// Method to delete a device.
  void (*delete_device)(void* device_info);

  // Implements TFE_CreatePackedTensorHandle when one of `handles` is on this
  // custom device.
  //
  // Many devices will want to simply return an "unimplemented" status
  // here. This is the default behavior if `pack` is null when passed to
  // TFE_RegisterCustomDevice.
  TFE_TensorHandle* (*pack)(TFE_Context* context, TFE_TensorHandle** handles,

void TF_DeleteAbstractTensor(TF_AbstractTensor*);

// TF_OutputList holds the list of TF_AbstractTensor that results from executing
// an operation, or provided to create a function.
// When executing an operation in an eager context, the expected number of
// outputs must be set beforehand with `TF_OutputListSetNumOutputs`.
typedef struct TF_OutputList TF_OutputList;
TF_OutputList* TF_NewOutputList();

// nodes. This may cause incorrectness in graph analyses, but is useful for
// simulation or scheduling.
// If aggressive_shape_inference is true, nodes are executed on the host to
// identify output values when possible and does other aggressive strategies.
// This may cause incorrectness in graph analyses, but is useful for simulation
// or scheduling.
// If include_input_tensor_values is true, the values of constant

//
// The lock for `graph` must be held when calling this function.
//
// Unless implementing advanced behavior, like custom gradient functions, you
// most likely need to call `TF_NewOperation` instead.
TF_CAPI_EXPORT extern TF_OperationDescription* TF_NewOperationLocked(
    TF_Graph* graph, const char* op_type, const char* oper_name);

// Operation will only be added to *graph when TF_FinishOperation() is

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);

  ~ParallelDevice();

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//
// Caller takes ownership of the result and must call TF_JoinThread on it
// eventually.

// values and hence receives 5 gradient values during backprop. However the
// gradient function uses only the first of those values and ignores the rest.
// The entry, "FusedBatchNorm": [1, 2, 3, 4], indicates that only the gradient
// corresponding to index 0 is used, and the gradient values at indices 1-4 are
// ignored (and hence can be None). The backprop algorithm can then leverage

// needs to be statically registered in some tests and utilities for building
// the API files at the time of creating the pip package. Hence, we need to
// expose this function so that this filesystem can be statically registered
// when needed.
void TF_InitPlugin(TF_FilesystemPluginInfo* info);

TF_CAPI_EXPORT unsigned char TF_GetXlaConstantFoldingDisabled();
TF_CAPI_EXPORT void TF_SetXlaConstantFoldingDisabled(
    unsigned char should_enable);

// Create a serialized tensorflow.ConfigProto proto, where:
//
// a) ConfigProto.optimizer_options.global_jit_level is set to ON_1 if
// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.

    TFE_TensorHandle* h, TF_Status* status);

// This function will block till the operation that produces `h` has
// completed. The memory returned might alias the internal memory used by
// TensorFlow. Hence, callers should not mutate this memory (for example by
// modifying the memory region pointed to by TF_TensorData() on the returned
// TF_Tensor).
TF_CAPI_EXPORT extern TF_Tensor* TFE_TensorHandleResolve(TFE_TensorHandle* h,