CausalConvWithState (CPU)
=========================

* Domain: ``com.microsoft``
* Device: ``CPU``
* Kernel: ``onnx_light_cpu::CausalConvWithState``
* Supported types: ``FLOAT``, ``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::ComputeShapeCausalConvWithState`
   * - Peak memory
     - :cpp:func:`onnx_light_cpu::ComputePeakMemoryCausalConvWithState`
   * - Fusion patterns
     - None
   * - Gradient
     - Not provided

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

**Version:** 1

Stateful causal 1D depthwise convolution. Supports dilation, optional SiLU/Swish, channels-first or sequence-major channels-last layout, and unwindowed recurrent state.

**Inputs**

- **input** (``T``): (B, C, L), or (B, L, d_1, ...) when channels_last=1.
- **weight** (``T``): (C, 1, K) depthwise weights.
- **bias** (``T``): Optional (C) bias.
- **past_state** (``T``): Optional previous state, same shape as present_state.

**Outputs**

- **output** (``T``): Same shape and type as input.
- **present_state** (``T``): (B, C, (K-1)*dilation), or (B, (K-1)*dilation, d_1, ...) when channels_last=1.

**Attributes**

- **activation** (``STRING``): none, silu or swish. (optional, default: ``none``)
- **channels_last** (``INT``): Sequence-major layout when 1. (optional, default: ``0``)
- **dilation** (``INT``): Positive spacing between kernel taps. (optional, default: ``1``)
- **ndim** (``INT``): Only 1 is supported on CPU. (optional, default: ``1``)
- **state_window** (``INT``): Only unwindowed state (0) is supported on CPU. (optional, default: ``0``)

**Type constraints**

- **T**: Matching floating-point inputs and outputs. Allowed types: tensor(bfloat16), tensor(float), tensor(float16).