// this to be called.
TF_CAPI_EXPORT void TF_InitMain(const char* usage, int* argc, char*** argv);

// Platform-specific implementation to return an unused port. (This should used
// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.

class ParallelTensor;
class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ASSERT_TF_OK(status_);
  auto length = tf_read_only_memory_region::Length(region);
  ASSERT_EQ(length, content_view.length());
  auto data =
      static_cast<const char*>(tf_read_only_memory_region::Data(region));
  ASSERT_TRUE(CompareSubString(0, content_view.length(), data, length));

  tf_read_only_memory_region::Cleanup(region);
  delete region;
}

  TFE_OpSetCancellationManager(op_.get(), wrap(cancellation_manager_), status);
  if (TF_GetCode(status) != TF_OK) return;

  // unwrap op_ and set step_id only if valid step id value was set.
  // Currently only required for non-TFRT use cases, e.g., EagerOp.
  if (step_id_.has_value()) {
    tensorflow::unwrap(op_.get())->SetStepId(step_id_.value());
  }

Windows)*:

```
$ pip install tensorflow
```

Other devices (DirectX and MacOS-metal) are supported using
[Device plugins](https://www.tensorflow.org/install/gpu_plugins#available_devices).

A smaller CPU-only package is also available:

```
$ pip install tensorflow-cpu
```

To update TensorFlow to the latest version, add `--upgrade` flag to the above
commands.

*Nightly binaries are available for testing using the

  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the
  // only reasonable way to make variables with no aliasing using the eager C
  // API.
  std::string no_sharing = "cd2c89b7-88b7-44c8-ad83-06c2a9158347";
  TFE_OpSetAttrString(op.get(), "shared_name", no_sharing.c_str(),

//
// Currently this is an error, and un-packing ParallelTensors must be performed
// explicitly by running a TPUReplicatedOutput operation on the parallel device.
//
// TODO(allenl): There are some use-cases that are only supported by copying to
// host at the moment (e.g. debug print on a tensor, .numpy(), etc.). We either
// need to return something here or address these use-cases one by one.

// other test cases. Define a static variable to switch it on and off.
bool GraphErrorInjectionPass::enabled_ = false;

// Test to ensure that a registered graph optimization pass is only executed
// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the