger Wang <ywang@roblox.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/siglip.py

@@ -4,6 +4,7 @@ within a vision language model."""
 import math
 from typing import Iterable, List, Optional, Tuple, Union
 
+import numpy as np
 import torch
 from PIL import Image
 from torch import nn
@@ -89,6 +90,24 @@ def dummy_image_for_siglip(
     return {"image": image if num_images == 1 else [image] * num_images}
 
 
+def dummy_video_for_siglip(
+    hf_config: SiglipVisionConfig,
+    num_frames: int,
+    *,
+    image_width_override: Optional[int] = None,
+    image_height_override: Optional[int] = None,
+):
+    pil_frame = dummy_image_for_siglip(
+        hf_config,
+        num_images=1,
+        image_width_override=image_width_override,
+        image_height_override=image_height_override)
+    np_frame = np.array(pil_frame["image"])
+    mm_data_per_video = np.repeat([np_frame], num_frames, axis=0)
+    mm_data = {"video": mm_data_per_video}
+    return mm_data
+
+
 def input_processor_for_siglip(
     model_config: ModelConfig,
     hf_config: SiglipVisionConfig,