Function gdn_prefill_regresident 

pub fn gdn_prefill_regresident(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    h_state: DevicePtr,
    query: DevicePtr,
    key: DevicePtr,
    value: DevicePtr,
    gate: DevicePtr,
    beta: DevicePtr,
    output: DevicePtr,
    batch_size: u32,
    seq_len: u32,
    num_k_heads: u32,
    num_v_heads: u32,
    k_dim: u32,
    v_dim: u32,
    qk_stride: u32,
    v_stride: u32,
    gb_stride: u32,
    stream: u64,
) -> Result<()>

Register-resident token-sequential prefill recurrence (warm-replay path).

Kernel gated_delta_rule_prefill_regresident: one WARP owns one v-column, holding the 128 k-rows of H in registers (4/lane) — no smem-H, no per-token barriers, >=2 CTA/SM. Token-equal to WY4 (cosine 1.0) and ~2.9x faster. Grid: (num_v_heads, batch, v_dim / 4) Block: (128, 1, 1) (4 warps/block). Requires k_dim == 128 and v_dim % 4 == 0.