* existing values in the execution request that are controlled by the toolchains. Hence, it is expected that this
     * method is called on a new/empty execution request before the caller mutates it to fit its needs.
     *
     * @param request The execution request to populate, must not be {@code null}.

 *
 * <p>{@code ByteSink} provides two kinds of methods:
 *
 * <ul>
 *   <li><b>Methods that return a stream:</b> These methods should return a <i>new</i>, independent
 *       instance each time they are called. The caller is responsible for ensuring that the
 *       returned stream is closed.
 *   <li><b>Convenience methods:</b> These are implementations of common operations that are

 *
 * <p>{@code CharSink} provides two kinds of methods:
 *
 * <ul>
 *   <li><b>Methods that return a writer:</b> These methods should return a <i>new</i>, independent
 *       instance each time they are called. The caller is responsible for ensuring that the
 *       returned writer is closed.
 *   <li><b>Convenience methods:</b> These are implementations of common operations that are

  // A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s
  // return a bad status the caller must run all of the `Join`s or any future

 *
 * <p>{@code CharSource} provides two kinds of methods:
 *
 * <ul>
 *   <li><b>Methods that return a reader:</b> These methods should return a <i>new</i>, independent
 *       instance each time they are called. The caller is responsible for ensuring that the
 *       returned reader is closed.
 *   <li><b>Convenience methods:</b> These are implementations of common operations that are

   * SES.schedule is called. Therefore Fire.run has to check for null. However, it should be visible
   * if Fire.run is called by delegate.addListener since addListener is called after the assignment
   * to timer, and importantly this is the main situation in which we need to be able to see the
   * write.
   *
   * 2. visibility of the writes to an afterDone() call triggered by cancel():
   *

 *
 * <p>{@code ByteSource} provides two kinds of methods:
 *
 * <ul>
 *   <li><b>Methods that return a stream:</b> These methods should return a <i>new</i>, independent
 *       instance each time they are called. The caller is responsible for ensuring that the
 *       returned stream is closed.
 *   <li><b>Convenience methods:</b> These are implementations of common operations that are

  //===--------------------------------------------------------------------===//
#if !defined(IS_MOBILE_PLATFORM)
  // Set up a multi-client distributed execution environment. Must be called on
  // all tasks in the cluster.
  // This call internally coordinates with other tasks to initialize the eager
  // context and TF server for multi-client execution.
  virtual absl::Status EnableCollectiveOps(const ServerDef& server_def) = 0;

   */
  final void init() {
    /*
     * requireNonNull is safe because this is called from the constructor after `futures` is set but
     * before releaseResources could be called (because we have not yet set up any of the listeners
     * that could call it, nor exposed this Future for users to call cancel() on).
     */
    requireNonNull(futures);

    // Corner case: List is empty.
    if (futures.isEmpty()) {

    // JVM shutdown will not be prevented from exiting after this service has stopped or failed.
    // Technically this listener is added after start() was called so it is a little gross, but it
    // is called within doStart() so we know that the service cannot terminate or fail concurrently
    // with adding this listener so it is impossible to miss an event that we are interested in.
    addListener(