ttnn.as_tensor

ttnn.as_tensor = Operation(python_fully_qualified_name='ttnn.as_tensor', function=<function as_tensor>, preprocess_golden_function_inputs=<function default_preprocess_golden_function_inputs>, golden_function=None, postprocess_golden_function_outputs=<function default_postprocess_golden_function_outputs>, is_cpp_operation=False, is_experimental=False)

Converts the torch.Tensor tensor into a ttnn.Tensor.

Parameters:

  • tensor (torch.Tensor) – the input tensor.

  • dtype (ttnn.DataType, optional) – The ttnn data type.

Keyword Arguments:

  • layout (ttnn.Layout, optional) – The ttnn layout. Defaults to ttnn.ROW_MAJOR_LAYOUT.

  • device (ttnn.MeshDevice, optional) – The ttnn device. Defaults to None.

  • memory_config (ttnn.MemoryConfig, optional) – The ttnn memory configuration. Defaults to None.

  • cache_file_name (str | pathlib.Path, optional) – The cache file name. Defaults to None.

  • preprocess (Callable[[ttnn.Tensor], ttnn.Tensor], optional) – The function to preprocess the tensor before serializing/converting to ttnn. Defaults to None.

  • mesh_mapper (ttnn.CppTensorToMesh, optional) – The TensorToMesh to define the mapping from torch to multi-device. Defaults to None.

  • use_device_tilizer (bool, optional) –

    The flag that toggles whether to use host vs. device tilizer. Defaults to False.

    • For Grayskull, the on-device tilizer will truncate mantissa bits for bfp* formats.

    • For Wormhole, the on-device tilizer will raise a runtime error (RTE) for bfp8 but will truncate for bfp4/2 formats.

Returns:

ttnn.Tensor – The resulting ttnn tensor.

Examples

>>> tensor = ttnn.as_tensor(torch.randn((2,3)), dtype=ttnn.bfloat16)
>>> print(tensor)
Tensor([[1.375, -1.30469, -0.714844],
    [-0.761719, 0.53125, -0.652344]], dtype=bfloat16)