// 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.

  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Custom device: inputs in same custom device works.
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcustom0.get()));
  num_retvals = 1;
  executed = false;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());

                    /*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_);

      ASSERT_EQ(errors::OK, s.code()) << s.message();
      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 ctx_;

 public:

//
// 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,

  std::unique_ptr<WritableFile> same_file;
  status = env_->NewAppendableFile(filepath, &same_file);
  if (!status.ok())
    GTEST_SKIP() << "NewAppendableFile() not supported: " << status;

  const std::string more_test_data("qwer");
  EXPECT_EQ(same_file->Append(more_test_data).code(), Code::OK);
  EXPECT_EQ(same_file->Flush().code(), Code::OK);
  EXPECT_EQ(same_file->Close().code(), Code::OK);

      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());

  TFE_TensorHandle* h1_task0 = TestMatrixTensorHandle(ctx);
  std::vector<TFE_TensorHandle*> handles_task0;
  if (heavy_load_on_streaming_rpc) {
    // Send 50 tensor copy requests to simulate that there have been some RPC
    // requests been enqueued.
    for (int i = 0; i < 50; ++i) {
      handles_task0.push_back(TestMatrixTensorHandle(ctx));
    }
  }
  const char task1_name[] = "/job:localhost/replica:0/task:1/device:CPU:0";