pub fn nvfp4_scale_bf16(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
data: DevicePtr,
scale: f32,
total: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: data[i] *= scale in place over total BF16 values; the caller
passes the down projection’s weight_scale_2.