pub fn gelu_backward<T: Triton, D: Float, const BLOCK_SIZE: i32>(
dy_ptr: T::Pointer<D>,
x_ptr: T::Pointer<D>,
dx_ptr: T::Pointer<D>,
n_elements: i32,
)where
T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
T::Pointer<D>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<D>>>,Expand description
Backward of the tanh-GELU approximation. Let inner = x + ax³, s = sigmoid(2cinner), t = tanh(cinner) = 2s-1 d/dx = 0.5(1 + t) + x * 0.5 * sech²(cinner) * c(1+3a*x²)