// Method to copy a tensor from the custom device to a target device.
  TFE_TensorHandle* (*copy_tensor_from_device)(TFE_Context* context,
                                               TFE_TensorHandle* tensor,
                                               const char* target_device_name,
                                               TF_Status* status,

  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // 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

// Set the target in TF_SessionOptions.options.
// target can be empty, a single entry, or a comma separated list of entries.
// Each entry is in one of the following formats :
// "local"
// ip:port
// host:port
TF_CAPI_EXPORT extern void TF_SetTarget(TF_SessionOptions* options,
                                        const char* target);

// Set the config in TF_SessionOptions.options.

// temporary storage. You can iterate through the list with TF_StringStreamNext.
// The caller is responsible for freeing the list (see TF_StringStreamDone).
TF_CAPI_EXPORT extern TF_StringStream* TF_GetLocalTempDirectories(void);

// Creates a temporary file name with an extension.
// The caller is responsible for freeing the returned pointer.
TF_CAPI_EXPORT extern char* TF_GetTempFileName(const char* extension);

  // it in this context.
  virtual AbstractOperation* CreateOperation() = 0;

  // Registers a function with this context, after this the function is
  // available to be called/referenced by its name in this context.
  virtual Status RegisterFunction(AbstractFunction*) = 0;
  // Remove a function. 'func' argument is the name of a previously added
  // FunctionDef. The name is in fdef.signature.name.

                          const char* device, TF_Status* status);
  // Dereferences the backing buffer for the variable. Note that since this can
  // fail (it runs operations), it must be called explicitly and the resulting
  // `status` checked.
  void Destroy(TFE_Context* context, TF_Status* status);

  // Reads from the variable.
  TensorHandlePtr Read(TFE_Context* context, TF_Status* status);

// The returned string is heap-allocated, and caller should call free() on it.
TF_CAPI_EXPORT extern const char* TF_GraphDebugString(TF_Graph* graph,
                                                      size_t* len);

// Returns the function content in a human-readable format, with length set in
// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
//

  in_grads.reserve(input_tensors.size());
  for (int target_index = 0; target_index < input_tensors.size();
       ++target_index) {
    const auto current_grad =
        accumulated_gradients_.find(input_tensors[target_index].GetID());
    if (current_grad == accumulated_gradients_.end()) {
      if (IsDtypeTrainable(input_tensors[target_index].GetDType())) {

// things like the available devices, resource manager etc.
// TFE_Context must outlive all tensor handles created using it. In other
// words, TFE_DeleteContext() must be called after all tensor handles have
// been deleted (with TFE_DeleteTensorHandle).
//
// TODO(ashankar): Merge with TF_Session?
typedef struct TFE_Context TFE_Context;

TF_CAPI_EXPORT extern TFE_Context* TFE_NewContext(

// A create function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void* create(TF_OpKernelConstruction* ctx) {
  TF_DataType type;
  TF_Status* s = TF_NewStatus();
  TF_OpKernelConstruction_GetAttrType(ctx, "foobar", &type, s);
  TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.