// 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

//
// If the device is successfully registered, `status` is set to TF_OK. Otherwise
// the device is not usable. In case of a bad status, `device.delete_device` is
// still called on `device_info` (i.e. the caller does not retain ownership).
//
// This API is highly experimental, and in particular is expected to change when
// it starts supporting operations with attributes and when tf.function support
// is added.

// The returned string is heap-allocated, and caller should call free() on it.
TF_CAPI_EXPORT extern const char* TF_GraphDebugString(TF_Graph* graph,
                                                      size_t* len);

// Returns the function content in a human-readable format, with length set in
// `len`. The format is subject to change in the future.
// The returned string is heap-allocated, and caller should call free() on it.
//

      new GradientTape<Gradient, BackwardFunction, TapeTensor>(false));
  AccumulatorCallState& call_state = call_state_.top();
  call_state.backward_tape = tape.get();
  auto pop_backward_tape =
      gtl::MakeCleanup([&call_state] { call_state.backward_tape = nullptr; });
  std::vector<Gradient*> forwardprop_aids;
  std::vector<int64_t> sources;
  std::unordered_set<int64_t> sources_set;

// Clears the internal caches in the TFE context. Useful when reseeding random
// ops.
TF_CAPI_EXPORT extern void TFE_ContextClearCaches(TFE_Context* ctx);

// Sets a thread-local device placement policy. After this call, other calls to
// TFE_Execute in the same thread will use the device policy specified here
// instead of the device policy used to construct the context. This has no
// effect on the device policy used by other program threads.

  //===--------------------------------------------------------------------===//
#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 Status EnableCollectiveOps(const ServerDef& server_def) = 0;

  /// returns total bytes read ( -1 in case of errors ). This method will set
  /// `status` to:
  ///
  /// 1) The error from the remote filesystem, if the read from the remote
  ///    filesystem failed.
  /// 2) `TF_FAILED_PRECONDITION` if the read from the remote filesystem
  /// succeeded,
  ///    but the read returned a partial block, and the LRU cache contained a