ExecutionState::kIdle;
  // Tells the worker thread that there is new work.
  tensorflow::condition_variable start_execute_;
  // The worker thread notifies that work has finished.
  tensorflow::condition_variable finished_execute_;
  // Notifies a StartExecute that the previous Join has finished.
  tensorflow::condition_variable finished_join_;

  TFE_TensorHandle* h0 = TestVariable(ctx, 1.0, task1_name);
  TFE_TensorHandle* h1 = TestVariable(ctx, 2.0, task2_name);
  TFE_TensorHandle* h2 = TestVariable(ctx, 3.0, task0_name);

  // Add a sync point to make sure that variables have been initialized
  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* tensor(size_t index) const { return tensors_[index].get(); }

  // If the `shape` argument to `FromTensorHandles` is specified, returns that.
  //
  // Otherwise if all of the tensors have the same shape, returns that via the
  // `shape` output argument. This blocks waiting for async tensors, may return
  // a delayed bad status encountered during async execution, and will return a

  ExpectScalarEq<float>(result_components[1].get(), 3.);
}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

void RegisterCollectiveMulFunction(TFE_Context* context,

        //
        // TODO(allenl): There may be smarter ways to do this copy in some
        // cases, i.e. with a collective broadcast. We'll need to be careful
        // about things that are taken as inputs on the host or on their
        // existing device (for multi-device functions).
        std::unique_ptr<ParallelTensor> parallel_tensor(
            parallel_device.CopyToParallelDevice(

                f'Log fragment:\n'
                f'```\n{log_fragment}\n```\n'
                f'{"=" * 140}')
        break
      k -= 1

    # A low-effort attempt to find the bazel command that triggered the
    # invocation.
    bazel_comm_min_line_i = (previous_end_line if previous_end_line is not None
                             else 0)
    while k > bazel_comm_min_line_i:
      backtrack_line = log_lines[k]

[tools](https://www.tensorflow.org/resources/tools),
[libraries](https://www.tensorflow.org/resources/libraries-extensions), and
[community](https://www.tensorflow.org/community) resources that lets
researchers push the state-of-the-art in ML and developers easily build and
deploy ML-powered applications.

TensorFlow was originally developed by researchers and engineers working within

    variable->Assign(context, initial_value.get(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  }

  // Read from the variable and verify that we have a parallel tensor.
  {
    TensorHandlePtr read = variable->Read(context, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    std::array<TensorHandlePtr, 2> components;

                                    fn_name, "'");
    output_tensors->emplace_back(node, idx);
  }
  return absl::OkStatus();
}

// Populates `body_nodes` with the nodes that will become function's body.
// Performs various checks.
Status ComputeBodyNodes(
    const TF_Graph* fn_body, const char* fn_name, int num_opers,
    const TF_Operation* const* opers,

    if (TF_GetCode(status) != TF_OK) {
      TF_DeleteStatus(status);
      return "";
    }
    TF_DeleteStatus(status);

    // We add a random value into `test_dir` to ensures that two consecutive
    // runs are unlikely to clash.
    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_int_distribution<> distribution;
    std::string rng_val = std::to_string(distribution(gen));