Function moe_w4a16_fused_gate_up_k64_m128 

pub fn moe_w4a16_fused_gate_up_k64_m128(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    a: DevicePtr,
    gate_packed_ptrs: DevicePtr,
    gate_scale_ptrs: DevicePtr,
    gate_scale2_vals: DevicePtr,
    up_packed_ptrs: DevicePtr,
    up_scale_ptrs: DevicePtr,
    up_scale2_vals: DevicePtr,
    c_gate: DevicePtr,
    c_up: DevicePtr,
    expert_offsets: DevicePtr,
    sorted_token_ids: DevicePtr,
    num_experts: u32,
    n_out: u32,
    k: u32,
    max_m_tiles_m128: u32,
    stream: u64,
) -> Result<()>

2026-09-25: moe_w4a16_fused_gate_up_k64_n128 with a 128-row M tile and a 256-thread block. max_m_tiles_m128 counts 128-row tiles: the call site halves the 64-row count (forward_prefill_routed.rs).