#ifndef TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_OPERATION_H_
#define TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_OPERATION_H_

#include <memory>

#include "absl/types/optional.h"
#include "absl/types/span.h"
#include "tensorflow/c/eager/abstract_operation.h"
#include "tensorflow/c/eager/immediate_execution_tensor_handle.h"
#include "tensorflow/c/tensor_interface.h"

// Causes the calling thread to block till all ops dispatched in this executor
// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.

class ParallelTensor;
class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

// to this graph.

// --------------------------------------------------------------------------
// API for driving Graph execution.

typedef struct TF_Session TF_Session;

// Return a new execution session with the associated graph, or NULL on
// error. Does not take ownership of any input parameters.
//

// TODO(b/154564140): Move this to its own header. This requires splitting
// c_api_experimental.h
struct TFE_ContextOptions {
  TF_SessionOptions session_options;
  // true if async execution is enabled.
  bool async = false;
  TFE_ContextDevicePlacementPolicy device_placement_policy{
      TFE_DEVICE_PLACEMENT_SILENT};
  // If true, use TFRT backend
  bool use_tfrt = false;

  // functions.
  virtual void MarkAsResult(Gradient* gradient) const = 0;

  // Deletes the input tensor.
  virtual void DeleteGradient(Gradient* gradient) const = 0;
};

// Traces the execution of operations, doing eager garbage collection, and
// exporting a full trace so other code can do backpropagation. Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class GradientTape {

limitations under the License.
==============================================================================*/

#ifndef TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_DISTRIBUTED_MANAGER_H_
#define TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_DISTRIBUTED_MANAGER_H_

#include <cstdint>
#include <string>

#include "tensorflow/core/platform/status.h"

namespace tsl {
class CoordinationServiceAgent;
}

// *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.
//
// For sync execution, if any of the inputs to `op` are not ready, this call

#ifndef TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_CONTEXT_H_
#define TENSORFLOW_C_EAGER_IMMEDIATE_EXECUTION_CONTEXT_H_

#include <functional>
#include <map>
#include <memory>
#include <vector>

#include "absl/types/optional.h"
#include "absl/types/span.h"
#include "tensorflow/c/eager/abstract_context.h"
#include "tensorflow/c/eager/immediate_execution_distributed_manager.h"

#include "tensorflow/core/platform/refcount.h"
#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Abstract interface to a Tensor handle in either tracing or immediate
// execution mode.
class AbstractTensorHandle : public core::RefCounted {
 protected:
  enum AbstractTensorHandleKind { kGraph, kMlir, kEager, kTfrt, kCustomDevice };