refactor: move video processing configuration into nested processor_kwargs for improved parameter organization

app.py

@@ -322,11 +322,16 @@ def predict(
             return_dict=True,
             return_tensors="pt",
             enable_thinking=thinking_mode,
-            downsample_mode="16x",
-            max_num_frames=max_frames,
-            stack_frames=1,
-            max_slice_nums=1 if any(it.get("type") == "video" for msg in messages for it in msg["content"]) else 9,
-            use_image_id=False if any(it.get("type") == "video" for msg in messages for it in msg["content"]) else True
+            processor_kwargs={
+                "downsample_mode": "16x",
+                "max_slice_nums": 1 if any(it.get("type") == "video" for msg in messages for it in msg["content"]) else 9,
+                "use_image_id": False if any(it.get("type") == "video" for msg in messages for it in msg["content"]) else True,
+                "videos_kwargs": {
+                    "max_num_frames": max_frames,
+                    "do_sample_frames": False, # Frames are already sampled by load_video
+                    "stack_frames": 1,
+                }
+            }
         ).to(model.device)
     
     for k, v in inputs.items():