// you want the latter, use TFE_TensorHandleBackingDeviceName. This function
// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.

                    int expected_max_outputs,
                    CancellationManager& cancellation_manager,
                    std::optional<int64_t> step_id = std::nullopt) const;

  // Blocks until the previous `StartExecute` has run `TFE_Execute` on each
  // device. If is_async=false (constructor argument) this means the ops have
  // run and have results. If is_async=true it means that all of the

                 uint64_t block_size, size_t max_bytes, uint64_t max_staleness,
                 uint64_t stat_cache_max_age, size_t stat_cache_max_entries)
    : gcs_client(gcs_client),
      compose(compose),
      block_cache_lock(),
      block_size(block_size) {
  file_block_cache = std::make_unique<RamFileBlockCache>(
      block_size, max_bytes, max_staleness,

void RecordMutation(TF_Graph* graph, const TF_Operation& op,
                    const char* mutation_type)
    TF_EXCLUSIVE_LOCKS_REQUIRED(graph->mu);

bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status)
    TF_LOCKS_EXCLUDED(session->graph->mu, session->mu);

std::string getTF_OutputDebugString(TF_Output node);

// Causes the calling thread to block till all ops dispatched in this executor
// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.

  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use

                    absl::optional<int64_t> step_id = absl::nullopt);
  // Block until the previous `StartExecute` operation has executed. Forwards
  // the status from `TFE_Execute` and returns outputs if the status is OK.
  std::vector<TensorHandlePtr> Join(TF_Status* status);

  // Block until all Ops finished running on the thread.
  void AsyncWait(TF_Status* status);

 private:
  void Run();

dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> 1712896733 +0000

bool ExtendSessionGraphHelper(TF_Session* session, TF_Status* status) {
  if (session->graph != nullptr) {
    // Take the graph lock before the session lock to avoid deadlock. This is
    // safe since session->graph does not change.
    session->graph->mu.lock();
    mutex_lock session_lock(session->mu);
    const Graph& graph = session->graph->graph;

    const string& mutation_warning = session->graph->sessions[session];

  // Add `devices` into context's device manager. Context's device manager
  // will take ownership and maintain devices' lifetime.
  virtual Status AddDevices(std::vector<std::unique_ptr<Device>> devices) = 0;

  // Block until all pending nodes are finished.
  virtual Status AsyncWait() = 0;

  // Add a function (serialized FunctionDef protocol buffer) so that it can
  // be executed as an op. Return error if the function with the same name