// threadpool, so that we avoid the possibility of running the runner_ in the
  // 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

  EXPECT_EQ(TF_GetCode(status.get()), TF_INVALID_ARGUMENT);
  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) {

  if (fflush(posix_file->handle) != 0)
    TF_SetStatusFromIOError(status, errno, posix_file->filename);
}

static void Sync(const TF_WritableFile* file, TF_Status* status) {
  // For historical reasons, this does the same as `Flush` at the moment.
  // TODO(b/144055243): This should use `fsync`/`sync`.
  Flush(file, status);
}

static void Close(const TF_WritableFile* file, TF_Status* status) {

                             serialized_update.size(), status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // The device of var_handle0 is local device which is the same before and
  // after cluster update. Remove resource with valid device should succeed.
  TFE_Op* op = TFE_NewOp(ctx, "DestroyResourceOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

      gtl::ArraySlice<AbstractTensorHandle*> output_gradients,
      absl::Span<AbstractTensorHandle*> result) const override;

  // Builds a tensor filled with ones with the same shape and dtype as `t`.
  Status BuildOnesLike(const TapeTensor& t,
                       AbstractTensorHandle** result) const override;

  // Looks up the ID of a Gradient.

//
// DeviceThread itself is thread-safe, in that StartExecute will block if there
// is a pending execution. Since StartExecute is equivalent to grabbing a lock,
// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:
  // Starts a background thread waiting for `StartExecute`.
  explicit DeviceThread(const std::string& device, const bool is_async,

    const char* tensor_handle_device =
        TFE_TensorHandleDeviceName(input, status);
    if (TF_GetCode(status) != TF_OK) return;
    if (named_device->name() == tensor_handle_device) {
      // We assume that any tensors already placed on this device are
      // ParallelTensors.
      typed_inputs.emplace_back(reinterpret_cast<ParallelTensor*>(
          TFE_TensorHandleDevicePointer(input, status)));

             : absl::OkStatus();
}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(
    const TF_Graph* fn_body, const char* fn_name, int ninputs,

      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Run a collective sum, so each component should now be the same.
  TensorHandlePtr reduced(
      CollectiveSum(context.get(), parallel_value.get(), 2, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

}

// Set server_def on the context, possibly updating it.
// TODO(b/291142876) Simplify TFE_ContextSetServerDefWithTimeoutAndRetries and
// TFE_ContextUpdateServerDefWithTimeout to be simple wrappers around the same
// C++ function.
// Retries are used for CreateContext calls, which is used in
// ParameterServerStrategy initialization to be robust to worker preemption.