TFE_DeleteTensorHandle(h1_task1);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_DeleteContext(ctx);

  TF_DeleteStatus(status);

  TFE_DeleteTensorHandle(h0_task0);
  TFE_DeleteTensorHandle(retvals[0]);

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

// Read the value of variable `var` and save it into `out_value`.

                                                  TF_Status* status) {
  tensorflow::EagerContext* context =
      tensorflow::ContextFromInterface(tensorflow::unwrap(ctx));
  auto collective_executor_handle = context->GetCollectiveExecutorHandle();
  collective_executor_handle->get()->StartAbort(status->status);
}

TF_CAPI_EXPORT extern void TFE_CollectiveOpsCheckPeerHealth(
    TFE_Context* ctx, const char* task, int64_t timeout_in_ms,

  EXPECT_THAT(TF_Message(status.get()), HasSubstr("assertion failed"));

  // Note that future collectives with the same context do not work at the
  // moment; once canceled, the collective executor requires the program to be
  // restarted / context to be reset.
}

TEST(PARALLEL_DEVICE_LIB, TestDifferentShapes) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(

        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(

  TFE_DeleteTensorHandle(h0);
  TFE_DeleteTensorHandle(h1);
  TFE_DeleteTensorHandle(h2);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TFE_ContextRemoveFunction(ctx, "AddVariablesFunction", status);
  TFE_DeleteContext(ctx);

                                        TF_Status* status) {
  status->status = executor->executor()->WaitForAllPendingNodes();
}

void TFE_ExecutorClearError(TFE_Executor* executor) {
  executor->executor()->ClearError();
}

void TFE_ContextSetExecutorForThread(TFE_Context* ctx, TFE_Executor* executor) {
  tensorflow::unwrap(ctx)->SetExecutorForThread(executor->executor());
}

  for (auto* h : handles_task2) {
    TFE_DeleteTensorHandle(h);
  }

  TFE_DeleteOp(matmul);

  TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_GetCode(status), TF_OK) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  if (func) {
    TFE_ContextRemoveFunction(ctx, "MatMulFunction", status);
  }
  TFE_DeleteContext(ctx);

  TF_DeleteFunction(fn);

  for (bool async : {false, true, false}) {
    TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/async, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

    TF_Tensor* t =

    if (state.iterations() >= state.max_iterations && async) {
      TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
      TFE_ExecutorWaitForAllPendingNodes(executor, status);
      ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
      TFE_DeleteExecutor(executor);
    }
  }
  TFE_DeleteOp(matmul);
  TFE_DeleteTensorHandle(m);
  TFE_DeleteContext(ctx);