fix(aero_realtime): disable torchvision fallback to prevent NCCL timeout

Temporarily remove torchvision from VIDEO_READER_BACKENDS during
fetch_video so that decord failures raise immediately instead of
falling back to torchvision which can hang 30+ min on long videos.

Author: kcz358

src/lmms_engine/datasets/iterable/aero_realtime_iterable_dataset.py

@@ -145,10 +145,14 @@ def _load_video_with_metadata(
         Uses qwen_vl_utils fetch_video with return_video_metadata=True to get
         frame indices and fps needed for timestamp computation.
 
+        Disables torchvision fallback to avoid 30-min hangs on corrupted
+        videos that would cause NCCL timeouts.
+
         Returns:
             Tuple of (frames, video_metadata, sample_fps).
         """
         from qwen_vl_utils import fetch_video
+        from qwen_vl_utils import vision_process as _vp
 
         if data_folder is not None:
             full_path = os.path.join(data_folder, video_path)
@@ -174,11 +178,19 @@ def _load_video_with_metadata(
                 f"Invalid video sampling strategy: {self.config.video_sampling_strategy}"
             )
 
-        video_inputs, sample_fps = fetch_video(
-            video_dict,
-            return_video_sample_fps=True,
-            return_video_metadata=True,
-        )
+        # Temporarily remove torchvision from backends to prevent slow fallback
+        # that can hang for 30+ minutes on problematic videos, causing NCCL timeout.
+        _tv_backup = _vp.VIDEO_READER_BACKENDS.pop("torchvision", None)
+        try:
+            video_inputs, sample_fps = fetch_video(
+                video_dict,
+                return_video_sample_fps=True,
+                return_video_metadata=True,
+            )
+        finally:
+            if _tv_backup is not None:
+                _vp.VIDEO_READER_BACKENDS["torchvision"] = _tv_backup
+
         frames, video_metadata = video_inputs
         frames = frames.numpy()
         return frames, video_metadata, sample_fps