Function w8a16_gemv 

pub fn w8a16_gemv(
    gpu: &dyn GpuBackend,
    kernel: KernelHandle,
    input: DevicePtr,
    weight: DevicePtr,
    row_scale: DevicePtr,
    output: DevicePtr,
    n: u32,
    k: u32,
    stream: u64,
) -> Result<()>

2026-09-25: W8A16 GEMV for M=1 decode: C[1,N] = A[1,K] @ dequant(B[N,K])^T.

row_scale is the 2D block scale [ceil(N/128), ceil(K/128)] FP32, one scale per 128x128 weight block, despite the parameter name. Each 256-thread block computes 4 outputs with 64 threads per output, so the grid is ceil(N/4) (kernels/gb10/common/w8a16_gemv.cu, and the same geometry in kernels/hopper/common/w8a16_gemv.cu).