pub fn launch_tmp<R: Runtime, AP: AttentionPrecision>(
client: &ComputeClient<R::Server>,
query: &TensorHandleRef<'_, R>,
key: &TensorHandleRef<'_, R>,
value: &TensorHandleRef<'_, R>,
mask: &Option<TensorHandleRef<'_, R>>,
out: &TensorHandleRef<'_, R>,
) -> Result<(), AttentionSetupError>