// shufflesharding and thus makes the distinguisher method of
  // associated flow schemas irrelevant.  This field has a default
  // value of 64.
  // +optional
  optional int32 queues = 1;

  // `handSize` is a small positive number that configures the
  // shuffle sharding of requests into queues.  When enqueuing a request
  // at this priority level the request's flow identifier (a string

      sort(
          insertionOrder,
          new Comparator<V>() {
            @Override
            public int compare(V left, V right) {
              // The indexes are small enough for the subtraction trick to be safe.
              return indexOfEntryWithValue(left) - indexOfEntryWithValue(right);
            }

            int indexOfEntryWithValue(V value) {

 *       SHORT is always discriminable from zero time, and always allows enough time for the small
 *       amounts of computation (creating a thread, calling a few methods, etc) needed to reach a
 *       timeout point. Similarly, a SMALL is always discriminable as larger than SHORT and smaller
 *       than MEDIUM. And so on. These constants are set to conservative values, but even so, if

threshold for this to be true is not precisely defined by law.

  If such an object file uses only numerical parameters, data
structure layouts and accessors, and small macros and small inline
functions (ten lines or less in length), then the use of the object
file is unrestricted, regardless of whether it is legally a derivative
work.  (Executables containing this object code plus portions of the

   * using a {@code LinkedHashSet} instead, at the cost of increased memory footprint, to get
   * deterministic iteration behavior.
   *
   * <p>This method is just a small convenience, either for {@code newHashSet(}{@link Arrays#asList
   * asList}{@code (...))}, or for creating an empty set then calling {@link Collections#addAll}.

    // as their length. These files are very small, so it's wasteful to allocate an 8KB buffer.
    int initialBufferSize = min(BUFFER_SIZE, max(128, Integer.highestOneBit(totalLen) * 2));
    // Starting with an 8k buffer, double the size of each successive buffer. Smaller buffers
    // quadruple in size until they reach 8k, to minimize the number of small reads for longer

      std::move(maybe_typed_outputs.value()));

  if (typed_outputs.size() > *num_outputs) {
    TF_SetStatus(status, TF_INTERNAL,
                 "The allocated output buffer was too small.");
    return;
  }

  for (int i = 0; i < typed_outputs.size(); ++i) {
    MaybeParallelTensorOwned typed_output(std::move(typed_outputs[i]));
    if (absl::holds_alternative<TensorHandlePtr>(typed_output)) {

 *
 * <p>Users of these utilities must be <i>extremely careful</i> not to mix up signed and unsigned
 * {@code int} values. When possible, it is recommended that the {@link UnsignedInteger} wrapper
 * class be used, at a small efficiency penalty, to enforce the distinction in the type system.
 *
 * <p>See the Guava User Guide article on <a
 * href="https://github.com/google/guava/wiki/PrimitivesExplained#unsigned-support">unsigned

 *
 * <p>Users of these utilities must be <i>extremely careful</i> not to mix up signed and unsigned
 * {@code int} values. When possible, it is recommended that the {@link UnsignedInteger} wrapper
 * class be used, at a small efficiency penalty, to enforce the distinction in the type system.
 *
 * <p>See the Guava User Guide article on <a
 * href="https://github.com/google/guava/wiki/PrimitivesExplained#unsigned-support">unsigned

   *     new ImmutableBiMap.Builder<String, Integer>()
   *         .put("one", 1)
   *         .put("two", 2)
   *         .put("three", 3)
   *         .buildOrThrow();
   * }</pre>
   *
   * <p>For <i>small</i> immutable bimaps, the {@code ImmutableBiMap.of()} methods are even more
   * convenient.
   *
   * <p>By default, a {@code Builder} will generate bimaps that iterate over entries in the order