---
title: "clika_runtime.nn.functional"
sidebar_label: "clika_runtime.nn.functional"
description: "The clika_runtime.nn.functional module, introspected from the installed wheel."
---

<!-- Generated by tools/api_reference/generate_api_docs.py. Do not edit. -->

Functional interface to the neural-network operators::

    import clika_runtime.nn.functional as F

    y = F.relu(x)
    p = F.softmax(y, dim=-1)

Use these when a computation has no state worth holding in a
:class:`~clika_runtime.nn.Module`: activations, padding, pooling,
normalization, losses, attention. Every function takes its input tensor
first and returns a new tensor; a trailing-underscore twin (``F.relu_``)
writes through its input instead of allocating. A mode argument takes a
name (``F.gelu(x, approximate="tanh")``) or the bound enum value.

Pooling and convolution run over channels-last tensors
(``[batch, *spatial, channels]``); pooling knobs take an int or a
per-dimension sequence. The dropout family is the identity outside
training; with ``training=True`` it draws its Bernoulli mask on the
tensor's device. ``rrelu`` uses its fixed mean slope outside training and
draws one slope per element, on the tensor's device, with
``training=True``. ``__all__`` lists every available name.

| Name | Kind |
| --- | --- |

| [functions](./functions.md) | module functions |