pub fn w4a16_gemv(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
input: DevicePtr,
weight: &QuantizedWeight,
output: DevicePtr,
n: u32,
k: u32,
stream: u64,
) -> Result<()>Expand description
W4A16 GEMV (M=1): C = A @ dequant(B) for single-row activations.
A: [1, K] BF16, B: NVFP4 packed, C: [1, N] BF16. 4 outputs/block, 64 threads (2 warps) per output. Cross-warp smem reduction.
Kernel: w4a16_gemv(A, B_packed, B_scale, scale2, C, N, K)
Grid: (ceil(N/4), 1, 1) Block: (256, 1, 1)