tensorflow::Session* session;
  TF_Graph* const graph;

  tensorflow::mutex mu TF_ACQUIRED_AFTER(TF_Graph::mu);
  int last_num_graph_nodes;

  // If true, TF_SessionRun and similar methods will call
  // ExtendSessionGraphHelper before running the graph (this is the default
  // public behavior). Can be set to false if the caller needs to call
  // ExtendSessionGraphHelper manually.

test:nonpip --config=nonpip_filters -- //tensorflow/... -//tensorflow/python/integration_testing/... -//tensorflow/compiler/tf2tensorrt/... -//tensorflow/core/tpu/... -//tensorflow/lite/... -//tensorflow/tools/toolchains/...

# "pip tests" run a similar suite of tests the "nonpip" tests, but do something
# odd to attempt to validate the quality of the pip package. The wheel is
# installed into a virtual environment, and then that venv is used to run all

// 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.

test:nonpip --config=nonpip_filters -- //tensorflow/... -//tensorflow/python/integration_testing/... -//tensorflow/compiler/tf2tensorrt/... -//tensorflow/core/tpu/... -//tensorflow/lite/... -//tensorflow/tools/toolchains/...

# "pip tests" run a similar suite of tests the "nonpip" tests, but do something
# odd to attempt to validate the quality of the pip package. The wheel is
# installed into a virtual environment, and then that venv is used to run all

// Description of the TensorFlow op to execute.
//
// Assumes that the provided 'ctx' outlives the returned TFE_Op, i.e.,
// TFE_DeleteOp() is called before TFE_DeleteContext().
//
// Very similar to TF_OperationDescription with some differences:
// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.

  /// cache is constructed. It returns total bytes read ( -1 in case of errors
  /// ). The `status` should be `TF_OK` as long as the read from the remote
  /// filesystem succeeded (similar to the semantics of the read(2) system
  /// call).
  typedef std::function<int64_t(const std::string& filename, size_t offset,
                                size_t buffer_size, char* buffer,

  ///
  /// This is accomplished by traversing directory tree rooted at `path` and
  /// deleting entries as they are encountered, from leaves to root. Each plugin
  /// is free to choose a different approach which obtains similar results.
  ///
  /// On successful deletion, `status` must be `TF_OK` and `*undeleted_files`
  /// and `*undeleted_dirs` must be 0. On unsuccessful deletion, `status` must

### Major Features and Improvements

*   <INSERT MAJOR FEATURE HERE, USING MARKDOWN SYNTAX>
*   <IF RELEASE CONTAINS MULTIPLE FEATURES FROM SAME AREA, GROUP THEM TOGETHER>

### Bug Fixes and Other Changes

* <SIMILAR TO ABOVE SECTION, BUT FOR OTHER IMPORTANT CHANGES / BUG FIXES>
* <IF A CHANGE CLOSES A GITHUB ISSUE, IT SHOULD BE DOCUMENTED HERE>
* <NOTES SHOULD BE GROUPED PER AREA>

* GPU

    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }
    return absl::OkStatus();
  }
  ~IdentityNGradientFunction() override {}
};

    int noutputs, const TF_Output* outputs, const char* const* output_names,
    const TF_FunctionOptions* opts, const char* description, TF_Status* status);

// Similar to TF_GraphToFunction but allows specifying control outputs of the
// function.
//
//  The arguments of TF_GraphToFunction have the same meaning, but the new
//  arguments are as follows:
//