Allow passing tensor arguments in reader constructors#6252

Open

rostan-t wants to merge 22 commits intoNVIDIA:mainfrom

rostan-t:ndd-reader-tensor-args

Commits on Mar 19, 2026

Support constructing readers with tensor args
rostan-t
committed
Detect when default values are passed when invoking a reader
rostan-t
committed
Fix call stack depth for reader invocations when reconstructing stack traces
rostan-t
committed
Add tests passing tensor arguments
rostan-t
committed
Disallow constructing a reader with batch kwargs
rostan-t
committed
Update signature of reader constructors to allow tensor arguments
rostan-t
committed
Update NumpyReader example to pass ROI in the reader constructor
rostan-t
committed
Apply suggestions from review
rostan-t
committed
Prevent processing again tensor args when not necessary in batch processing
rostan-t
committed
Fix test_video_resize_tensor_args_partial
rostan-t
committed
Cache processed tensor args passed in the constructor
rostan-t
committed
Fix caller_depth handling. Remove special case for readers
rostan-t
committed
Pass scalar arguments directly to reader constructors and copy external tensors
rostan-t
committed
Set _raw_tensor_args instead of _tensor_args in reader constructor
rostan-t
committed
Fix tensor arg tracking in reader op constructor
rostan-t
committed
Properly use reader tensor args in TorchData integration
rostan-t
committed
Fix signature of reader constructors
rostan-t
committed
Fix tensor arg handling in reader op constructor
rostan-t
committed
Fix typos
rostan-t
committed
Copy all tensors passed to constructors and perform only broadcast in _process_tensor_args
rostan-t
committed
Revert the change to the caller depth.
rostan-t
committed
Perform dtype conversion of reader constructor arguments
rostan-t
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Allow passing tensor arguments in reader constructors#6252

Allow passing tensor arguments in reader constructors#6252
rostan-t wants to merge 22 commits intoNVIDIA:mainfrom
rostan-t:ndd-reader-tensor-args

Commits on Mar 19, 2026

Support constructing readers with tensor args

Detect when default values are passed when invoking a reader

Fix call stack depth for reader invocations when reconstructing stack traces

Add tests passing tensor arguments

Disallow constructing a reader with batch kwargs

Update signature of reader constructors to allow tensor arguments

Update NumpyReader example to pass ROI in the reader constructor

Apply suggestions from review

Prevent processing again tensor args when not necessary in batch processing

Fix test_video_resize_tensor_args_partial

Cache processed tensor args passed in the constructor

Fix caller_depth handling. Remove special case for readers

Pass scalar arguments directly to reader constructors and copy external tensors

Set _raw_tensor_args instead of _tensor_args in reader constructor

Fix tensor arg tracking in reader op constructor

Properly use reader tensor args in TorchData integration

Fix signature of reader constructors

Fix tensor arg handling in reader op constructor

Fix typos

Copy all tensors passed to constructors and perform only broadcast in _process_tensor_args

Revert the change to the caller depth.

Perform dtype conversion of reader constructor arguments