Module gelu 

GELU activation (pytorch_tanh variant) GPU dispatch.

Computes: 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))

This is the exact variant used by Gemma 4. It is not the erf-based GELU approximation.

Statics

GELU_SHADER_SOURCE

MSL source for the GELU kernels (embedded at compile time).

Functions

dispatch_gelu

Dispatch a GELU activation on the GPU.

register

Register GELU shader sources with the given kernel registry.