* <p><b>Context-sensitive interpretation:</b>
     * A dependency with this path type will not necessarily be placed on the class-path.
     * There are two circumstances where the dependency may nevertheless be placed somewhere else:
     * </p>
     * <ul>
     *   <li>If {@link #MODULES} path type is also set, then the dependency can be placed either on the
     *       class-path or on the module-path, but only one of those. The choice is up to the plugin,

  if (*requested_placement == '\0') {
    TF_SetStatus(
        status, TF_INTERNAL,
        "Ops must be placed on the parallel device explicitly, or their inputs "
        "first un-packed. Got an un-placed op with an input placed on the "
        "parallel device.");
    return;
  }
  TFE_Context* context = TFE_OpGetContext(original_op, status);

   * overkill for most Atomics because they are usually irregularly
   * scattered in memory and thus don't interfere much with each
   * other. But Atomic objects residing in arrays will tend to be
   * placed adjacent to each other, and so will most often share
   * cache lines (with a huge negative performance impact) without
   * this precaution.
   *
   * In part because Cells are relatively large, we avoid creating

   * overkill for most Atomics because they are usually irregularly
   * scattered in memory and thus don't interfere much with each
   * other. But Atomic objects residing in arrays will tend to be
   * placed adjacent to each other, and so will most often share
   * cache lines (with a huge negative performance impact) without
   * this precaution.
   *
   * In part because Cells are relatively large, we avoid creating

    /**
     * The ordered list of the flattened dependency nodes.
     */
    private final List<Node> nodes;

    /**
     * The file paths of all dependencies, regardless on which Java tool option those paths should be placed.
     */
    private final List<Path> paths;

    /**
     * The file paths of all dependencies, dispatched according the Java options where to place them.
     */

   * overkill for most Atomics because they are usually irregularly
   * scattered in memory and thus don't interfere much with each
   * other. But Atomic objects residing in arrays will tend to be
   * placed adjacent to each other, and so will most often share
   * cache lines (with a huge negative performance impact) without
   * this precaution.
   *
   * In part because Cells are relatively large, we avoid creating

    /**
     * The ordered list of the flattened dependency nodes.
     */
    private final List<Node> nodes;

    /**
     * The file paths of all dependencies, regardless on which Java tool option those paths should be placed.
     */
    private final List<Path> paths;

    /**
     * The file paths of all dependencies, dispatched according the Java options where to place them.
     */

  /// 3) `TF_OUT_OF_RANGE` if the read from the remote filesystem succeeded, but
  ///    the file contents do not extend past `offset` and thus nothing was
  ///    placed in `out`.
  /// 4) `TF_OK` otherwise (i.e. the read succeeded, and at least one byte was
  /// placed
  ///    in `buffer`).
  ///
  /// Caller is responsible for allocating memory for `buffer`.
  /// `buffer` will be left unchanged in case of errors.

     * nodes. The traversal order among descendant nodes (particularly aunts and nieces) are
     * determined by the {@code InsertionOrder} parameter: nieces are placed at the FRONT before
     * aunts for pre-order; while in BFS they are placed at the BACK after aunts.
     */
    private Iterator<N> topDown(Iterator<? extends N> startNodes, InsertionOrder order) {
      Deque<Iterator<? extends N>> horizon = new ArrayDeque<>();

namespace parallel_device {

using TensorHandlePtr = tensorflow::Safe_TFE_TensorHandlePtr;

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,