python_sugared_value.cpp

#include <torch/csrc/jit/python/python_sugared_value.h>

#include <pybind11/pytypes.h>
#include <torch/csrc/Dtype.h>
#include <torch/csrc/Layout.h>
#include <torch/csrc/MemoryFormat.h>
#include <torch/csrc/jit/frontend/schema_matching.h>
#include <torch/csrc/jit/python/module_python.h>
#include <climits>
#include <memory>
#include <sstream>
#include <string>
#include <tuple>
#include <vector>

#include <Python.h>

namespace torch {
namespace jit {

std::string typeString(py::handle h) {
  return py::str(h.get_type().attr("__name__"));
}

c10::optional<StrongFunctionPtr> as_function(const py::object& obj) {
  if (py::isinstance<StrongFunctionPtr>(obj)) {
    return py::cast<StrongFunctionPtr>(obj);
  }
  return c10::nullopt;
}

FunctionSchema PythonValue::getSchema(
    const size_t n_args,
    const size_t n_binders,
    const SourceRange& loc) {
  auto annotations = py::module::import("torch.jit.annotations");
  const auto callable = moduleSelf_ ? py::getattr(self, "original_fn") : self;

  // Make sure the function is not a class instantiation (e.g. `Exception()`)
  annotations.attr("check_fn")(callable, loc);
  auto is_vararg = py::cast<bool>(annotations.attr("is_vararg")(callable));

  auto signature = annotations.attr("get_signature")(
      callable, rcb ? *rcb : py::none(), loc, bool(moduleSelf_));
  std::vector<Argument> args, rets;

  auto py_param_names = annotations.attr("get_param_names")(callable, n_args);
  auto param_names = py::cast<std::vector<std::string>>(py_param_names);
  auto names_it = param_names.begin();
  if (moduleSelf_) {
    if (param_names.size() == 0) {
      throw ErrorReport(loc)
          << "Non-static method does not have a self argument";
    }

    // If there is a `self` parameter on the callable, skip it on the names list
    args.emplace_back(Argument(*names_it, moduleSelf_->type(), {}, {}, false));
    ++names_it;
  }
  if (signature.is_none()) {
    // No type signature was provided on the callable, so make a default
    // signature where each argument is typed as a Tensor
    for (; names_it != param_names.end(); ++names_it) {
      args.emplace_back(Argument(
          /*name=*/*names_it,
          /*type=*/TensorType::get(),
          /*N=*/c10::nullopt,
          /*default_value=*/c10::nullopt,
          /*kwarg_only=*/false));
    }

    // Use as many outputs as are requested to make the return type
    TypePtr ret_type = TensorType::get();
    if (n_binders == 0) {
      ret_type = NoneType::get();
    } else if (n_binders > 1) {
      std::vector<TypePtr> tuple_values(n_binders, ret_type);
      ret_type = TupleType::create(std::move(tuple_values));
    }
    rets.emplace_back(Argument("0", ret_type, {}, {}, false));
  } else {
    // Use the provided type signature
    std::vector<TypePtr> arg_types;
    TypePtr ret_type;
    std::tie(arg_types, ret_type) =
        py::cast<std::pair<std::vector<TypePtr>, TypePtr>>(signature);

    // arg_types does not include self but param_names does, so adjust for that
    // if needed
    TORCH_INTERNAL_ASSERT(
        arg_types.size() == param_names.size() - (moduleSelf_ ? 1 : 0));

    auto types_it = arg_types.begin();
    for (; types_it != arg_types.end(); ++types_it, ++names_it) {
      args.emplace_back(
          /*name=*/*names_it,
          /*type=*/std::move(*types_it),
          /*N=*/c10::nullopt,
          /*default_value=*/c10::nullopt,
          /*kwarg_only=*/false);
    }
    rets.push_back(Argument("0", std::move(ret_type), {}, {}, false));
  }

  std::string name;
  if (py::hasattr(self, "__qualname__")) {
    // Use the qualified name if possible
    name = py::str(py::getattr(self, "__qualname__"));
  } else if (py::hasattr(self, "__name__")) {
    name = py::str(py::getattr(self, "__name__"));
  }
  return FunctionSchema(name, "", std::move(args), std::move(rets), is_vararg);
}

std::shared_ptr<SugaredValue> PythonValue::call(
    const SourceRange& loc,
    GraphFunction& m,
    at::ArrayRef<NamedValue> args,
    at::ArrayRef<NamedValue> kwargs,
    size_t n_binders) {
  std::vector<NamedValue> argsWithSelf;
  if (moduleSelf_) {
    argsWithSelf.emplace_back(NamedValue("self", moduleSelf_));
  }
  argsWithSelf.insert(argsWithSelf.end(), args.begin(), args.end());

  auto schema = getSchema(argsWithSelf.size(), n_binders, loc);
  auto inputs = toValues(*m.graph(), argsWithSelf);

  MatchedSchema matched_schema =
      matchSchema(schema, loc, *m.graph(), argsWithSelf, kwargs);

  // If if a function is marked as dropped,
  // we throw an exception if it is invoked.
  if (py::cast<bool>(py::module::import("torch._jit_internal")
                         .attr("should_drop")(self))) {
    auto g = m.graph();
    auto err_msg = insertConstant(
        *g,
        IValue(
            "This Python function is annotated to be ignored and cannot be run"));
    g->insert(prim::RaiseException, {err_msg}, {}, loc);
    return std::make_shared<SimpleValue>(
        g->insertNode(g->createUninitialized(matched_schema.return_types.at(0)))
            ->output());
  }

  // Release the function object so we can wrap it in a PythonOp
  py::object func = self;
  std::string cconv(inputs.size(), 'd');
  Node* new_node = m.graph()->insertNode(
      m.graph()->createPythonOp(THPObjectPtr(func.release().ptr()), cconv, {}));

  new_node->setSourceRange(loc);
  for (auto& i : matched_schema.inputs)
    new_node->addInput(i);

  Value* output =
      new_node->addOutput()->setType(matched_schema.return_types.at(0));
  return std::make_shared<SimpleValue>(output);
}

std::string PythonValue::kind() const {
  std::stringstream ss;
  ss << "python value of type '" << typeString(self) << "'";
  return ss.str();
}

std::vector<std::shared_ptr<SugaredValue>> PythonValue::asTuple(
    const SourceRange& loc,
    GraphFunction& m,
    const c10::optional<size_t>& size_hint) {
  const std::string type_str = typeString(self);
  std::stringstream ss;
  ss << kind() << " cannot be used as a tuple";
  checkForAddToConstantsError(ss);
  throw ErrorReport(loc) << ss.str();
}

std::shared_ptr<SugaredValue> PythonValue::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  const std::string type_str = typeString(self);
  std::stringstream ss;
  ss << "attribute lookup is not defined on " << kind();
  checkForAddToConstantsError(ss);
  throw ErrorReport(loc) << ss.str();
}

py::object PythonValue::getattr(
    const SourceRange& loc,
    const std::string& name) {
  try {
    return py::getattr(self, name.c_str());
  } catch (py::error_already_set& e) {
    throw ErrorReport(loc) << "object has no attribute " << name;
  }
}

void PythonValue::checkForAddToConstantsError(std::stringstream& ss) {
  auto nn = py::module::import("torch.nn");
  if (py::isinstance(self, nn.attr("ModuleList")) ||
      py::isinstance(self, nn.attr("Sequential"))) {
    ss << ". Did you forget to add it to __constants__? ";
  }
}

std::shared_ptr<SugaredValue> PythonModuleValue::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  py::object member = getattr(loc, field);
  // note: is_constant = true because we consider that global properties
  // on modules like math.pi or torch.float to be constants
  // even though it is possible, though rare, for someone to mutate them
  return toSugaredValue(member, m, loc, /*is_constant=*/true);
}

#if !defined(USE_ROCM)
std::shared_ptr<SugaredValue> CUDAPythonModuleValue::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  // List of all the cuda operators which are supported in JIT
  const std::unordered_set<std::string> cuda_ops = {
      "current_stream",
      "default_stream",
      "current_device",
      "set_device",
      "device_index",
      "device_count",
      "set_stream",
      "synchronize"};

  if (cuda_ops.find(field) != cuda_ops.end()) {
    // Both current_device and set_device API's are a part of c10::cuda
    // namespace. Hence, to resolve the conflict for jit, we append _ to both
    // these APIs.
    if (field == "current_device" || field == "set_device") {
      return std::make_shared<BuiltinFunction>(
          Symbol::cuda("_" + field), c10::nullopt);
    } else {
      return std::make_shared<BuiltinFunction>(
          Symbol::cuda(field), c10::nullopt);
    }
  }

  if (field == "Stream" || field == "Event") {
    auto class_type = getCustomClass("__torch__.torch.classes.cuda." + field);
    return std::make_shared<ClassValue>(class_type);
  }

  py::object member = getattr(loc, field);
  // note: is_constant = true because we consider that global properties
  // on modules like math.pi or torch.float to be constants
  // even though it is possible, though rare, for someone to mutate them
  return toSugaredValue(member, m, loc, /*is_constant=*/true);
}
#endif

Value* ModuleValue::asValue(const SourceRange& loc, GraphFunction& m) {
  return self_;
}

SugaredValuePtr ModuleValue::asTupleValue(
    const SourceRange& loc,
    GraphFunction& m) {
  if (concreteType_->getIterableModuleKind() == IterableModuleKind::LIST) {
    auto dict = getSugaredDict(loc, m);
    auto mods = dict->getModules();
    return mods;
  }
  throw ErrorReport(loc)
      << "Only ModuleList or Sequential modules can be used as tuple";
}

bool ModuleValue::areAllSubmodulesSubtypeOf(
    const TypePtr& ty,
    std::ostream* why_not) const {
  const auto& self_type = concreteType_->getJitType()->expect<ClassType>();
  for (size_t i = 0; i < self_type->numAttributes(); ++i) {
    const auto& attr_type = self_type->getAttribute(i);
    if (attr_type->is_module()) {
      std::stringstream ss;
      if (!attr_type->isSubtypeOfExt(ty, &ss)) {
        if (why_not) {
          *why_not << "Attribute " << self_type->getAttributeName(i)
                   << " is not of annotated type " << ty->annotation_str()
                   << ": " << ss.str();
        }

        return false;
      }
    }
  }

  return true;
}

SugaredValuePtr ModuleValue::getitem(
    const SourceRange& loc,
    GraphFunction& m,
    Value* idx,
    TypePtr type_hint) {
  if (concreteType_->getIterableModuleKind() == IterableModuleKind::LIST) {
    if (type_hint) {
      // Check that all submodules comply with the type hint.
      std::stringstream ss;
      if (!areAllSubmodulesSubtypeOf(type_hint, &ss)) {
        throw ErrorReport(loc) << ss.str();
      }

      // Emit a prim::ModuleContainerIndex operator. This is needed because
      // it's difficult to construct a list in the graph representing the
      // ModuleList and use aten::__getitem__ ops to index into it because
      // any call to ModuleList.setitem would invalidate that emitted list.
      auto graph = m.graph();
      auto* getitem_node = graph->insertNode(
          graph->create(prim::ModuleContainerIndex, {self_, idx}));
      getitem_node->output(0)->setType(type_hint);
      return std::make_shared<SimpleValue>(getitem_node->output(0));
    } else {
      return getSugaredDict(loc, m)->getModules()->getitem(
          loc, m, idx, type_hint);
    }
  } else if (
      concreteType_->getIterableModuleKind() == IterableModuleKind::DICT) {
    if (auto ivalue = toIValue(idx)) {
      auto sd = getSugaredDict(loc, m);
      auto idx_str = ivalue->toStringRef();
      auto keys_iter = sd->keys_;
      auto module_values_iter = sd->modules_;
      for (size_t i = 0; i < keys_iter->tup_.size(); ++i) {
        auto key = keys_iter->tup_.at(i);
        auto key_str = toIValue(key->asValue(loc, m))->toStringRef();
        if (key_str == idx_str) {
          return module_values_iter->tup_.at(i);
        }
      }
      throw ErrorReport(loc) << "Key Error, " << idx_str;
    } else if (type_hint) {
      // Check that all submodules comply with the type hint.
      std::stringstream ss;
      if (!areAllSubmodulesSubtypeOf(type_hint, &ss)) {
        throw ErrorReport(loc) << ss.str();
      }

      // Emit a prim::ModuleContainerIndex operator. This is needed because
      // it's difficult to construct a dict in the graph representing the
      // ModuleDict and use aten::__getitem__ ops to index into it because
      // any call to ModuleDict.setAttr would invalidate that emitted dict.
      auto graph = m.graph();
      auto* getitem_node = graph->insertNode(
          graph->create(prim::ModuleContainerIndex, {self_, idx}));
      getitem_node->output(0)->setType(type_hint);
      return std::make_shared<SimpleValue>(getitem_node->output(0));
    }
    throw ErrorReport(loc)
        << "Unable to extract string literal index. "
        << "ModuleDict indexing is only supported with string literals. "
        << "Enumeration of ModuleDict is supported, e.g. 'for k, v in self.items(): ...'";
  }
  throw ErrorReport(loc)
      << "Only ModuleList, Sequential, and ModuleDict modules are subscriptable";
}

void checkInterface(
    const SourceRange& loc,
    GraphFunction& m,
    const std::shared_ptr<ModuleValue>& self,
    const std::string& field) {
  if (self->asValue(loc, m)->type()->cast<InterfaceType>()) {
    throw ErrorReport(loc)
        << "Could not compile " << field
        << "() because module is an interface type. Please file issue.";
  }
}

void recurseThroughNestedModules(
    const SourceRange& loc,
    GraphFunction& m,
    std::vector<SugaredValuePtr>& keys,
    std::vector<SugaredValuePtr>& values,
    std::shared_ptr<ModuleValue>& self,
    const std::string& prefix,
    const std::string& field) {
  auto prefix_value =
      std::make_shared<SimpleValue>(insertConstant(*m.graph(), prefix));

  keys.push_back(prefix_value);
  values.push_back(self);

  checkInterface(loc, m, self, field);
  auto module_dict = self->getSugaredDict(loc, m);
  auto keys_iter = module_dict->keys_;
  auto module_values_iter = module_dict->modules_;
  for (size_t i = 0; i < keys_iter->tup_.size(); ++i) {
    std::shared_ptr<SugaredValue> module_sugared_value =
        module_values_iter->tup_.at(i);
    auto module_value =
        std::dynamic_pointer_cast<ModuleValue>(module_sugared_value);

    auto keys_value = keys_iter->tup_.at(i);
    auto key_string = toIValue(keys_value->asValue(loc, m))->toStringRef();
    std::string submodule_prefix = prefix;
    if (prefix != "") {
      submodule_prefix = prefix + ".";
    }
    submodule_prefix += key_string;
    recurseThroughNestedModules(
        loc, m, keys, values, module_value, submodule_prefix, field);
  };
}

std::shared_ptr<SugaredDict> ModuleValue::getSugaredNamedBufferDict(
    const SourceRange& loc,
    GraphFunction& m) {
  std::vector<std::string> paramNames;
  std::vector<SugaredValuePtr> values;

  const auto& selfType = concreteType_->getJitType()->expect<ClassType>();
  for (size_t i = 0; i < selfType->numAttributes(); ++i) {
    if (selfType->is_buffer(i)) {
      paramNames.push_back(selfType->getAttributeName(i));
    }
  }

  std::vector<SugaredValuePtr> keys;
  for (const auto& name : paramNames) {
    auto name_v =
        std::make_shared<SimpleValue>(insertConstant(*m.graph(), name));
    m.graph()->insertGetAttr(self_, name);
    values.push_back(tryGetAttr(loc, m, name));
    keys.push_back(name_v);
  }

  return std::make_shared<SugaredDict>(
      std::make_shared<ModuleValue>(self_, concreteType_),
      std::make_shared<SugaredTupleValue>(keys),
      std::make_shared<SugaredTupleValue>(values));
}

std::shared_ptr<SugaredDict> ModuleValue::getSugaredDict(
    const SourceRange& loc,
    GraphFunction& m) {
  std::vector<std::string> submoduleNames;
  const auto& selfType = concreteType_->getJitType()->expect<ClassType>();
  for (size_t i = 0; i < selfType->numAttributes(); ++i) {
    const auto& attrType = selfType->getAttribute(i);
    if (attrType->is_module()) {
      submoduleNames.push_back(selfType->getAttributeName(i));
    }
  }

  std::vector<SugaredValuePtr> keys;
  std::vector<SugaredValuePtr> values;
  for (const auto& name : submoduleNames) {
    auto name_v =
        std::make_shared<SimpleValue>(insertConstant(*m.graph(), name));
    Value* module_v = m.graph()->insertGetAttr(self_, name);
    auto mod_v = std::make_shared<ModuleValue>(
        module_v, concreteType_->findSubmoduleConcreteType(name));

    keys.push_back(name_v);
    values.push_back(mod_v);
  }

  return std::make_shared<SugaredDict>(
      std::make_shared<ModuleValue>(self_, concreteType_),
      std::make_shared<SugaredTupleValue>(keys),
      std::make_shared<SugaredTupleValue>(values));
}

std::shared_ptr<SugaredValue> SugaredDict::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  // Recursive compilation does not maintain module aliasing,
  // so we do not add uniqueness checks on
  // "children"/"named_children"/"modules"/"named_modules"
  checkInterface(loc, m, self_, field);
  if (field == "keys") {
    return std::make_shared<ModuleDictMethod>(keys_, "keys");
  } else if (field == "values" || field == "children") {
    return std::make_shared<ModuleDictMethod>(modules_, field);
  } else if (
      field == "items" || field == "named_children" ||
      field == "named_buffers") {
    auto iterator = std::make_shared<IterableTree>();
    iterator->addChild(loc, m, keys_);
    iterator->addChild(loc, m, modules_);
    return std::make_shared<ModuleDictMethod>(iterator, field);
  } else if (field == "named_modules" || field == "modules") {
    std::vector<SugaredValuePtr> keys;
    std::vector<SugaredValuePtr> values;
    recurseThroughNestedModules(loc, m, keys, values, self_, "", field);
    if (field == "modules") {
      return std::make_shared<ModuleDictMethod>(
          std::make_shared<SugaredTupleValue>(values), field);
    } else {
      auto iterator = std::make_shared<IterableTree>();
      iterator->addChild(loc, m, std::make_shared<SugaredTupleValue>(keys));
      iterator->addChild(loc, m, std::make_shared<SugaredTupleValue>(values));
      return std::make_shared<ModuleDictMethod>(iterator, field);
    }
  }
  TORCH_INTERNAL_ASSERT(false);
}

std::shared_ptr<SugaredEnumClass> createSugaredEnumClassFromObj(
    const py::object& obj,
    GraphFunction& m,
    const SourceRange& loc) {
  auto annotation_type = py::module::import("torch.jit.annotations")
                             .attr("try_ann_to_type")(obj, loc);
  TORCH_INTERNAL_ASSERT(!annotation_type.is_none());
  auto type = py::cast<TypePtr>(annotation_type);
  auto enum_type = type->expect<EnumType>();
  return std::make_shared<SugaredEnumClass>(enum_type);
}

// helper function for instantiating a SugaredValue from an IValue
std::shared_ptr<SugaredValue> toSugaredValue(
    const IValue& v,
    GraphFunction& m,
    const SourceRange& loc) {
  if (v.isTuple()) {
    auto tp = v.toTuple();
    std::vector<Value*> values;
    values.reserve(tp->elements().size());
    for (const auto& e : tp->elements()) {
      values.push_back(toSugaredValue(e, m, loc)->asValue(loc, m));
    }
    return toSimple(
        m.graph()->insertNode(m.graph()->createTuple(values))->output());
  } else {
    return toSimple(m.graph()->insertConstant(v, loc));
  }
}

// This method controls how we desugar attribute lookups on ScriptModules
std::shared_ptr<SugaredValue> ModuleValue::tryGetAttr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  // 1. Look inside Module object for the field.
  const auto& selfType_ = concreteType_->getJitType();
  if (selfType_->cast<InterfaceType>()) {
    return std::make_shared<SimpleValue>(self_)->attr(loc, m, field);
  }

  const auto& selfType = selfType_->expect<ClassType>();

  if (selfType->hasAttribute(field) &&
      selfType->getAttribute(field)->is_module()) {
    // ...if it's a submodule, return it as a new ModuleValue.
    if (const auto submoduleConcreteType =
            concreteType_->findSubmoduleConcreteType(field)) {
      return std::make_shared<ModuleValue>(
          m.graph()->insertGetAttr(self_, field), submoduleConcreteType);
    }

    return std::make_shared<ModuleValue>(
        m.graph()->insertGetAttr(self_, field),
        ConcreteModuleType::fromJitType(selfType->getAttribute(field)));
  } else if (selfType->hasAttribute(field) || selfType->findMethod(field)) {
    // ...otherwise, methods, parameters, attributes, and buffers are all
    // first class so they get returned as SimpleValues
    return std::make_shared<SimpleValue>(self_)->attr(loc, m, field);
  } else if (selfType->hasConstant(field)) {
    auto v = selfType->getConstant(field);
    return toSugaredValue(v, m, loc);
  }

  // 2. Special case: for module dicts we manually desugar items(), keys(),
  // values() calls into the appropriate method.
  if (concreteType_->getIterableModuleKind() == IterableModuleKind::DICT) {
    if (field == "items" || field == "keys" || field == "values") {
      return getSugaredDict(loc, m)->attr(loc, m, field);
    }
  }

  if (field == "named_modules" || field == "modules" || field == "children" ||
      field == "named_children") {
    return getSugaredDict(loc, m)->attr(loc, m, field);
  }

  if (field == "named_buffers") {
    return getSugaredNamedBufferDict(loc, m)->attr(loc, m, field);
  }

  // 3. Check if this is the name of an overloaded method.

  // This can also be a call to a non-script module, or a plain
  // python method. If so return this as a python value.
  if (const auto overloads = concreteType_->findOverloads(field)) {
    return std::make_shared<MethodValue>(self_, *overloads);
  }

  // 4. Check if it's a function attribute.
  if (const auto fnAttr = concreteType_->findFunctionAttribute(field)) {
    return std::make_shared<FunctionValue>(*fnAttr);
  } else if (const auto builtin = concreteType_->findBuiltinFunction(field)) {
    return std::make_shared<BuiltinFunction>(*builtin, /*self=*/c10::nullopt);
  }

  // 5. Check if it's an attribute of the original Python class that this
  // ScriptModule was derived from. The only class attributes we handle are
  // methods.
  const auto maybePyClass = concreteType_->getPyClass();
  if (!maybePyClass) {
    // ConcreteType doesn't always have an originating Python class, e.g. if it
    // was derived from a serialized ScriptModule. In this case, we've exhausted
    // our options for attr lookup.
    return nullptr;
  }
  py::object unboundMethod = py::getattr(
      *maybePyClass, field.c_str(), pybind11::cast<pybind11::none>(Py_None));

  if (py::isinstance<py::function>(unboundMethod)) {
    bool isStaticFn =
        py::cast<bool>(py::module::import("torch._jit_internal")
                           .attr("is_static_fn")(*maybePyClass, field.c_str()));
    if (isStaticFn) {
      // Functions within the module annotated with @staticmethod do not need
      // binding.
      py::object staticFn =
          py::module::import("torch._jit_internal")
              .attr("get_static_fn")(*maybePyClass, field.c_str());
      return toSugaredValue(staticFn, m, loc);
    }
    // For Python methods that we're trying to call directly, we need to bind
    // the method to a self. (see the documentation for lazy_bind in Python for
    // more info).
    bool isIgnoredFn =
        py::cast<bool>(py::module::import("torch._jit_internal")
                           .attr("is_ignored_fn")(unboundMethod));
    if (isIgnoredFn) {
      // Create a generated ScriptModule type with module_ set as cpp_module
      auto boundMethod = py::module::import("torch.jit._recursive")
                             .attr("lazy_bind")(concreteType_, unboundMethod);
      TORCH_CHECK(py::isinstance<py::function>(boundMethod));
      auto rcb =
          py::module::import("torch._jit_internal")
              .attr("createResolutionCallbackFromClosure")(unboundMethod);
      return std::make_shared<PythonValue>(boundMethod, rcb, self_);
    }

    // If we reach here, it's because this is a "normal" method that just hasn't
    // been compiled yet (directly exported methods would have been returned by
    // step 1). Just compile it.
    auto stub =
        py::module::import("torch.jit._recursive")
            .attr("compile_unbound_method")(concreteType_, unboundMethod);
    TORCH_INTERNAL_ASSERT(!stub.is_none());
    // Look up the attribute again, it will be available as a compiled method.
    return attr(loc, m, field);
  }

  return nullptr;
}

bool ModuleValue::hasAttr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  return tryGetAttr(loc, m, field) != nullptr;
}

std::shared_ptr<SugaredValue> ModuleValue::call(
    const SourceRange& loc,
    GraphFunction& caller,
    at::ArrayRef<NamedValue> args,
    at::ArrayRef<NamedValue> kwargs,
    size_t n_binders) {
  c10::ClassTypePtr class_type = concreteType_->getJitType()->cast<ClassType>();
  bool have_pre_hooks =
      class_type && class_type->getForwardPreHooks().size() != 0;
  bool have_hooks = class_type && class_type->getForwardHooks().size() != 0;

  std::vector<Value*> arg_values;
  std::vector<NamedValue> pre_hook_result;
  Value* forward_input = nullptr;
  std::shared_ptr<Graph> calling_graph = caller.graph();

  if (have_pre_hooks || have_hooks) {
    // convert forward args into tuple for forward hooks
    // (the input of eager hooks are always tuples)
    for (const auto& sv : args) {
      arg_values.push_back(sv.value(*calling_graph));
    }
    forward_input =
        calling_graph->insertNode(calling_graph->createTuple(arg_values))
            ->output();
  }

  // call pre_hooks
  if (have_pre_hooks) {
    for (const auto& hook : class_type->getForwardPreHooks()) {
      TORCH_INTERNAL_ASSERT(forward_input != nullptr);
      Value* pre_hook_output =
          FunctionValue(hook)
              .call(
                  loc,
                  caller,
                  {NamedValue(self_), NamedValue(forward_input)},
                  kwargs,
                  n_binders)
              ->asValue(loc, caller);
      if (pre_hook_output->type() != NoneType::get()) {
        if (pre_hook_output->type()->kind() != TypeKind::TupleType) {
          pre_hook_output =
              calling_graph
                  ->insertNode(calling_graph->createTuple({pre_hook_output}))
                  ->output();
        }
        forward_input = pre_hook_output;
      }
    }
    // de-tuple pre_hook output for forward
    at::ArrayRef<Value*> output_nodes =
        calling_graph
            ->insertNode(calling_graph->createTupleUnpack(forward_input))
            ->outputs();
    for (auto& output_node : output_nodes) {
      pre_hook_result.emplace_back(NamedValue(output_node));
    }
    if (args.size() != 0) { // only replace input if it existed
      args = pre_hook_result;
    }
  }

  // call forward
  std::shared_ptr<SugaredValue> forwardSV =
      attr(loc, caller, "forward")->call(loc, caller, args, kwargs, n_binders);
  Value* forward_output = forwardSV->asValue(loc, caller);

  // call hooks
  if (have_hooks) {
    for (const auto& hook : class_type->getForwardHooks()) {
      Value* forward_hook_output = FunctionValue(hook)
                                       .call(
                                           loc,
                                           caller,
                                           {NamedValue(self_),
                                            NamedValue(forward_input),
                                            NamedValue(forward_output)},
                                           kwargs,
                                           n_binders)
                                       ->asValue(loc, caller);
      if (forward_hook_output->type() != NoneType::get()) {
        forward_output = forward_hook_output;
      }
    }
  }

  return std::make_shared<SimpleValue>(forward_output);
}

// This method controls how we desugar attribute lookups on ScriptModules.
std::shared_ptr<SugaredValue> ModuleValue::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  if (auto attr = tryGetAttr(loc, m, field)) {
    return attr;
  }

  // Check if it's a property.
  auto prop =
      concreteType_->getJitType()->expectRef<ClassType>().getProperty(field);
  if (prop) {
    return MethodValue(self_, prop->getter->name())
        .call(loc, m, {}, {}, /*n_binders=*/1);
  }

  // We don't define this attr. Bailout with a hint to the user.
  std::string hint;
  if (auto failureReason = concreteType_->findFailedAttribute(field)) {
    hint = *failureReason;
  } else if (concreteType_->isIgnoredAttribute(field)) {
    hint = "attribute was ignored during compilation";
  }

  throw ErrorReport(loc)
      << "Module '"
      << concreteType_->getJitType()->expectRef<ClassType>().name()->name()
      << "'"
      << " has no attribute '" << field << "' " << hint;
}

SugaredValuePtr ModuleValue::iter(const SourceRange& loc, GraphFunction& m) {
  const auto iterableModuleKind = concreteType_->getIterableModuleKind();
  if (iterableModuleKind == IterableModuleKind::NONE) {
    throw ErrorReport(loc)
        << "Only constant Sequential, ModueList, or ModuleDict can be used as an iterable";
  }

  auto module_dict = getSugaredDict(loc, m);
  if (iterableModuleKind == IterableModuleKind::DICT) {
    return module_dict->keys_;
  } else if (iterableModuleKind == IterableModuleKind::LIST) {
    return module_dict->modules_;
  } else {
    TORCH_INTERNAL_ASSERT(false);
  }
}

std::shared_ptr<SugaredValue> PythonClassValue::attr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  // Resolve values from the Python object first (e.g. for static methods on
  // this type, resolve them as functions)
  if (auto* fn = type_->findStaticMethod(field)) {
    return std::make_shared<FunctionValue>(fn);
  }
  auto py_attr = py::getattr(py_type_, field.c_str(), py::none());
  if (!py_attr.is_none()) {
    return toSugaredValue(py_attr, m, loc);
  }

  return ClassValue::attr(loc, m, field);
}

bool PythonClassValue::hasAttr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field) {
  try {
    py::getattr(py_type_, field.c_str());
    return true;
  } catch (py::error_already_set& e) {
    return false;
  }
}

void ModuleValue::setAttr(
    const SourceRange& loc,
    GraphFunction& m,
    const std::string& field,
    Value* newValue) {
  // Forward to SimpleValue::setAttr
  SimpleValue simple(self_);
  simple.setAttr(loc, m, field, newValue);
}

std::shared_ptr<SugaredValue> BooleanDispatchValue::call(
    const SourceRange& loc,
    GraphFunction& caller,
    at::ArrayRef<NamedValue> args,
    at::ArrayRef<NamedValue> kwargs,
    size_t n_binders) {
  c10::optional<bool> result;
  Graph& graph = *(caller.graph());

  auto index = py::cast<size_t>(dispatched_fn_["index"]);
  auto arg_name = py::str(dispatched_fn_["arg_name"]);

  ErrorReport error(loc);
  if (index < args.size()) {
    // Dispatch flag is in arg list
    result = constant_as<bool>(args.at(index).value(graph));
    error << "Argument for boolean dispatch at position " << index
          << " was not constant";
  } else if (auto i = findInputWithName(arg_name, kwargs)) {
    // Dispatch flag is in kwargs
    result = constant_as<bool>(kwargs[*i].value(graph));
    error << "Keyword argument '" << arg_name
          << "' for boolean dispatch at position was not constant";
  } else {
    // Didn't find dispatch flag, so use default value
    result = py::cast<bool>(dispatched_fn_["default"]);
    TORCH_INTERNAL_ASSERT(result);
  }

  if (!result.has_value()) {
    throw error;
  }

  std::shared_ptr<SugaredValue> value;
  if (*result) {
    value = toSugaredValue(dispatched_fn_["if_true"], caller, loc);
  } else {
    value = toSugaredValue(dispatched_fn_["if_false"], caller, loc);
  }
  return value->call(loc, caller, args, kwargs, n_binders);
}

std::shared_ptr<SugaredValue> PythonExceptionValue::call(
    const SourceRange& loc,
    GraphFunction& caller,
    at::ArrayRef<NamedValue> args,
    at::ArrayRef<NamedValue> kwargs,
    size_t /*n_binders*/) {
  Value* error_message = nullptr;
  if (args.size() == 0) {
    error_message = insertConstant(*caller.graph(), "", loc);
  } else if (args.size() == 1) {
    error_message = args.at(0).value(*caller.graph());
  } else {
    std::vector<Value*> message_values;
    message_values.reserve(args.size() + kwargs.size());

    for (const auto& inp : args) {
      message_values.push_back(inp.value(*caller.graph()));
    }
    for (const auto& kwarg_inp : kwargs) {
      message_values.push_back(kwarg_inp.value(*caller.graph()));
    }
    error_message =
        caller.graph()
            ->insertNode(caller.graph()->createTuple(message_values))
            ->output();
  }

  return std::make_shared<ExceptionMessageValue>(error_message);
}

bool isNamedTupleClass(const py::object& obj) {
  auto tuple_type = reinterpret_cast<PyObject*>(&PyTuple_Type);
  int is_tuple_class = PyObject_IsSubclass(obj.ptr(), tuple_type);
  if (is_tuple_class == -1) {
    PyErr_Clear();
    return false;
  }
  return is_tuple_class == 1 && py::hasattr(obj, "_fields");
}

TypePtr registerNamedTuple(const py::object& obj, const SourceRange& loc) {
  TORCH_INTERNAL_ASSERT(isNamedTupleClass(obj));
  auto qualifiedName = c10::QualifiedName(py::cast<std::string>(
      py::module::import("torch._jit_internal").attr("_qualified_name")(obj)));

  py::object props = py::module::import("torch._jit_internal")
                         .attr("_get_named_tuple_properties")(obj);

  std::string unqualName;
  std::vector<std::string> field_names;
  std::vector<TypePtr> field_types;
  std::vector<py::object> objects;

  std::tie(unqualName, field_names, field_types, objects) = py::cast<std::tuple<
      std::string,
      std::vector<std::string>,
      std::vector<TypePtr>,
      std::vector<py::object>>>(props);

  std::vector<IValue> field_defaults;
  auto min_default_idx = field_names.size() - objects.size();
  for (size_t i = min_default_idx, j = 0; i < field_names.size(); ++i, ++j) {
    py::object o = objects[j];
    auto type = tryToInferType(objects[j]);
    IValue ival = toIValue(objects[j], type.type());
    TORCH_CHECK(
        ival.tagKind() != "Tensor",
        "Tensors are"
        " not supported as default NamedTuple fields. Their "
        "mutability could lead to potential memory aliasing "
        "problems");
    field_defaults.emplace_back(ival);
  }

  auto tt = TupleType::createNamed(
      qualifiedName, field_names, field_types, field_defaults);
  if (auto type = get_python_cu()->get_type(qualifiedName)) {
    TORCH_CHECK(
        type->isSubtypeOf(tt), "Can't redefine NamedTuple: ", tt->repr_str());
    return type;
  }
  get_python_cu()->register_type(tt);
  return tt;
}

bool isEnumClass(py::object obj) {
  auto enum_type_obj =
      py::cast<py::object>(py::module::import("enum").attr("Enum"));
  int ret = PyObject_IsSubclass(obj.ptr(), enum_type_obj.ptr());
  if (ret == -1) {
    PyErr_Clear();
    return false;
  }
  return ret == 1;
}

std::shared_ptr<SugaredValue> createSimpleEnumValue(
    const py::object& obj,
    GraphFunction& m,
    const SourceRange& loc) {
  auto enum_class = obj.attr("__class__");
  auto enum_type =
      py::cast<TypePtr>(py::module::import("torch.jit.annotations")
                            .attr("try_ann_to_type")(enum_class, loc));
  auto enum_ivalue = toIValue(obj, enum_type);
  return toSimple(m.graph()->insertConstant(enum_ivalue, loc));
}

std::shared_ptr<SugaredValue> PythonSliceClass::call(
    const SourceRange& loc,
    GraphFunction& caller,
    at::ArrayRef<NamedValue> args,
    at::ArrayRef<NamedValue> kwargs,
    size_t /*n_binders*/) {
  if (!kwargs.empty()) {
    throw ErrorReport(loc) << "Slice does not accept any keyword arguments";
  }

  static constexpr int64_t default_start = 0;
  static constexpr int64_t default_stop = std::numeric_limits<int64_t>::max();
  static constexpr int64_t default_step = 1;
  Graph& graph = *(caller.graph());

  auto ValOr = [&](Value* given, int64_t default_val) {
    if (!given || given->type()->isSubtypeOf(*NoneType::get())) {
      return graph.insertConstant(default_val, loc);
    }
    return given;
  };

  Value* start = nullptr;
  Value* stop = nullptr;
  Value* step = nullptr;
  size_t n = args.size();
  // Slice's constructor signature is Slice(start=None, stop, step=None)
  if (n == 1) {
    // Case where only `stop` is specified.
    start = ValOr(nullptr, default_start);
    stop = ValOr(args[0].value(graph), default_stop);
    step = ValOr(nullptr, default_step);
  } else if (n == 2) {
    // Case where `start` and `stop` are specified.
    start = ValOr(args[0].value(graph), default_start);
    stop = ValOr(args[1].value(graph), default_stop);
    step = ValOr(nullptr, default_step);
  } else if (n == 3) {
    // Case where `start`, `stop` and `step` are all specified.
    start = ValOr(args[0].value(graph), default_start);
    stop = ValOr(args[1].value(graph), default_stop);
    step = ValOr(args[2].value(graph), default_step);
  } else {
    throw ErrorReport(loc) << "slice accepts exactly 1, 2 or 3 arguments, got: "
                           << n;
  }

  return std::make_shared<SliceValue>(start, stop, step);
}

std::shared_ptr<SugaredValue> toSugaredValue(
    py::object obj,
    GraphFunction& m,
    const SourceRange& loc,
    bool is_constant) {
  // directly create SimpleValues when possible, because they are first-class
  // and can be re-assigned. Otherwise, this would be invalid:
  // f = python_constant
  // while ...
  //   f = f + 1
  auto& g = *m.graph();
  if (is_constant) {
    if (py::isinstance<py::bool_>(obj)) {
      return toSimple(g.insertConstant(py::cast<bool>(obj), loc));
    } else if (py::isinstance<py::int_>(obj)) {
      return toSimple(g.insertConstant(py::cast<int64_t>(obj), loc));
    } else if (py::isinstance<py::float_>(obj)) {
      return toSimple(g.insertConstant(py::cast<double>(obj), loc));
    } else if (PyComplex_CheckExact(obj.ptr())) {
      auto c_obj = py::cast<std::complex<double>>(obj.ptr());
      return toSimple(
          g.insertConstant(static_cast<c10::complex<double>>(c_obj), loc));
    } else if (py::isinstance<py::str>(obj)) {
      return toSimple(g.insertConstant(py::cast<std::string>(obj), loc));
    } else if (obj.is(py::none())) {
      return toSimple(g.insertConstant(IValue(), loc));
    } else if (THPDevice_Check(obj.ptr())) {
      auto device = reinterpret_cast<THPDevice*>(obj.ptr());
      return toSimple(g.insertConstant(device->device));
    } else if (THPLayout_Check(obj.ptr())) {
      auto layout = reinterpret_cast<THPLayout*>(obj.ptr());
      const auto v = static_cast<int64_t>(layout->layout);
      return toSimple(g.insertConstant(v, loc));
    } else if (THPMemoryFormat_Check(obj.ptr())) {
      auto memory_format = reinterpret_cast<THPMemoryFormat*>(obj.ptr());
      const auto v = static_cast<int64_t>(memory_format->memory_format);
      return toSimple(g.insertConstant(v, loc));
    } else if (THPDtype_Check(obj.ptr())) {
      auto dtype = reinterpret_cast<THPDtype*>(obj.ptr());
      const auto v = static_cast<int64_t>(dtype->scalar_type);
      return toSimple(g.insertConstant(v, loc));
    } else if (THPQScheme_Check(obj.ptr())) {
      auto qscheme = reinterpret_cast<THPQScheme*>(obj.ptr());
      const auto v = static_cast<uint8_t>(qscheme->qscheme);
      return toSimple(g.insertConstant(v, loc));
    } else if (py::isinstance<py::tuple>(obj)) {
      py::tuple tup = obj;
      std::vector<Value*> values;
      values.reserve(tup.size());
      for (py::handle t : tup) {
        py::object obj = py::reinterpret_borrow<py::object>(t);
        values.push_back(toSugaredValue(obj, m, loc, true)->asValue(loc, m));
      }
      return toSimple(
          m.graph()->insertNode(m.graph()->createTuple(values))->output());
    }
  }

  bool isRpcAvailable = py::cast<bool>(
      py::module::import("torch.distributed.rpc").attr("is_available")());

  if (auto callee = as_function(obj)) {
    return std::make_shared<FunctionValue>(callee->function_);
  } else if (py::isinstance<py::module>(obj)) {
#ifndef USE_ROCM
    std::string obj_name = py::cast<py::str>(py::getattr(obj, "__name__"));
    if (obj_name.compare("torch.cuda") == 0) {
      return std::make_shared<CUDAPythonModuleValue>(obj);
    }
#endif
    return std::make_shared<PythonModuleValue>(obj);
  } else if (
      obj.ptr() == py::module::import("torch.jit").attr("_fork").ptr() ||
      obj.ptr() == py::module::import("torch.jit").attr("fork").ptr()) {
    return SpecialFormValue::create(prim::fork);
  } else if (
      obj.ptr() == py::module::import("torch.jit").attr("annotate").ptr()) {
    return SpecialFormValue::create(prim::annotate);
  } else if (
      obj.ptr() == py::module::import("torch.jit").attr("isinstance").ptr()) {
    return SpecialFormValue::create(prim::isinstance);
#ifdef USE_RPC
    // RPC module is only avaialble when build flag "USE_DISTRIBUTED" is on.
  } else if (
      isRpcAvailable &&
      obj.ptr() ==
          py::module::import("torch.distributed.rpc").attr("rpc_async").ptr()) {
    return SpecialFormValue::create(prim::rpc_async);
  } else if (
      isRpcAvailable &&
      obj.ptr() ==
          py::module::import("torch.distributed.rpc").attr("rpc_sync").ptr()) {
    return SpecialFormValue::create(prim::rpc_sync);
  } else if (
      isRpcAvailable &&
      // RPC module is only avaialble  when build flag "USE_DISTRIBUTED" is on.
      obj.ptr() ==
          py::module::import("torch.distributed.rpc").attr("remote").ptr()) {
    return SpecialFormValue::create(prim::rpc_remote);
#endif
  } else if (auto callee = as_module(obj)) {
    throw ErrorReport(loc) << "Cannot call a ScriptModule that is not"
                           << " a submodule of the caller";
  }

  py::object builtin_name =
      py::module::import("torch.jit._builtins").attr("_find_builtin")(obj);
  if (!builtin_name.is_none()) {
    return std::make_shared<BuiltinFunction>(
        Symbol::fromQualString(py::str(builtin_name)), c10::nullopt);
  }

  if (py::cast<bool>(py::module::import("torch._jit_internal")
                         .attr("_is_exception")(obj))) {
    return std::make_shared<PythonExceptionValue>(obj);
  }

  if (py::isinstance<py::function>(obj)) {
    if (typeString(obj) == "builtin_function_or_method") {
      throw ErrorReport(loc) << "Python builtin " << py::str(obj)
                             << " is currently not supported in Torchscript";
    }
  }

  py::object dispatched_fn = py::module::import("torch._jit_internal")
                                 .attr("_try_get_dispatched_fn")(obj);
  if (!dispatched_fn.is_none()) {
    return std::make_shared<BooleanDispatchValue>(std::move(dispatched_fn));
  }

  if (py::isinstance<ScriptClass>(obj)) {
    auto script_class = py::cast<ScriptClass>(obj);
    return std::make_shared<PythonClassValue>(
        script_class.class_type_.type_->expect<ClassType>(), obj);
  }

  if (isNamedTupleClass(obj)) {
    auto tuple_type = registerNamedTuple(obj, loc)->expect<TupleType>();
    return std::make_shared<NamedTupleConstructor>(tuple_type);
  }

  if (isEnumClass(obj)) {
    return createSugaredEnumClassFromObj(obj, m, loc);
  }

  auto enum_type = py::module::import("enum").attr("Enum");
  py::bool_ is_enum_value = py::isinstance(obj, enum_type);
  if (py::cast<bool>(is_enum_value)) {
    return createSimpleEnumValue(obj, m, loc);
  }

  py::bool_ is_class = py::module::import("inspect").attr("isclass")(obj);
  if (py::cast<bool>(is_class)) {
    py::str qualifiedName =
        py::module::import("torch._jit_internal").attr("_qualified_name")(obj);
    auto pyCu = get_python_cu();
    auto qualname = c10::QualifiedName(qualifiedName);

    if (auto classType = pyCu->get_class(qualname)) {
      return std::make_shared<PythonClassValue>(classType, obj);
    } else {
      // If we can't get the source code for the type, it's implemented in C and
      // probably part of the standard library, so give up and leave it as a
      // call to Python
      bool can_compile_class =
          py::cast<bool>(py::module::import("torch._jit_internal")
                             .attr("can_compile_class")(obj));
      if (can_compile_class) {
        // Register class
        auto rcb = py::module::import("torch._jit_internal")
                       .attr("createResolutionCallbackForClassMethods")(obj);
        py::module::import("torch.jit._script")
            .attr("_recursive_compile_class")(obj, loc);

        // Return class
        auto newClassType = pyCu->get_class(qualname);
        AT_ASSERT(
            newClassType,
            "Class '",
            qualifiedName,
            "' should have been compiled but was not");
        return std::make_shared<PythonClassValue>(newClassType, obj);
      }
    }
  }

  py::bool_ isFunction = py::module::import("inspect").attr("isfunction")(obj);
  if (py::cast<bool>(isFunction)) {
    auto overloads =
        py::module::import("torch.jit._script").attr("_get_overloads")(obj);
    if (!overloads.is_none()) {
      auto compiled_fns = py::cast<std::vector<StrongFunctionPtr>>(overloads);
      return std::make_shared<FunctionValue>(std::move(compiled_fns));
    }

    auto compiled_fn = py::module::import("torch.jit._recursive")
                           .attr("try_compile_fn")(obj, loc);
    if (auto callee = as_function(compiled_fn)) {
      return std::make_shared<FunctionValue>(*callee);
    }
  }

  py::bool_ isMethod = py::module::import("inspect").attr("ismethod")(obj);
  // methods here have been explicitly annotated to not be compiled,
  // so they do not have the same overload and compile checks as for functions
  if (isFunction || isMethod) {
    auto rcb = py::module::import("torch._jit_internal")
                   .attr("createResolutionCallbackFromClosure")(obj);
    return std::make_shared<PythonValue>(obj, rcb);
  }

  if (obj.is(py::module::import("builtins").attr("slice"))) {
    return std::make_shared<PythonSliceClass>();
  }

  return std::make_shared<PythonValue>(obj);
}
} // namespace jit
} // namespace torch