// 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

  // functions (and hence the tensors they keep alive). Instead, everything
  // is deleted in ~GradientTape. Persistent GradientTapes are useful when
  // users want to compute multiple gradients over the same tape.
  explicit GradientTape(bool persistent) : persistent_(persistent) {}
  ~GradientTape() {
    for (const auto& pair : op_tape_) {