// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.
// c) ConfigProto.device_count is set to `num_cpu_devices`.
TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig(
    unsigned char enable_xla_compilation, unsigned char gpu_memory_allow_growth,
    unsigned int num_cpu_devices);

// Create a serialized tensorflow.RunOptions proto, where RunOptions.trace_level

// If source and destination are the same device, then this creates a new handle
// that shares the underlying buffer. Otherwise, it currently requires at least
// one of the source or destination devices to be CPU (i.e., for the source or
// destination tensor to be placed in host memory).
// If async execution is enabled, the copy may be enqueued and the call will
// return "non-ready" handle. Else, this function returns after the copy has

  // Find and return a function record added by its name.
  virtual core::RefCountPtr<FunctionRecord> FindRecord(
      const string& name) const = 0;

  // Return the ParsedName of Host CPU device.
  virtual const DeviceNameUtils::ParsedName& HostCPUParsedName() const = 0;
  virtual const string& HostCPUName() const = 0;

  // Configure soft device placement policy.

    }
  });
  if (!s.ok()) {
    return s;
  }

  std::unordered_map<int64_t, int64_t> gradients_size;
  // TODO(apassos) multiple threads could be dequeuing from op_stack at the same
  // time, for better CPU backprop performance.
  VLOG(1) << "Initial stack:";
  if (VLOG_IS_ON(1)) {
    for (auto t : op_stack) {
      VLOG(1) << "  " << t;
    }
  }
  while (!op_stack.empty()) {