pub fn gdn_decode_chunk3(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
h_state: DevicePtr,
query: DevicePtr,
key: DevicePtr,
value: DevicePtr,
gate: DevicePtr,
beta: DevicePtr,
output: DevicePtr,
h_state_inter0: DevicePtr,
h_state_inter1: DevicePtr,
batch_size: 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<()>Expand description
Fused 3-token GDN decode (K=3 speculative verification).
Processes exactly 3 tokens through GDN in a single kernel launch. Saves 2 intermediate H states (H_1, H_2) for rollback on draft rejection. 4 passes vs 6 for 3× sequential decode.
Kernel: gated_delta_rule_chunk3(h_state, query, key, value, gate, beta, output, h_inter0, h_inter1, batch_size, num_k_heads, num_v_heads, k_dim, v_dim, qk_stride, v_stride, gb_stride)
Grid: (num_v_heads, batch, 1) Block: (128, 1, 1)