# First run tfcompile to generate the list of out_nodes.
        #
        # Here and below, we set CUDA_VISIBLE_DEVICES='' to prevent the code we
        # launch from using any GPUs which might be present.  This is important
        # because builds may run concurrently with tests, and tests need to be
        # able to assume that they have control of the full GPU.
        out_nodes_file = "out_nodes_" + freeze_name
        native.genrule(