pub fn moe_silu_mul(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
gate: DevicePtr,
up: DevicePtr,
output: DevicePtr,
total_elements: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: output[i] = silu(gate[i]) * up[i] over total_elements BF16 values.