Fix video processing channel format (#41603)

fix

Author: zucchini-nlp

src/transformers/video_processing_utils.py

@@ -57,12 +57,12 @@
     VideoInput,
     VideoMetadata,
     group_videos_by_shape,
+    infer_channel_dimension_format,
     is_valid_video,
     load_video,
     make_batched_metadata,
     make_batched_videos,
     reorder_videos,
-    to_channel_dimension_format,
 )
 
 
@@ -338,10 +338,16 @@ def _prepare_input_videos(
         for video in videos:
             # `make_batched_videos` always returns a 4D array per video
             if isinstance(video, np.ndarray):
-                video = to_channel_dimension_format(video, ChannelDimension.FIRST, input_data_format)
                 # not using F.to_tensor as it doesn't handle (C, H, W) numpy arrays
                 video = torch.from_numpy(video).contiguous()
 
+            # Infer the channel dimension format if not provided
+            if input_data_format is None:
+                input_data_format = infer_channel_dimension_format(video)
+
+            if input_data_format == ChannelDimension.LAST:
+                video = video.permute(0, 3, 1, 2).contiguous()
+
             if device is not None:
                 video = video.to(device)