// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleBackingDeviceName(

TF_CAPI_EXPORT extern void TFE_ContextAsyncWait(TFE_Context* ctx,
                                                TF_Status* status);

// This function will block till the operation that produces `h` has
// completed. This is only valid on local TFE_TensorHandles. The pointer
// returned will be on the device in which the TFE_TensorHandle resides (so e.g.
// for a GPU tensor this will return a pointer to GPU memory). The pointer is

        dtype_(dtype) {}

  const ParallelDevice& device_;
  std::vector<TensorHandlePtr> tensors_;
  // Parallel tensors are immutable but compute their shape lazily unless it is
  // provided on construction. The optional has a value if the lazy computation
  // has been completed or the shape was provided on construction.
  mutable absl::optional<std::vector<int64_t>> shape_;
  const TF_DataType dtype_;
};

  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.
  DEVICE_PLACEMENT_SILENT = 2,
  // Placement policy which silently copies int32 tensors but not other dtypes.
  DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
};

  ///
  /// Replaces `dst` if it exists. In case of error, both `src` and `dst` keep
  /// the same state as before the call.
  ///
  /// Plugins:
  ///   * Must set `status` to `TF_OK` if rename was completed.
  ///   * Must set `status` to `TF_NOT_FOUND` if one of the parents entries in
  ///     either `src` or `dst` doesn't exist or if the specified `src` path
  ///     doesn't exist.