TF_DeleteGraph);

  TF_SessionOptions* opts = TF_NewSessionOptions();
  TF_Session* sess = TF_NewSession(graph.get(), opts, s);
  TF_DeleteSessionOptions(opts);

  const string gpu_device_name = GPUDeviceName(sess);
  TF_DeleteSession(sess, s);
  CHECK_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  return gpu_device_name;
}

# This is the same as the official TensorFlow builds.
# See https://developer.nvidia.com/cuda-gpus#compute
# `compute_XY` enables PTX embedding in addition to SASS. PTX
# is forward compatible beyond the current compute capability major
# release while SASS is only forward compatible inside the current
# major release. Example: sm_80 kernels can run on sm_89 GPUs but

  if (prefix == nullptr || strlen(prefix) == 0)
    prefix = "/job:localhost/replica:0/task:0";

  tensorflow::SessionOptions sess_options;
  (*sess_options.config.mutable_device_count())["CPU"] = num_cpus;
  status->status =
      tensorflow::DeviceFactory::AddCpuDevices(sess_options, prefix, &devices);

  // Remove the device that has the host device name since host device is alreay
  // in an initialized context.

    `CUDA_ILLEGAL_ADDRESS` failures.

    Google discovered in mid-December 2017 that the PTX-to-SASS compiler in CUDA
    9 and CUDA 9.1 sometimes does not properly compute the carry bit when
    decomposing 64-bit address calculations with large offsets (e.g. `load [x +
    large_constant]`) into 32-bit arithmetic in SASS.

    As a result, these versions of `ptxas` miscompile most XLA programs which