Import torchcodec lazily in processing_utils (#50325)

The module-level import loaded torchcodec (and its FFmpeg libs) for every
processor, and could break importing processing_utils when torchcodec is
installed but not importable. Import it only in the audio-from-video path,
after the version check, which now raises ImportError.

Author: eustlb

src/transformers/processing_utils.py

@@ -57,7 +57,6 @@
     direct_transformers_import,
     hf_api,
     is_torch_available,
-    is_torchcodec_available,
     is_torchcodec_greater_or_equal,
     list_repo_templates,
     logging,
@@ -78,9 +77,6 @@
 from .video_utils import VideoInput, VideoMetadataType, make_batched_videos
 
 
-if is_torchcodec_available():
-    import torchcodec
-
 if is_torch_available():
     import torch
 
@@ -2571,7 +2567,10 @@ def _resolve_audio_from_video(
             )
 
         if not is_torchcodec_greater_or_equal("0.17"):
-            raise ValueError("Loading audio track from video file requires torchcodec>=0.17 installed!")
+            raise ImportError(
+                "`load_audio_from_video=True` requires torchcodec>=0.17. Install it with `pip install -U torchcodec`."
+            )
+        from torchcodec.decoders import get_container_metadata
 
         standalone_audio = iter(audio)
         audio_resolved = []
@@ -2584,7 +2583,7 @@ def _resolve_audio_from_video(
 
             # Only `torchcodec` can decode audio track from all types of video file formats
             # Don't raise a load-error if video is silent and instead skip it
-            cont_metadata = torchcodec.decoders.get_container_metadata(videos[video_idx])
+            cont_metadata = get_container_metadata(videos[video_idx])
             if not any(stream.media_type == "audio" for stream in cont_metadata.streams):
                 continue