onnx-tensorrt
onnx-tensorrt copied to clipboard
BitShift
When is BitShift supported?
In a torch -> onnx -> onnxruntime[TensorRTExecutionProvider] pipeline, pytorch casts values to uint64 to satisfy type constrains on the BitShift operator.
However, this causes an error in the onnxruntime backend: WeightsContext.cpp:454: Found unsupported datatype (13) when importing initializer: mask_3
TensorRT only knows int64 (https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/FoundationalTypes/DataType.html#datatype).
Any idea how to work around this ?