downloaded_block = true;
          block->state = FetchState::FINISHED;
        } else {
          block->state = FetchState::ERROR;
        }
        block->cond_var.SignalAll();
        return;
      case FetchState::FETCHING:
        block->cond_var.WaitWithTimeout(&block->mu, absl::Minutes(1));
        if (block->state == FetchState::FINISHED) {
          return TF_SetStatus(status, TF_OK, "");

  int32_t init_timeout_in_ms = 300000;
  TFE_Context* ctx_0 =
      CreateContext(serialized_server_def_0,
                    /*isolate_session_state=*/false, init_timeout_in_ms);
  TFE_Context* ctx_1 =
      CreateContext(serialized_server_def_1,
                    /*isolate_session_state=*/false, init_timeout_in_ms);

  // Remote device on `worker2`.
  const char remote_device[] = "/job:localhost/replica:0/task:2/device:CPU:0";

  typedef std::pair<std::string, size_t> Key;

  /// \brief The state of a block.
  ///
  /// A block begins in the CREATED stage. The first thread will attempt to read
  /// the block from the filesystem, transitioning the state of the block to
  /// FETCHING. After completing, if the read was successful the state should
  /// be FINISHED. Otherwise the state should be ERROR. A subsequent read can

      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;

    {
      tensorflow::mutex_lock l(execution_mutex_);
      while (execution_state_ == ExecutionState::kIdle ||
             execution_state_ == ExecutionState::kHasResult) {
        start_execute_.wait(l);
      }
      if (execution_state_ == ExecutionState::kShuttingDown) {
        return;
      } else if (execution_state_ == ExecutionState::kReadyToExecute) {
        // op_outputs_ may have been std::moved

      TF_SetPayload(&s, tensorflow::CoordinationErrorPayloadKey().data(),
                    error.SerializeAsString().c_str());
    }
    *state_iter = std::move(s);
    ++state_iter;
  }
  status->status = absl::OkStatus();
}

void TFE_WaitAtBarrier(TFE_Context* ctx, const char* barrier_id,
                       int64_t barrier_timeout_in_ms, TF_Status* status) {

TFE_Context* CreateContext(const std::string& serialized_server_def,
                           bool isolate_session_state,
                           int64_t init_timeout_in_ms) {
  TF_Status* status = TF_NewStatus();
  TFE_ContextOptions* opts = TFE_NewContextOptions();
  opts->session_options.options.config.set_isolate_session_state(
      isolate_session_state);
  TFE_ContextOptionsSetAsync(opts, static_cast<unsigned char>(false));

      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of

  // Advance the fake timestamp so that "a" becomes stale via its first block.
  env->SetNowSeconds(now + 2);
  // The pruning thread periodically compares env->NowSeconds() with the oldest
  // block's timestamp to see if it should evict any files. At the current fake
  // timestamp of `now` + 2, file "a" is stale because its first block is stale,
  // but file "b" is not stale yet. Thus, once the pruning thread wakes up (in

//    }

#define GO_MAJOR 0
#define GO_MINOR 0
#define GO_PATCH 1

#ifdef __cplusplus
extern "C" {
#endif

// TF_TriState is the C API typedef for tri-state.
typedef enum TF_TriState {
  TF_TriState_Default = 0,
  TF_TriState_Off,
  TF_TriState_On,
} TF_TriState;

// TF_GrapplerItem represents a combination of a graph, one of more fetch nodes,