Struct GdnPrefillBuffers 

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

Device pointers to full-sequence GDN input/output buffers.

Used by the two-phase SSM prefill: phase 1 writes GDN inputs here, phase 2 reads them for the single-launch GDN kernel, phase 3 reads output.

Uses a packed QKV layout matching the conv1d output: each token occupies conv_dim contiguous BF16 elements as [Q(key_dim) | K(key_dim) | V(value_dim)]. This allows simple contiguous memcpy from per-chunk conv1d output buffers. The GDN kernel reads Q/K/V via stride parameters (qk_stride = conv_dim, v_stride = conv_dim) to index into the packed layout.

Fields

qkv: DevicePtr

Packed Q/K/V: [total_len, conv_dim] BF16. Layout per token: [Q(key_dim) | K(key_dim) | V(value_dim)].

gate_beta: DevicePtr

Interleaved gate/beta: [total_len, 2*num_v_heads] FP32. Layout per token: [gate(nv) | beta(nv)].

output: DevicePtr

GDN recurrence output: [total_len, value_dim] BF16.

z: DevicePtr

Z gate for gated RMS norm: [total_len, value_dim] BF16.

total_len: usize

Total number of tokens across all chunks.