pub fn moe_decode_atomic_c4_finalize(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
output: DevicePtr,
routed_accum: DevicePtr,
shared_out: DevicePtr,
input: DevicePtr,
gate_weight: DevicePtr,
hidden: u32,
num_tokens: u32,
include_shared: bool,
stream: u64,
) -> Result<()>