import torch before decord to fix random_device could not be read

setup.py

@@ -187,7 +187,7 @@
     "transformers>=4.42.0",  # Pins numpy < 2
     "zstandard",
     "polars[timezone]>=0.20.0",
-    # "decord==0.6.0",
+    "decord==0.6.0",
 ]
 
 

src/datasets/features/video.py

@@ -78,6 +78,11 @@ def encode_example(self, value: Union[str, bytes, dict, np.ndarray, "VideoReader
             `dict` with "path" and "bytes" fields
         """
         if config.DECORD_AVAILABLE:
+            # We need to import torch first, otherwise later it can cause issues
+            # e.g. "RuntimeError: random_device could not be read"
+            # when running `torch.tensor(value).share_memory_()`
+            if config.TORCH_AVAILABLE:
+                import torch  # noqa
             from decord import VideoReader
 
         else:
@@ -129,6 +134,11 @@ def decode_example(self, value: dict, token_per_repo_id=None) -> "VideoReader":
             raise RuntimeError("Decoding is disabled for this feature. Please use Video(decode=True) instead.")
 
         if config.DECORD_AVAILABLE:
+            # We need to import torch first, otherwise later it can cause issues
+            # e.g. "RuntimeError: random_device could not be read"
+            # when running `torch.tensor(value).share_memory_()`
+            if config.TORCH_AVAILABLE:
+                import torch  # noqa
             from decord import VideoReader
         else:
             raise ImportError("To support decoding videos, please install 'decord'.")
@@ -302,6 +312,11 @@ def _patched_get_batch(self: "VideoReader", *args, **kwargs):
 
 def patch_decord():
     if config.DECORD_AVAILABLE:
+        # We need to import torch first, otherwise later it can cause issues
+        # e.g. "RuntimeError: random_device could not be read"
+        # when running `torch.tensor(value).share_memory_()`
+        if config.TORCH_AVAILABLE:
+            import torch  # noqa
         import decord.video_reader
         from decord import VideoReader
     else:

src/datasets/formatting/jax_formatter.py

@@ -106,6 +106,11 @@ def _tensorize(self, value):
             if isinstance(value, PIL.Image.Image):
                 value = np.asarray(value)
         elif config.DECORD_AVAILABLE and "decord" in sys.modules:
+            # We need to import torch first, otherwise later it can cause issues
+            # e.g. "RuntimeError: random_device could not be read"
+            # when running `torch.tensor(value).share_memory_()`
+            if config.TORCH_AVAILABLE:
+                import torch  # noqa
             from decord import VideoReader
 
             if isinstance(value, VideoReader):

src/datasets/formatting/np_formatter.py

@@ -63,6 +63,11 @@ def _tensorize(self, value):
             if isinstance(value, PIL.Image.Image):
                 return np.asarray(value, **self.np_array_kwargs)
         elif config.DECORD_AVAILABLE and "decord" in sys.modules:
+            # We need to import torch first, otherwise later it can cause issues
+            # e.g. "RuntimeError: random_device could not be read"
+            # when running `torch.tensor(value).share_memory_()`
+            if config.TORCH_AVAILABLE:
+                import torch  # noqa
             from decord import VideoReader
 
             if isinstance(value, VideoReader):

src/datasets/formatting/tf_formatter.py

@@ -70,6 +70,11 @@ def _tensorize(self, value):
             if isinstance(value, PIL.Image.Image):
                 value = np.asarray(value)
         elif config.DECORD_AVAILABLE and "decord" in sys.modules:
+            # We need to import torch first, otherwise later it can cause issues
+            # e.g. "RuntimeError: random_device could not be read"
+            # when running `torch.tensor(value).share_memory_()`
+            if config.TORCH_AVAILABLE:
+                import torch  # noqa
             from decord import VideoReader
             from decord.bridge import to_tensorflow