Struct CombinedGrads 

#[non_exhaustive]
pub struct CombinedGrads<B: Backend> {
    pub d_x_bnlhp: F<B, 5>,
    pub d_dt_discretized_bhnl: F<B, 4>,
    pub d_b_bnlhr: F<B, 5>,
    pub d_c_bnlhr: F<B, 5>,
    pub d_d_h: F<B, 1>,
    pub d_initial_state_bhpr: F<B, 4>,
    pub d_a_decay_h: F<B, 1>,
}

Per-input gradients produced by combined_backward (one field per differentiable forward input).

Fields (Non-exhaustive)

Non-exhaustive structs could have additional fields added in future. Therefore, non-exhaustive structs cannot be constructed in external crates using the traditional Struct { .. } syntax; cannot be matched against without a wildcard ..; and struct update syntax will not work.

d_x_bnlhp: F<B, 5>

Gradient of the input x.

d_dt_discretized_bhnl: F<B, 4>

Gradient of the discretised step Δ (dt).

d_b_bnlhr: F<B, 5>

Gradient of the input projection B.

d_c_bnlhr: F<B, 5>

Gradient of the output projection C.

d_d_h: F<B, 1>

Gradient of the per-head skip term D.

d_initial_state_bhpr: F<B, 4>

Gradient of the initial SSM state.

d_a_decay_h: F<B, 1>

Gradient of the per-head decay rate A (as a_decay_h).