pub fn w8a8_gemv(
gpu: &dyn GpuBackend,
kernels: &W8a8Kernels,
w: &W8a8Weight,
scratch: &W8a8Scratch,
rows: usize,
out: DevicePtr,
ldc: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-28: The GEMV over the activation w8a8_act_quant left in
scratch: out[m, 0..n] (row pitch ldc BF16 elements) for m < rows.