const void** buf, size_t* len);

// TODO(josh11b,mrry): Change Session to be able to use a Graph*
// directly, instead of requiring us to serialize to a GraphDef and
// call Session::Extend().
bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status) {
  if (session->graph != nullptr) {
    // Take the graph lock before the session lock to avoid deadlock. This is

           /*dep=*/{"assert"}},
      });
  TF_ASSERT_OK(ContextFromInterface(unwrap(context.get()))
                   ->AddFunctionDef(assert_and_collective));

  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> call_op(
      TFE_NewOp(context.get(), "AssertAndCollective", status.get()),
      TFE_DeleteOp);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
  std::unique_ptr<ParallelTensor> reduced_values =

  CheckRemoteMatMulExecutesOK(ctx, remote_device_name, local_device_name);

  // Adding a non-existent remote worker to cluster def. This should cause the
  // UpdateServerDef call to fail.
  tensorflow::ClusterDef* cluster_def = server_def.mutable_cluster();
  tensorflow::JobDef* job_def = cluster_def->mutable_job(0);
  int port = tensorflow::testing::PickUnusedPortOrDie();

  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name) {
  return new TF_AttrBuilder(op_name);
}

    if (!enabled_) {
      return absl::OkStatus();
    }
    if (first_call_) {
      first_call_ = false;
      return absl::OkStatus();
    }
    return tensorflow::errors::Internal("Graph pass runs for more than once!");
  }

 private:
  bool first_call_ = true;
};

// After the graph pass is registered, it takes effect globally and can affect

struct TF_VariableInfo {
  TF_VariableInfo() = delete;
  // TF_VariableInfo is constructed here by TensorFlow, and will be passed to
  // plugin as a opaque pointer. Plugin will need to call C APIs below to
  // operate on TF_VariableInfo (such as allocate temp tensor for the `var` held
  // by the underlying tensorflow::VariableInfo.
  TF_VariableInfo(int index, const std::string& name, tensorflow::Var* var) {

  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;

  //   Outputs:
  std::vector<TensorHandlePtr> op_outputs_ TF_GUARDED_BY(execution_mutex_);
  // TF_Status is an incomplete type and so can't be stack allocated. To avoid
  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);

// license that can be found in the LICENSE file.

// This .cc file will be automatically compiled by the go tool and
// included in the package.

#include <string>
#include "main.h"

std::string Caller::call() {
	if (callback_ != 0)
		return callback_->run();
	return "";

      case FetchState::ERROR:
        // TF_FALLTHROUGH_INTENDED
      case FetchState::CREATED:
        block->state = FetchState::FETCHING;
        block->mu.Unlock();  // Release the lock while making the API call.
        block->data.clear();
        block->data.resize(block_size_, 0);
        int64_t bytes_transferred;
        bytes_transferred = block_fetcher_(key.first, key.second, block_size_,