// 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.

      new ParallelTensor(parallel_device, std::move(components), dtype));
}

absl::Status ParallelTensor::Shape(const std::vector<int64_t>** shape) const {
  if (!shape_.has_value()) {
    TF_Status status;
    PartialTensorShape combined_shape;
    TF_RETURN_IF_ERROR(unwrap(tensors_[0].get())->Shape(&combined_shape));

    for (const TensorHandlePtr& component : tensors_) {
      PartialTensorShape component_shape;

  std::vector<PartialTensorShape> shapes;
  shapes.reserve(num_shapes);
  for (int i = 0; i < num_shapes; ++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]));
    }
  }
  desc->node_builder.Attr(attr_name, shapes);
}

    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();
  }

  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual absl::Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.

*   Tracing/timeline support for distributed runtime (no GPU profiler yet).
*   C API gives access to inferred shapes with `TF_GraphGetTensorNumDims` and
    `TF_GraphGetTensorShape`.
*   Shape functions for core ops have moved to C++ via
    `REGISTER_OP(...).SetShapeFn(...)`. Python shape inference RegisterShape
    calls use the C++ shape functions with `common_shapes.call_cpp_shape_fn`. A
    future release will remove `RegisterShape` from python.

cpovirk <******@****.***> 1721404771 -0700

  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;
  absl::Status s = handles[0]->Shape(&shape);
  ASSERT_TRUE(s.ok());
  EXPECT_EQ(0, shape->size());
}

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

namespace tensorflow {

std::string AbstractTensorHandle::DebugString() const {
  PartialTensorShape shape;
  absl::Status s = Shape(&shape);
  std::string shape_string;
  if (!s.ok()) {
    shape_string = "<error computing shape>";
  } else {
    shape_string = shape.DebugString();
  }
  return absl::StrCat("TensorHandle(shape=", shape_string,
                      ", dtype=", DataType_Name(DataType()),

namespace tensorflow {

std::string ImmediateExecutionTensorHandle::DebugString() const {
  PartialTensorShape shape;
  std::string shape_string;
  if (Shape(&shape).ok()) {
    shape_string = shape.DebugString();
  } else {
    shape_string = "<error computing shape>";
  }
  std::string value_string;
  if (!SummarizeValue(value_string).ok()) {
    value_string = "<error computing value>";
  }