*   XLA reduction emitter is deterministic when the environment variable
        `TF_DETERMINISTIC_OPS` is set to "true" or "1". This extends
        deterministic `tf.nn.bias_add` back-prop functionality (and therefore
        also deterministic back-prop of bias-addition in Keras layers) to
        include when XLA JIT compilation is enabled.
    *   Fix problem, when running on a CUDA GPU and when either environment

    // executions not forwarded to backward_tape should be ignored.
    bool accumulating;
  };
  // A deque-backed stack, whose element references are not invalidated by
  // pushes and pops at the back.
  std::stack<AccumulatorCallState> call_state_;
};

// Template instantiations here

inline bool IsDtypeTrainable(DataType dtype) {
  switch (dtype) {
    case DT_HALF: