tf.nn.softmax_cross_entropy_with_logits(labels, logits))

    grads = tape.gradient(loss_value, model.trainable_variables)
    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    optimizer.apply_gradients(zip(grads, model.trainable_variables))
    return accuracy, loss_value

  @tf.function
  def distributed_train_step(dist_inputs):

        value {
          type_url: "type.googleapis.com/tensorflow.AssetFileDef"
          value: "\n\t\n\007Const:0\022\007foo.txt"
        }
      }
    }
  }
  collection_def {
    key: "trainable_variables"
    value {
      bytes_list {
        value: "\n\003a:0\022\010a/Assign\032\010a/read:0"
        value: "\n\003b:0\022\010b/Assign\032\010b/read:0"
        value: "\n\003c:0\022\010c/Assign\032\010c/read:0"

                // - `model.keras_api.layers.1.kernel`
                // - `model.variables.0`
                // - `model.keras_api.layers.1.keras_api.trainable_variables.0`
                // - ... 10 more long aliases ending in digits ...
                return std::make_tuple(isdigit(a.back()), a.size(), a) <
                       std::make_tuple(isdigit(b.back()), b.size(), b);

        `apply_gradients()` or `minimize()` call. If your workflow calls
        optimizer to update different parts of model in multiple stages,
        please call `optimizer.build(model.trainable_variables)` before the
        training loop.
    *   **Performance regression on `ParameterServerStrategy`.** This could be
        significant if you have many PS servers. We are aware of this issue and