moe_expert_gemv_silu_down

spark_model::layers::ops

Function moe_expert_gemv_silu_down 

Source 
pub fn moe_expert_gemv_silu_down(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    gate_out: DevicePtr,
    up_out: DevicePtr,
    packed_ptrs: DevicePtr,
    scale_ptrs: DevicePtr,
    scale2_vals: DevicePtr,
    output: DevicePtr,
    expert_indices: DevicePtr,
    n: u32,
    k: u32,
    top_k: u32,
    stream: u64,
) -> Result<()>
Expand description

Fused SiLU+down expert GEMV: computes silu(gate)*up inline as activation.

Eliminates separate silu_mul kernel. Reads both gate_out and up_out, computes silu(gate)*up per element, then GEMV with down weights.

Grid: (ceil(N/4), top_k, 1) Block: (128, 1, 1)