Function w4a16_decode_gemv 

pub fn w4a16_decode_gemv(
    gpu: &dyn GpuBackend,
    gemv: KernelHandle,
    gemv_sw: KernelHandle,
    use_sw: bool,
    input: DevicePtr,
    weight: &QuantizedWeight,
    output: DevicePtr,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

Decode GEMV: software-pipelined single-warp when the lever and handle agree.