Function stable_softmax

src/transformers/src/transformers/tf_utils.py:51–72

Stable wrapper that returns the same output as `tf.nn.softmax`, but that works reliably with XLA on CPU. It is meant as a workaround for the [following issue](https://github.com/tensorflow/tensorflow/issues/55682), and will be removed after it gets fixed. The arguments and outputs are t

(logits: tf.Tensor, axis: Optional[int] = None, name: Optional[str] = None)

def stable_softmax(logits: tf.Tensor, axis: Optional[int] = None, name: Optional[str] = None) -> tf.Tensor:
    """
    Stable wrapper that returns the same output as `tf.nn.softmax`, but that works reliably with XLA on CPU. It is
    meant as a workaround for the [following issue](https://github.com/tensorflow/tensorflow/issues/55682), and will be
    removed after it gets fixed. The arguments and outputs are the same as `tf.nn.softmax`, and relies on the fact that
    `softmax(x) = softmax(x + c)` (see https://ogunlao.github.io/2020/04/26/you_dont_really_know_softmax.html).

    Args:
        logits (`tf.Tensor`):
            Must be one of the following types: half, float32, float64.
        axis (`int`, *optional*):
            The dimension softmax would be performed on. The default is -1 which indicates the last dimension.
        name (`str`, *optional*):
            A name for the operation.

    Returns:
        `tf.Tensor`:
            A Tensor. Has the same type and shape as logits.
    """
    # TODO: When the issue linked above gets sorted, add a check on TF version here and use the original function if
    # it has the fix. After we drop the support for unfixed versions, remove this function.
    return tf.nn.softmax(logits=logits + 1e-9, axis=axis, name=name)


def functional_layernorm(inputs, weight, bias, epsilon=1e-5, axis=-1):