Struct GdnPrefillBuffers 

pub struct GdnPrefillBuffers {
    pub qkv: DevicePtr,
    pub gate_beta: DevicePtr,
    pub output: DevicePtr,
    pub z: DevicePtr,
    pub total_len: usize,
}

2026-09-25: Whole-prompt GDN buffers for the split SSM prefill: prefill_phase1 writes the inputs, prefill_gdn_full runs the recurrence into output, and prefill_phase3 reads output and z.

Each token’s QKV is packed as conv_dim BF16 elements, [Q(key_dim) | K(key_dim) | V(value_dim)], and the GDN kernels read Q, K and V with a stride of conv_dim.

Fields

qkv: DevicePtr

2026-09-25: Packed Q/K/V, [total_len, conv_dim] BF16.

gate_beta: DevicePtr

2026-09-25: Gate and beta, [total_len, 2 * num_v_heads] FP32: per token [gate(nv) | beta(nv)].

output: DevicePtr

2026-09-25: GDN recurrence output, [total_len, value_dim] BF16.

z: DevicePtr

2026-09-25: Z gate for the gated RMSNorm, [total_len, value_dim] BF16.

total_len: usize

2026-09-25: Token rows the buffers hold for this call.