core.activations#

Module Contents#

Functions#

squared_relu

Squared ReLU activation

tanh_soft_clamp

Tanh Soft Clamp to precondition activation inputs.

situ

Sigmoid-Tanh Unit from Kimi K3: s * tanh(x / s) * sigmoid(x)

situ_glu

SiTU-GLU: situ(x_gate, gate_scale) * (x_linear + linear_offset).

quick_gelu

Quick GELU activation

fast_gelu

Fast GELU activation

API#

core.activations.squared_relu(x: torch.Tensor) torch.Tensor#

Squared ReLU activation

core.activations.tanh_soft_clamp(x: torch.Tensor, scale: float) torch.Tensor#

Tanh Soft Clamp to precondition activation inputs.

core.activations.situ(x: torch.Tensor, scale: float) torch.Tensor#

Sigmoid-Tanh Unit from Kimi K3: s * tanh(x / s) * sigmoid(x)

core.activations.situ_glu(
x: torch.Tensor,
gate_scale: float,
linear_scale: Optional[float] = None,
linear_offset: float = 0.0,
) torch.Tensor#

SiTU-GLU: situ(x_gate, gate_scale) * (x_linear + linear_offset).

core.activations.quick_gelu(x: torch.Tensor) torch.Tensor#

Quick GELU activation

core.activations.fast_gelu(x: torch.Tensor) torch.Tensor#

Fast GELU activation