TFE_CancellationManager*, TFE_CancellationToken token,
    const TFE_CancelCallback* c_callback, const char* callback_name);
TF_CAPI_EXPORT extern bool TFE_CancellationManagerDeregisterCallback(
    TFE_CancellationManager*, TFE_CancellationToken token);
TF_CAPI_EXPORT extern bool TFE_CancellationManagerTryDeregisterCallback(
    TFE_CancellationManager*, TFE_CancellationToken token);
TF_CAPI_EXPORT extern void TFE_DeleteCancellationManager(

void TF_OutputListPushBack(TF_OutputList* o, TF_AbstractTensor* tensor,
                           TF_Status*);

// TF_ExecuteOperation will, if in eager mode, execute, if in graph mode, maybe
// capture some inputs and then add a node in the graph. The output tensors are
// returned through the provided TF_OutputList.
// Any active tape will observe the effects of this execution.
void TF_ExecuteOperation(TF_AbstractOp* op, int num_inputs,

// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:
// streamz->GetCell("label-value")->IncrementBy(1);
//
// The test code can then read and test the value of that counter:
//
// auto* reader = TFE_MonitoringNewCounterReader("name");
// test();

//   "/job:localhost/replica:0/task:0/device:CUSTOM:0"
// 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

                                                void (*work_func)(void*),
                                                void* param);

// Waits for the given thread to finish execution, then deletes it.
TF_CAPI_EXPORT extern void TF_JoinThread(TF_Thread* thread);

// \brief Load a dynamic library.
//
// Pass "library_filename" to a platform-specific mechanism for dynamically

  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future
  // `StartExecute`s will deadlock).
  //

  // (1) If type == TF_ATTR_STRING and is_list == 0
  //     then total_size is the byte size of the string
  //     valued attribute.
  // (2) If type == TF_ATTR_STRING and is_list == 1
  //     then total_size is the cumulative byte size
  //     of all the strings in the list.
  // (3) If type == TF_ATTR_SHAPE and is_list == 0
  //     then total_size is the number of dimensions
  //     of the shape valued attribute, or -1

        bool accumulating)
        : backward_tape(backward_tape), accumulating(accumulating) {}
    // Set temporarily while in the Accumulate method; if backward_tape is not
    // nullptr then we forward op executions to it so Accumulate can compute a
    // backward pass on its backward function.
    //
    // Not owned by the ForwardAccumulator. The method which sets
    // `backward_tape` keeps ownership.

// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
//
// Do not return const char*, because some foreign language binding
// (e.g. swift) cannot then call free() on the returned pointer.
TF_CAPI_EXPORT extern char* TF_FunctionDebugString(TF_Function* func,
                                                   size_t* len);

TFE_ContextGetDevicePlacementPolicy(TFE_Context* ctx);

// A tensorflow.ServerDef specifies remote workers (in addition to the current
// workers name). Operations created in this context can then be executed on
// any of these remote workers by setting an appropriate device.
//
// If the following is set, all servers identified by the
// ServerDef must be up when the context is created.