ttnn.softcap

ttnn.softcap(input_tensor: ttnn.Tensor, beta: float, *, memory_config: ttnn.MemoryConfig = None, output_tensor: ttnn.Tensor = None, sub_core_grids: ttnn.CoreRangeSet = None) → ttnn.Tensor

Applies softcap to input_tensor element-wise with beta.

Bounds the input smoothly to +/-beta: near-linear well inside beta, saturating at the limits. Known as soft capping (e.g. Gemma logit softcapping); also the up half of Moonshot’s SiTU activation.

\[\mathrm{output\_tensor}_i = \verb|beta| \cdot \tanh(\mathrm{input\_tensor}_i / \verb|beta|)\]
Parameters:

  • input_tensor (ttnn.Tensor)the input tensor.

  • beta (float)The beta parameter. Bounds the output to +/-beta. Must be non-zero.

Keyword Arguments:

  • memory_config (ttnn.MemoryConfig, optional)Memory configuration for the operation. Defaults to None.

  • output_tensor (ttnn.Tensor, optional)preallocated output tensor. Defaults to None.

  • sub_core_grids (ttnn.CoreRangeSet, optional)sub core grids for the operation. Defaults to None.

Returns:

ttnn.Tensorthe output tensor.

Note

Supported dtypes and layouts:

Dtypes

Layouts

BFLOAT16, BFLOAT8_B

TILE, ROW_MAJOR