Enable nested default hooks (#70932)

Summary:
When default hooks are set, they are pushed onto a stack.
When nesting context-manager, only the inner-most hooks will
be applied.

There is special care needed to update the TLS code. See also https://github.com/pytorch/pytorch/issues/70940 (i.e. do we need to be storing the enabled flag as well?)

Fixes https://github.com/pytorch/pytorch/issues/70134

Pull Request resolved: https://github.com/pytorch/pytorch/pull/70932

Reviewed By: mruberry

Differential Revision: D33530370

Pulled By: albanD

fbshipit-source-id: 3197d585d77563f36c175d3949115a0776b309f4

aten/src/ATen/SavedTensorHooks.cpp

@@ -1,13 +1,12 @@
 #include <ATen/SavedTensorHooks.h>
 #include <c10/util/Exception.h>
+#include <stack>
 
 namespace at {
 
 namespace {
   // PyObject is defined in c10/util/python_stub.h
-  // Reference counting is handled by the caller of `set_hooks`.
-  thread_local PyObject* pack_hook_(nullptr);
-  thread_local PyObject* unpack_hook_(nullptr);
+  thread_local std::stack<std::pair<PyObject*, PyObject*>> stack;
 
   // This flag is set to true the first time default hooks are registered
   // and left at true for the rest of the execution.
@@ -20,20 +19,32 @@ void SavedTensorDefaultHooks::enable() {
   is_enabled = true;
 }
 
-void SavedTensorDefaultHooks::set_hooks(PyObject* pack_hook, PyObject* unpack_hook) {
-  if (!is_enabled) {
-    TORCH_INTERNAL_ASSERT(pack_hook == nullptr && unpack_hook == nullptr);
-    return;
+void SavedTensorDefaultHooks::push_hooks(PyObject* pack_hook, PyObject* unpack_hook) {
+  // Reference counting is handled by the caller of `push_hooks`
+  TORCH_INTERNAL_ASSERT(is_enabled);
+  TORCH_INTERNAL_ASSERT(pack_hook != nullptr && unpack_hook != nullptr);
+  stack.push(std::make_pair(pack_hook, unpack_hook));
 }
-  pack_hook_ = pack_hook;
-  unpack_hook_ = unpack_hook;
+
+void SavedTensorDefaultHooks::pop_hooks() {
+  // Reference counting is handled by the caller of `pop_hooks`
+  TORCH_INTERNAL_ASSERT(is_enabled && !stack.empty());
+  stack.pop();
 }
 
 std::pair<PyObject*, PyObject*> SavedTensorDefaultHooks::get_hooks() {
-  if (!is_enabled) {
+  if (!is_enabled || stack.empty()) {
     return std::make_pair(nullptr, nullptr);
   }
-  return std::make_pair(pack_hook_, unpack_hook_);
+  return stack.top();
+}
+
+std::stack<std::pair<PyObject*, PyObject*>> SavedTensorDefaultHooks::get_stack() {
+  return stack;
+}
+
+void SavedTensorDefaultHooks::set_stack(std::stack<std::pair<PyObject*, PyObject*>> stack_) {
+  stack = stack_;
 }
 
 }

aten/src/ATen/SavedTensorHooks.h

@@ -2,15 +2,19 @@
 
 #include <c10/macros/Export.h>
 #include <c10/util/python_stub.h>
+#include <stack>
 
 #include <utility>
 
 namespace at {
 
 struct TORCH_API SavedTensorDefaultHooks {
-  static void set_hooks(PyObject* pack_hook, PyObject* unpack_hook);
+  static void push_hooks(PyObject* pack_hook, PyObject* unpack_hook);
+  static void pop_hooks();
   static std::pair<PyObject*, PyObject*> get_hooks();
   static void enable();
+  static std::stack<std::pair<PyObject*, PyObject*>> get_stack();
+  static void set_stack(std::stack<std::pair<PyObject*, PyObject*>>);
 };
 
 } // namespace at

aten/src/ATen/ThreadLocalState.cpp

@@ -15,7 +15,8 @@ ThreadLocalState::ThreadLocalState()
       functorch_tls_(functorch::getCopyOfFuncTorchTLS()),
       autograd_tls_(c10::AutogradState::get_tls_state()) {
   rf_tls_ = at::get_record_function_tls_();
-  saved_tensors_default_hooks_ = SavedTensorDefaultHooks::get_hooks();
+
+  saved_tensors_default_hooks_ = at::SavedTensorDefaultHooks::get_stack();
 
   bumped_record_all_functions_ = at::checkRecordAllFunctions();
   python_mode_state_ = at::impl::PythonModeTLS::get_state();
@@ -36,9 +37,7 @@ void ThreadLocalState::setThreadLocalState(
 
   at::set_record_function_tls_(state.rf_tls_);
 
-  SavedTensorDefaultHooks::set_hooks(
-      state.saved_tensors_default_hooks_.first,
-      state.saved_tensors_default_hooks_.second);
+  at::SavedTensorDefaultHooks::set_stack(state.saved_tensors_default_hooks_);
 
   c10::ThreadLocalDebugInfo::_forceCurrentDebugInfo(state.debug_info_);
 

aten/src/ATen/ThreadLocalState.h

@@ -1,5 +1,7 @@
 #pragma once
 
+#include <stack>
+
 #include <c10/core/InferenceMode.h>
 #include <c10/core/impl/LocalDispatchKeySet.h>
 #include <c10/util/Exception.h>
@@ -54,7 +56,7 @@ class TORCH_API ThreadLocalState {
   std::shared_ptr<TorchDispatchTypeObject> python_mode_state_;
 
   // TLS for saved tensors default hooks
-  std::pair<PyObject*, PyObject*> saved_tensors_default_hooks_;
+  std::stack<std::pair<PyObject*, PyObject*>> saved_tensors_default_hooks_;
 
   // Whether pre-sampling RecordFunction optimization was enabled
   bool bumped_record_all_functions_ = false;

test/test_autograd.py

@@ -5906,12 +5906,52 @@ for shape in [(1,), ()]:
             y.sum().backward()
             self.assertEqual(a.grad, y)
 
-    def test_setting_default_saved_variable_hooks_twice_should_fail(self):
-        with self.assertRaisesRegex(RuntimeError, "Setting default hooks but they have already been set. "):
+    def test_setting_default_saved_variable_hooks_twice_should_not_fail(self):
         with torch.autograd.graph.saved_tensors_hooks(lambda x: x, lambda x: x):
             with torch.autograd.graph.saved_tensors_hooks(lambda x: x, lambda x: x):
                 pass
 
+    def test_setting_default_saved_variable_hooks_twice_should_use_inner(self):
+        with torch.autograd.graph.saved_tensors_hooks(lambda x: 3 * x, lambda x: 3 * x):
+            b = torch.randn(5, requires_grad=True)
+            with torch.autograd.graph.saved_tensors_hooks(lambda x: 5 * x, lambda x: 5 * x):
+                a = torch.randn(5, requires_grad=True)
+                y = a * a
+            z = b * b
+        y.sum().backward()
+        z.sum().backward()
+        self.assertEqual(2 * 5 * 5 * a, a.grad)
+        self.assertEqual(2 * 3 * 3 * b, b.grad)
+
+    def test_save_on_cpu_and_checkpoint(self):
+        a = torch.randn(2, 2, requires_grad=True)
+
+        b = a.pow(2).pow(2).pow(2).pow(2)
+        b.sum().backward()
+        b_grad = a.grad.clone()
+        a.grad.zero_()
+
+        with torch.autograd.graph.save_on_cpu():
+            h = a.pow(2)
+            h = checkpoint(lambda x: x.pow(2).pow(2), h, use_reentrant=False)
+            c = h.pow(2)
+        c.sum().backward()
+        c_grad = a.grad.clone()
+        a.grad.zero_()
+
+        def f(a):
+            h = a.pow(2)
+            with torch.autograd.graph.save_on_cpu():
+                h = h.pow(2).pow(2)
+            return h.pow(2)
+
+        d = checkpoint(f, a, use_reentrant=False)
+        d.sum().backward()
+        d_grad = a.grad.clone()
+
+        self.assertEqual(b_grad, c_grad)
+        self.assertEqual(b_grad, d_grad)
+
     def test_pack_hook_with_inplace_modification_should_fail(self):
         a = torch.randn(5, requires_grad=True)
 

torch/_C/_autograd.pyi

@@ -93,8 +93,8 @@ def _record_function_with_args_exit(handle: torch.Tensor) -> None: ...
 def _supported_activities() -> Set[ProfilerActivity]: ...
 def _enable_record_function(enable: bool) -> None: ...
 def _set_empty_test_observer(is_global: bool, sampling_prob: float) -> None: ...
-def _register_saved_tensors_default_hooks(pack_hook: Callable, unpack_hook: Callable) -> None: ...
-def _reset_saved_tensors_default_hooks() -> None: ...
+def _push_saved_tensors_default_hooks(pack_hook: Callable, unpack_hook: Callable) -> None: ...
+def _pop_saved_tensors_default_hooks() -> None: ...
 
 def _enable_profiler_legacy(config: ProfilerConfig) -> None: ...
 def _disable_profiler_legacy() -> List[List[ProfilerEvent]]: ...

torch/autograd/__init__.py

@@ -307,7 +307,7 @@ from torch._C._autograd import (DeviceType, ProfilerActivity, ProfilerState, Pro
                                 _enable_record_function, _set_empty_test_observer, kineto_available,
                                 _record_function_with_args_enter, _record_function_with_args_exit,
                                 _supported_activities, _add_metadata_json, SavedTensor,
-                                _register_saved_tensors_default_hooks, _reset_saved_tensors_default_hooks)
+                                _push_saved_tensors_default_hooks, _pop_saved_tensors_default_hooks)
 
 from torch._C._autograd import (_ProfilerResult, _KinetoEvent,
                                 _prepare_profiler, _enable_profiler, _disable_profiler)

torch/autograd/graph.py

@@ -58,21 +58,21 @@ class saved_tensors_hooks():
         to undefined behavior.
 
     .. warning ::
-        Only one pair of hooks is allowed at a time. Recursively nesting this
-        context-manager is not yet supported.
+        Only one pair of hooks is allowed at a time. When recursively nesting this
+        context-manager, only the inner-most pair of hooks will be applied.
     """
     def __init__(self, pack_hook: Callable[[torch.Tensor], Any], unpack_hook: Callable[[Any], torch.Tensor]):
         self.pack_hook = pack_hook
         self.unpack_hook = unpack_hook
 
     def __enter__(self):
-        torch._C._autograd._register_saved_tensors_default_hooks(self.pack_hook, self.unpack_hook)
+        torch._C._autograd._push_saved_tensors_default_hooks(self.pack_hook, self.unpack_hook)
 
     def __exit__(self, *args: Any):
-        torch._C._autograd._reset_saved_tensors_default_hooks()
+        torch._C._autograd._pop_saved_tensors_default_hooks()
 
 
-class save_on_cpu():
+class save_on_cpu(saved_tensors_hooks):
     """Context-manager under which tensors saved by the forward pass will be
     stored on cpu, then retrieved for backward.
 
@@ -129,11 +129,4 @@ class save_on_cpu():
             device, tensor = packed
             return tensor.to(device, non_blocking=pin_memory)
 
-        self.pack_hook = pack_to_cpu
-        self.unpack_hook = unpack_from_cpu
-
-    def __enter__(self):
-        torch._C._autograd._register_saved_tensors_default_hooks(self.pack_hook, self.unpack_hook)
-
-    def __exit__(self, *args: Any):
-        torch._C._autograd._reset_saved_tensors_default_hooks()
+        super().__init__(pack_to_cpu, unpack_from_cpu)

torch/csrc/autograd/init.cpp

@@ -303,11 +303,11 @@ PyObject* THPAutograd_initExtension(PyObject* _unused, PyObject *unused) {
   m.def("_clear_callbacks", []() {
     at::clearCallbacks();
   });
-  m.def("_register_saved_tensors_default_hooks", [](py::function &pack_hook, py::function &unpack_hook) {
-    torch::autograd::PyDefaultSavedVariableHooks::set_hooks(pack_hook, unpack_hook);
+  m.def("_push_saved_tensors_default_hooks", [](py::function &pack_hook, py::function &unpack_hook) {
+    torch::autograd::PyDefaultSavedVariableHooks::push_hooks(pack_hook, unpack_hook);
   });
-  m.def("_reset_saved_tensors_default_hooks", []() {
-    torch::autograd::PyDefaultSavedVariableHooks::reset_hooks();
+  m.def("_pop_saved_tensors_default_hooks", []() {
+    torch::autograd::PyDefaultSavedVariableHooks::pop_hooks();
   });
 
   _C_m.def("_register_py_class_for_device", [](const std::string& device, py::object python_type_class) {

torch/csrc/autograd/python_saved_variable_hooks.cpp

@@ -46,25 +46,21 @@ namespace torch { namespace autograd {
     }
   }
 
-  void PyDefaultSavedVariableHooks::set_hooks(py::function &pack_hook, py::function &unpack_hook) {
-    PyObject *pack_hook_(nullptr), *unpack_hook_(nullptr);
-    std::tie(pack_hook_, unpack_hook_) = at::SavedTensorDefaultHooks::get_hooks();
-    TORCH_CHECK(!pack_hook_ && !unpack_hook_,
-        "Setting default hooks but they have already been set. "
-        "Hint: only one pair of hooks is allowed at a time.");
+  void PyDefaultSavedVariableHooks::push_hooks(py::function &pack_hook, py::function &unpack_hook) {
     at::SavedTensorDefaultHooks::enable();
-    at::SavedTensorDefaultHooks::set_hooks(pack_hook.release().ptr(), unpack_hook.release().ptr());
+    at::SavedTensorDefaultHooks::push_hooks(pack_hook.release().ptr(), unpack_hook.release().ptr());
   }
 
-  void PyDefaultSavedVariableHooks::reset_hooks() {
+  void PyDefaultSavedVariableHooks::pop_hooks() {
     PyObject *pack_hook(nullptr), *unpack_hook(nullptr);
     std::tie(pack_hook, unpack_hook) = at::SavedTensorDefaultHooks::get_hooks();
+    TORCH_INTERNAL_ASSERT(pack_hook != nullptr && unpack_hook != nullptr);
     if (Py_IsInitialized()) {
       py::gil_scoped_acquire gil;
       Py_XDECREF(pack_hook);
       Py_XDECREF(unpack_hook);
     }
-    at::SavedTensorDefaultHooks::set_hooks(nullptr, nullptr);
+    at::SavedTensorDefaultHooks::pop_hooks();
   }
 
   std::unique_ptr<SavedVariableHooks> PyDefaultSavedVariableHooks::get_hooks() {

torch/csrc/autograd/python_saved_variable_hooks.h

@@ -24,8 +24,8 @@ private:
 };
 
 struct PyDefaultSavedVariableHooks {
-  static void set_hooks(py::function &pack_hook, py::function &unpack_hook);
-  static void reset_hooks();
+  static void push_hooks(py::function &pack_hook, py::function &unpack_hook);
+  static void pop_hooks();
   static std::unique_ptr<SavedVariableHooks> get_hooks();
 };