* appropriate {@link DiscreteDomain} to {@link ContiguousSet#create}.
 *
 * <h3>Types of ranges</h3>
 *
 * <p>Each end of the range may be bounded or unbounded. If bounded, there is an associated
 * <i>endpoint</i> value, and the range is considered to be either <i>open</i> (does not include the

  @CheckForNull
  abstract C greatestValueBelow(DiscreteDomain<C> domain);

  /*
   * The canonical form is a BelowValue cut whenever possible, otherwise ABOVE_ALL, or
   * (only in the case of types that are unbounded below) BELOW_ALL.
   */
  Cut<C> canonical(DiscreteDomain<C> domain) {
    return this;
  }

  // note: overridden by {BELOW,ABOVE}_ALL
  @Override
  public int compareTo(Cut<C> that) {

 * appropriate {@link DiscreteDomain} to {@link ContiguousSet#create}.
 *
 * <h3>Types of ranges</h3>
 *
 * <p>Each end of the range may be bounded or unbounded. If bounded, there is an associated
 * <i>endpoint</i> value, and the range is considered to be either <i>open</i> (does not include the

  @CheckForNull
  abstract C greatestValueBelow(DiscreteDomain<C> domain);

  /*
   * The canonical form is a BelowValue cut whenever possible, otherwise ABOVE_ALL, or
   * (only in the case of types that are unbounded below) BELOW_ALL.
   */
  Cut<C> canonical(DiscreteDomain<C> domain) {
    return this;
  }

  // note: overridden by {BELOW,ABOVE}_ALL
  @Override
  public int compareTo(Cut<C> that) {

import java.util.concurrent.TimeUnit;
import org.checkerframework.checker.nullness.qual.Nullable;

/**
 * An unbounded {@linkplain BlockingQueue blocking queue} that uses the same ordering rules as class
 * {@link PriorityQueue} and supplies blocking retrieval operations. While this queue is logically
 * unbounded, attempted additions may fail due to resource exhaustion (causing {@code

   *
   * <p><b>Warning:</b> If the input future does not already implement {@code ListenableFuture}, the
   * returned future will emulate {@link ListenableFuture#addListener} by taking a thread from an
   * internal, unbounded pool at the first call to {@code addListener} and holding it until the
   * future is {@linkplain Future#isDone() done}.
   *
   * <p>Prefer to create {@code ListenableFuture} instances with {@link SettableFuture}, {@link

   *
   * <p><b>Warning:</b> If the input future does not already implement {@code ListenableFuture}, the
   * returned future will emulate {@link ListenableFuture#addListener} by taking a thread from an
   * internal, unbounded pool at the first call to {@code addListener} and holding it until the
   * future is {@linkplain Future#isDone() done}.
   *
   * <p>Prefer to create {@code ListenableFuture} instances with {@link SettableFuture}, {@link

   * Returns the minimal range with the given boundary types for which all values in this set are
   * {@linkplain Range#contains(Comparable) contained} within the range.
   *
   * <p>Note that this method will return ranges with unbounded endpoints if {@link BoundType#OPEN}
   * is requested for a domain minimum or maximum. For example, if {@code set} was created from the
   * range {@code [1..Integer.MAX_VALUE]} then {@code set.range(CLOSED, OPEN)} must return {@code

   * Returns the minimal range with the given boundary types for which all values in this set are
   * {@linkplain Range#contains(Comparable) contained} within the range.
   *
   * <p>Note that this method will return ranges with unbounded endpoints if {@link BoundType#OPEN}
   * is requested for a domain minimum or maximum. For example, if {@code set} was created from the
   * range {@code [1..Integer.MAX_VALUE]} then {@code set.range(CLOSED, OPEN)} must return {@code

  // 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;
}

// Consumes references to the tensors in the gradient_tensors list and returns