// from a placeholder node "arg_tensor_enqueue_<tensor_id>".
//
// `tensor` is still owned by the caller. This call will be blocked if the queue
// has reached its capacity, and will be unblocked when the queued tensors again
// drop below the capacity due to dequeuing.
//
// Tensors are dequeued via the corresponding TF dequeue op.
// TODO(hongm): Add support for `timeout_ms`.

// A value of 0 removes the limit, which is the behavior of TensorFlow 2.11.
// When is_async is false, the value is ignored.
TF_CAPI_EXPORT extern TFE_Executor* TFE_NewExecutor(
    bool is_async, bool enable_streaming_enqueue, int in_flight_nodes_limit);

// Deletes the eager Executor without waiting for enqueued nodes. Please call
// TFE_ExecutorWaitForAllPendingNodes before calling this API if you want to
// make sure all nodes are finished.

    // While the Accumulate method is running (accumulating is True), any op
    // 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

// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.
//
// If async execution is enabled, the call may simply enqueue the execution
// and return "non-ready" handles in `retvals`. Note that any handles contained
// in 'op' should not be mutated till the kernel execution actually finishes.
//

#include "tensorflow/core/common_runtime/eager/eager_executor.h"

struct TFE_Executor {
  explicit TFE_Executor(bool async, bool enable_streaming_enqueue,
                        int in_flight_nodes_limit)
      : owned_executor(new tensorflow::EagerExecutor(
            async, enable_streaming_enqueue, in_flight_nodes_limit)) {}

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)
      : owned_executor(nullptr), unowned_executor(executor) {}