ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;

                    char* buffer, TF_Status* status) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  char* dst = buffer;
  int64_t read = 0;

  while (n > 0) {
    // Some platforms, notably macs, throw `EINVAL` if `pread` is asked to read
    // more than fits in a 32-bit integer.
    size_t requested_read_length;
    if (n > INT32_MAX)
      requested_read_length = INT32_MAX;
    else

  // 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

                             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);

  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.
  dlm_tensor->dl_tensor.strides = stride_arr->data();

  dlm_tensor->dl_tensor.byte_offset =
      0;  // TF doesn't handle the strides and byte_offsets here

        // just copy-off but includes a sum) and consideration of performance.
        //
        // 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).

      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.

                    /*expected_shape*/ make_shape({kUnknownDim, kUnknownDim}));

  TFE_DeleteOp(matmul_op);
  // TODO(bgogul): Add some death tests where status is not OK.
}

TEST_F(ShapeInferenceTest, InfersShapesFromInputTensors) {
  // Prepare some tensors for shape.
  TF_Tensor* tensor_1X6 = Int32Tensor({1, 6});
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);

//
// 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,

             : 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,