* java.util.concurrent.atomic} package specification for description of the properties of atomic
 * variables. An {@code AtomicDouble} is used in applications such as atomic accumulation, and
 * cannot be used as a replacement for a {@link Double}. However, this class does extend {@code
 * Number} to allow uniform access by tools and utilities that deal with numerically-based classes.
 *

  // TODO(srbs): It seems like this is used only for performance optimization
  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.
  // Revisit and fix.
  return 1;
}

 *   <li>values automatically wrapped in {@linkplain WeakReference weak} or {@linkplain
 *       SoftReference soft} references
 *   <li>notification of evicted (or otherwise removed) entries
 *   <li>accumulation of cache access statistics
 * </ul>
 *
 * <p>These features are all optional; caches can be created using all or none of them. By default,

 *   <li>values automatically wrapped in {@linkplain WeakReference weak} or {@linkplain
 *       SoftReference soft} references
 *   <li>notification of evicted (or otherwise removed) entries
 *   <li>accumulation of cache access statistics
 * </ul>
 *
 * <p>These features are all optional; caches can be created using all or none of them. By default,

    *   Add support for `add_metric` in the graph function mode.
    *   Updating cosine similarity loss - removed the negate sign from cosine
        similarity.
    *   Changed default for gradient accumulation for TPU embeddings to true.
    *   Adds summary trace API for collecting graph and profile information.
    *   The `precision_mode` argument to `TrtGraphConverter` is now case
        insensitive.

    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.addAll(MANY_VALUES);
    Stats stats = accumulator.snapshot();
    accumulator.add(999.999); // should do nothing to the snapshot
    return stats;
  }

  private static Stats buildLongManyValuesStatsSnapshot() {
    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.addAll(LONG_MANY_VALUES);

  public void testCountOverflow_doesNotThrow() {
    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.add(ONE_VALUE);
    for (int power = 1; power < Long.SIZE - 1; power++) {
      accumulator.addAll(accumulator.snapshot());
    }
    // Should overflow without throwing.
    accumulator.addAll(accumulator.snapshot());
    assertThat(accumulator.count()).isLessThan(0L);
  }

  public void testMean() {

    PairedStatsAccumulator accumulator = new PairedStatsAccumulator();
    accumulator.add(ONE_VALUE, OTHER_ONE_VALUE);
    for (int power = 1; power < Long.SIZE - 1; power++) {
      accumulator.addAll(accumulator.snapshot());
    }
    // Should overflow without throwing.
    accumulator.addAll(accumulator.snapshot());
    assertThat(accumulator.count()).isLessThan(0L);
  }

    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.addAll(MANY_VALUES);
    Stats stats = accumulator.snapshot();
    accumulator.add(999.999); // should do nothing to the snapshot
    return stats;
  }

  private static Stats buildLongManyValuesStatsSnapshot() {
    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.addAll(LONG_MANY_VALUES);

  public void testCountOverflow_doesNotThrow() {
    StatsAccumulator accumulator = new StatsAccumulator();
    accumulator.add(ONE_VALUE);
    for (int power = 1; power < Long.SIZE - 1; power++) {
      accumulator.addAll(accumulator.snapshot());
    }
    // Should overflow without throwing.
    accumulator.addAll(accumulator.snapshot());
    assertThat(accumulator.count()).isLessThan(0L);
  }

  public void testMean() {