.. _op_ai_rt_DelayedInitializer:

DelayedInitializer
==================

- **Domain**: ``ai.rt``
- **Since version**: 1

Defers materialization of a tensor stored in an external weights file.

In onnx-light, ```load_device```` must be either ````"cpu"```` or ````"file"``` and
```runtime_device```` must be ````"cpu"````. When ````load_device```` is ````"cpu"```, the
kernel loads the tensor bytes during kernel initialization and returns a CPU
copy at execution time. When ```load_device```` is ````"file"```, initialization does
not touch the file and execution loads the tensor bytes directly from
```filename```` at byte ````offset```.

The static output shape comes from the required ```shape``` attribute and the
element type comes from the required ```dtype``` attribute.

**Outputs**

- **output** (*T*): Tensor produced by the delayed initializer.

**Attributes**

- **dtype** (*int*): Element type of the output tensor, encoded as a TensorProto::DataType value.
- **filename** (*string*): Filename containing the serialized tensor payload.
- **load_device** (*string*): Device where the initializer is first loaded.
- **offset** (*int*): Byte offset of the tensor payload within ``filename``.
- **runtime_device** (*string*): Device where the initializer is moved at runtime.
- **shape** (*int[]*): Shape of the output tensor.

**Type Constraints**

- **T**: Constrain output to tensor types backed by raw byte storage.
  Allowed types: tensor(bool), tensor(complex128), tensor(complex64), tensor(double), tensor(float), tensor(float16), tensor(int16), tensor(int32), tensor(int64), tensor(int8), tensor(uint16), tensor(uint32), tensor(uint64), tensor(uint8).

Examples
--------

**test_cc_delayedinitializer_cpu**

.. code-block:: text

    Node:
      ai.rt.DelayedInitializer() -> (y)
      Attributes:
        shape = [3]
        dtype = 1
        load_device = "cpu"
        runtime_device = "cpu"
        filename = "/tmp/onnx_light_backend_delayedinitializer_cpu.bin"
        offset = 0

.. code-block:: text

    Inputs:

    Outputs:
      y: shape=(3,), dtype=float32
        [3., 4., 5.]

**test_cc_delayedinitializer_file**

.. code-block:: text

    Node:
      ai.rt.DelayedInitializer() -> (y)
      Attributes:
        shape = [2]
        dtype = 1
        load_device = "file"
        runtime_device = "cpu"
        filename = "/tmp/onnx_light_backend_delayedinitializer_file.bin"
        offset = 8

.. code-block:: text

    Inputs:

    Outputs:
      y: shape=(2,), dtype=float32
        [ 1.5, -2. ]