Struct HrmCore 

pub struct HrmCore {
    pub t_inner: usize,
    pub n_cycles: usize,
    /* private fields */
}

The two-timescale core. L iterates fast against the injected input; H steps back once per cycle to integrate L’s state. Output is zH + reps, a residual refinement of the embeddings.

Fields

t_inner: usize

n_cycles: usize

Implementations

impl HrmCore

pub fn new( h: usize, layers_l: usize, layers_h: usize, t_inner: usize, n_cycles: usize, heads: usize, vb: VarBuilder<'_>, ) -> Result<Self>

pub fn forward( &self, reps: &Tensor, mask: Option<&Tensor>, grad_last_step: bool, ) -> Result<Tensor>

Refine reps [B,T,h] → [B,T,h]. The N_cycles of refinement run detached (the reference’s no_grad), and only the final L+H step carries gradient — the one-step gradient approximation that makes recurrent depth trainable at this size.