class oneflow.nn.RMSLayerNorm(hidden_size, eps=1e-06)

Construct a layernorm module in the T5 style. No bias and no subtraction of mean.

T5 uses a layer_norm which only scales and doesn’t shift, which is also known as Root Mean

Square Layer Normalization thus varience is calculated

w/o mean and there is no bias. Additionally we want to make sure that the accumulation for

half-precision inputs is done in fp32.

  • hidden_size (int) – number of features in the hidden state

  • eps – a value added to the denominator for numerical stability. Default: 1e-6

  • Input: \((N, *)\)

  • Output: \((N, *)\) (same shape as input)

For example:

>>> import oneflow as flow

>>> x = flow.randn(2, 4, 3)
>>> m = flow.nn.RMSLayerNorm(3)
>>> y = m(x)
>>> y.size()
oneflow.Size([2, 4, 3])