// is done.
// TODO(agarwal): change num_retvals to int from int*.
TF_CAPI_EXPORT extern void TFE_Execute(TFE_Op* op, TFE_TensorHandle** retvals,
                                       int* num_retvals, TF_Status* status);

// Add a function (serialized FunctionDef protocol buffer) to ctx so
// that it can be invoked using TFE_Execute.
TF_CAPI_EXPORT extern void TFE_ContextAddFunctionDef(

  virtual Status SetDeviceName(const char* name) = 0;

  virtual Status AddInput(AbstractTensorHandle* input) = 0;
  virtual Status AddInputList(
      absl::Span<AbstractTensorHandle* const> inputs) = 0;
  virtual Status Execute(absl::Span<AbstractTensorHandle*> retvals,
                         int* num_retvals) = 0;

  virtual Status SetAttrString(const char* attr_name, const char* data,
                               size_t length) = 0;

    if (TF_GetCode(s) != TF_OK) {
      return;
    }
  }
  int num_outputs = unwrap(o)->expected_num_outputs;
  tsl::Set_TF_Status_from_Status(
      s, unwrap(op)->Execute(
             absl::MakeSpan(reinterpret_cast<AbstractTensorHandle**>(
                                unwrap(o)->outputs.data()),
                            unwrap(o)->outputs.size()),
             &num_outputs));
}

  // TF_OK. Bad statuses are forwarded from underlying `TFE_Execute` calls, or
  // if sanity checks on dtypes/metadata fail.
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> Execute(
      TFE_Context* context, const std::vector<ParallelTensor*>& inputs,
      const char* operation_name, const TFE_OpAttrs* attributes,
      int expected_max_outputs, TF_Status* status) const;

  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

  void Execute(TFE_Context* context, const char* operation_name,

#include "tensorflow/c/c_api_experimental.h"
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/c/eager/c_api_experimental.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

  custom_device.copy_tensor_from_device = &CopyTensorFromLoggingDevice;
  custom_device.delete_device = &DeleteLoggingDevice;
  custom_device.execute = &LoggingDeviceExecute;
  LoggingDevice* device = new LoggingDevice;
  device->arrived_flag = arrived_flag;
  device->executed_flag = executed_flag;
  device->device_name = name;
  device->underlying_device = "/job:localhost/replica:0/task:0/device:CPU:0";

// Test to ensure that a registered graph optimization pass is only executed
// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops
// span the local and remote workers. If the graph optimization pass is executed

#include "tensorflow/c/tf_status_internal.h"
#include "tensorflow/core/lib/core/status_test_util.h"
#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {

// and `underlying_devices` is
//   {"/job:localhost/replica:0/task:0/device:GPU:0",
//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//