//        |         |
    //      Const_0    Const_1
    //
    const float const0_val[] = {1.0, 2.0, 3.0, 4.0};
    const float const1_val[] = {1.0, 0.0, 0.0, 1.0};
    TF_Operation* const0 = FloatConst2x2(graph_, s_, const0_val, "Const_0");
    TF_Operation* const1 = FloatConst2x2(graph_, s_, const1_val, "Const_1");
    TF_Operation* matmul = MatMul(graph_, s_, const0, const1, "MatMul");

  using tensorflow::RecordMutation;
  tensorflow::AttrValue attr_val;
  if (!attr_val.ParseFromArray(attr_value_proto->data,
                               attr_value_proto->length)) {
    status->status = absl::InvalidArgumentError("Invalid AttrValue proto");
    return;
  }

  mutex_lock l(graph->mu);
  op->node.AddAttr(attr_name, attr_val);
  RecordMutation(graph, *op, "setting attribute");
}

        to store weights, allowing a dramatic speedup for large sparse models.
    *   Enable the Keras compile API `experimental_run_tf_function` flag by
        default. This flag enables single training/eval/predict execution path.
        With this 1. All input types are converted to `Dataset`. 2. When
        distribution strategy is not specified this goes through the no-op

                         'Assuming to be a scripting mistake.' %
                         (var_name, n_ask_attempts))

  if resolve_symlinks:
    val = os.path.realpath(val)
  environ_cp[var_name] = val
  return val


def set_clang_cuda_compiler_path(environ_cp):
  """Set CLANG_CUDA_COMPILER_PATH."""
  default_clang_path = '/usr/lib/llvm-17/bin/clang'
  if not os.path.exists(default_clang_path):