Function kquant_mmvq_w 

pub fn kquant_mmvq_w(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    w_blocks: DevicePtr,
    y_q8: DevicePtr,
    out_bf16: DevicePtr,
    n: u32,
    k: u32,
    m: u32,
    stream: u64,
) -> Result<()>

kquant_mmvq_*_w: the warp-per-row GEMV (four rows a block, no shared memory), same arguments as kquant_mmvq.