Struct HyperConnectionArgs 

pub struct HyperConnectionArgs<'a, T: Element> {
    pub x_expanded: TensorRef<'a, T, 3>,
    pub rmsnorm_weight: TensorRef<'a, bf16, 1>,
    pub h_pre: TensorRef<'a, f32, 1>,
    pub h_post: TensorRef<'a, f32, 1>,
    pub h_res: TensorRef<'a, f32, 2>,
    pub out: TensorMut<'a, T, 3>,
}

Args bundle for a HyperConnectionPlan launch.

Fields

x_expanded: TensorRef<'a, T, 3>

Residual-stream input — [B, n, C] row-major contiguous.

rmsnorm_weight: TensorRef<'a, bf16, 1>

RMSNorm gamma — [C] bf16. Always bf16 regardless of T (matches upstream floatX typedef).

h_pre: TensorRef<'a, f32, 1>

Pre-mixing logits — [n] f32. The kernel passes them through sigmoid internally.

h_post: TensorRef<'a, f32, 1>

Post-mixing logits — [n] f32. The kernel passes them through 2 * sigmoid(.) internally.

h_res: TensorRef<'a, f32, 2>

Pre-Sinkhorn residual mixing matrix — [n, n] f32. The kernel passes it through Sinkhorn-Knopp iteration to project onto the doubly-stochastic manifold before mixing.

out: TensorMut<'a, T, 3>

Output — [B, n, C] row-major contiguous, same dtype as input.