BiasGelu (CPU)
==============

* Domain: ``com.microsoft``
* Device: ``CPU``
* Kernel: ``onnx_light_cpu::BiasGelu``
* Supported types: ``FLOAT``, ``DOUBLE``, ``FLOAT16``, ``BFLOAT16``
* Opset bounds: since opset 1

Operator support
----------------

.. list-table::
   :header-rows: 1
   :widths: 30 70

   * - Capability
     - Implementation
   * - Shape inference
     - :cpp:func:`onnx_light_cpu::ComputeShapeBiasGelu`
   * - Peak memory
     - :cpp:func:`onnx_light_cpu::ComputePeakMemoryBiasGelu`
   * - Fusion patterns
     - :cpp:class:`onnx_light_cpu::BiasGeluFusionPattern`
   * - Gradient
     - :cpp:func:`onnx_light_cpu::RegisterCustomOperatorGradients`

LightOpSchema
~~~~~~~~~~~~~

**Version:** 1

Adds a rank-1 bias to the last input dimension and applies the exact Gaussian Error Linear Unit.

**Inputs**

- **A** (``T``): The input tensor.
- **B** (``T``): The rank-1 bias whose length matches the last dimension of A.

**Outputs**

- **C** (``T``): The output tensor, with the same shape and type as A.

**Type constraints**

- **T**: Constrain inputs and output to floating-point tensors. Allowed types: tensor(bfloat16), tensor(double), tensor(float), tensor(float16).