/*input_tensors*/ {},
                    /*expected_shape*/ make_shape({3, kUnknownDim}));

  // Infer shape when some dimensions are unknown.
  CheckOutputShapes(
      matmul_op,
      /*input_shapes*/ {make_shape({kUnknownDim, 2}), make_shape({2, 4})},
      /*input_tensors*/ {},
      /*expected_shape*/ make_shape({kUnknownDim, 4}));

  // Infer shape when everything is unknown.

      shape.dims = nullptr;
      continue;
    }
    shape.dims = new int64_t[shape.num_dims];
    for (size_t j = 0; j < shape.num_dims; ++j) {
      shape.dims[j] = c.Value(c.Dim(shape_handle, j));
    }
  }
  if (output_shapes != nullptr) *output_shapes = output_shapes_result;

  // TODO(bgogul): Set output_resource_shapes_and_types.
}

  std::vector<int64_t>* shape_arr = &tf_dlm_tensor_ctx->shape;
  std::vector<int64_t>* stride_arr = &tf_dlm_tensor_ctx->strides;
  shape_arr->resize(ndim);
  stride_arr->resize(ndim, 1);
  for (int i = 0; i < ndim; i++) {
    (*shape_arr)[i] = tensor->dim_size(i);
  }
  for (int i = ndim - 2; i >= 0; --i) {
    (*stride_arr)[i] = (*shape_arr)[i + 1] * (*stride_arr)[i + 1];
  }

    std::vector<PartialTensorShape> shapes;
    shapes.reserve(num_values);
    for (int i = 0; i < num_values; ++i) {
      if (num_dims[i] < 0) {
        shapes.emplace_back();
      } else {
        shapes.emplace_back(ArraySlice<int64_t>(
            reinterpret_cast<const int64_t*>(dims[i]), num_dims[i]));
      }
    }
    op_->node_builder.Attr(attr_name, shapes);
    return absl::OkStatus();
  }

  auto outputs = parallel_device.Join(
      /*expected_output_shapes=*/{PartialTensorShape({})}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  const std::vector<int64_t>* shape;
  Status s = handles[0]->Shape(&shape);
  ASSERT_TRUE(s.ok());
  EXPECT_EQ(0, shape->size());
}

//     if you want shape inference to consider the input tensors of the
//     op for shape inference.
//   - The types need not be set in `input_shapes` as it is not used.
//   - The number of `input_tensors` should be the same as the number of items
//     in `input_shapes`.
//
// The results are returned in `output_shapes` and
// `output_resource_shapes_and_types`. The caller is responsible for freeing the

  // devices of a ParallelDevice. If called, ParallelTensor::Shape inspects
  // `components` to determine a shape.
  static std::unique_ptr<ParallelTensor> FromTensorHandles(
      const ParallelDevice& parallel_device,
      std::vector<TensorHandlePtr> components, TF_Status* status);
  // Uses the provided shape without additional checks, which avoids blocking
  // when ParallelTensor::Shape is called.

      if (combined_shape.dims() < 0 ||
          combined_shape.dims() != component_shape.dims()) {
        PartialTensorShape first_shape;
        TF_RETURN_IF_ERROR(unwrap(tensors_[0].get())->Shape(&first_shape));
        return errors::Unimplemented(absl::StrCat(
            "Computing the shape of a ParallelTensor when the components do "
            "not all have the same rank is not supported. One tensor had "
            "shape ",

  bool found_dtype = false;
  bool found_shape = false;
  for (const auto& attr : node_def.attr()) {
    if (attr.first == "dtype") {
      if (attr.second.type() == tensorflow::DT_INT32) {
        found_dtype = true;
      } else {
        return false;
      }
    } else if (attr.first == "shape") {
      found_shape = true;
    }
  }
  return found_dtype && found_shape;
}

    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    TFE_Op* shape_op = ShapeOp(ctx, hgpu);
    TFE_OpSetDevice(shape_op, gpu_device_name.c_str(), status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
    TFE_TensorHandle* retvals[1];
    int num_retvals = 1;
    TFE_Execute(shape_op, &retvals[0], &num_retvals, status.get());