Enable PythonOp for --enable_training_torch_interop build (#12539)

* enable PythonOp by default when --enable_training_torch_interop is enabled during build

* clean up

* fix

* fix comment

* fix

* fix tests

* fix fallback test

* pylint format

* refine based on comments

orttraining/orttraining/python/orttraining_pybind_state.cc

@@ -486,6 +486,13 @@ void addObjectMethodsForTraining(py::module& m, ExecutionProviderRegistrationFn
     pool.UnRegisterFunctions();
 #endif
   });
+  m.def("is_torch_interop_default_on", []() -> bool {
+#ifdef ENABLE_TRAINING_TORCH_INTEROP
+    return true;
+#else
+        return false;
+#endif
+  });
 
   py::class_<TrainingConfigurationResult> config_result(m, "TrainingConfigurationResult", "pbdoc(Configuration result for training.)pbdoc");
   config_result.def(py::init())

orttraining/orttraining/python/training/ortmodule/_custom_autograd_function.py

@@ -8,47 +8,81 @@ class Enabler(object):
     def __init__(self):
         self._state = False
 
+        # This is to indicate whether custom autograd support has been enabled or not (despite of the current state)
+        # in current process.
+        self._already_enabled = False
+
     @property
     def state(self):
         return self._state
 
+    @property
+    def already_enabled(self):
+        return self._already_enabled
+
     @state.setter
     def state(self, val):
         self._state = val
+        if self._already_enabled is False and val is True:
+            self._already_enabled = True
 
 
 custom_autograd_function_enabler = Enabler()
 
+# Legacy API to enable the custom autograd, keep its name with default value for compatibility.
+def enable_custom_autograd_support(enable=True):
 
-def enable_custom_autograd_support():
-    # Initialize static objects needed to run custom autograd.Function's.
-
-    from onnxruntime.capi._pybind_state import (
-        register_forward_runner,
-        register_backward_runner,
-        unregister_python_functions,
-    )
-    from torch.onnx import register_custom_op_symbolic
-    from ._custom_autograd_function_exporter import _export
-    from ._custom_autograd_function_runner import call_python_forward_function, call_python_backward_function
-    from onnxruntime.training.ortmodule.torch_cpp_extensions import torch_interop_utils
     import atexit
 
-    register_forward_runner(call_python_forward_function)
-    register_backward_runner(call_python_backward_function)
+    from torch.onnx import register_custom_op_symbolic, unregister_custom_op_symbolic
 
-    # Unregister all python functions automatically upon normal interpreter termination.
-    atexit.register(unregister_python_functions)
-    # Clear all gradient functions, to avoid a deadlock issue.
-    # Check the called function for more detailed comments.
-    atexit.register(torch_interop_utils.clear_all_grad_fns)
+    from onnxruntime.capi._pybind_state import (
+        register_backward_runner,
+        register_forward_runner,
+        unregister_python_functions,
+    )
+    from onnxruntime.training.ortmodule.torch_cpp_extensions import torch_interop_utils
 
-    try:
-        # This is for the latest Pytorch nightly after this commit:
-        # https://github.com/pytorch/pytorch/commit/11bc435622e6b7207bbf37ed1aafe999e1f296ec
-        register_custom_op_symbolic("prim::PythonOp", _export, 1)
-    except:
-        # This applies to Pytorch 1.9 and 1.9.1.
-        register_custom_op_symbolic("::prim_PythonOp", _export, 1)
+    from ._custom_autograd_function_exporter import _export
+    from ._custom_autograd_function_runner import call_python_backward_function, call_python_forward_function
 
-    custom_autograd_function_enabler.state = True
+    if enable is True:
+        if custom_autograd_function_enabler.already_enabled is False:
+            # Initialize static objects needed to run custom autograd.Function's.
+            register_forward_runner(call_python_forward_function)
+            register_backward_runner(call_python_backward_function)
+
+            # Unregister all python functions automatically upon normal interpreter termination.
+            atexit.register(unregister_python_functions)
+            # Clear all gradient functions, to avoid a deadlock issue.
+            # Check the called function for more detailed comments.
+            atexit.register(torch_interop_utils.clear_all_grad_fns)
+
+        try:
+            # This is for the latest Pytorch nightly after this commit:
+            # https://github.com/pytorch/pytorch/commit/11bc435622e6b7207bbf37ed1aafe999e1f296ec
+            register_custom_op_symbolic("prim::PythonOp", _export, 1)
+        except:
+            # This applies to Pytorch 1.9 and 1.9.1.
+            register_custom_op_symbolic("::prim_PythonOp", _export, 1)
+
+        custom_autograd_function_enabler.state = True
+    else:
+        if custom_autograd_function_enabler.already_enabled is True:
+            # We don't need remove the registered runner because it won't be used if we disable the feature.
+            # But we need unregister the PythonOp custom operator function.
+            try:
+                # This is for the latest Pytorch nightly after this commit:
+                # https://github.com/pytorch/pytorch/commit/11bc435622e6b7207bbf37ed1aafe999e1f296ec
+                unregister_custom_op_symbolic("prim::PythonOp", 1)
+            except:
+                # This applies to Pytorch 1.9 and 1.9.1.
+                unregister_custom_op_symbolic("::prim_PythonOp", 1)
+
+        custom_autograd_function_enabler.state = False
+
+
+from onnxruntime.capi._pybind_state import is_torch_interop_default_on
+
+# Enable the custom autograd by default when PythonOp backend support is enabled during build.
+enable_custom_autograd_support(is_torch_interop_default_on())

orttraining/orttraining/python/training/ortmodule/_graph_execution_manager.py

@@ -22,7 +22,6 @@ from onnxruntime.tools.symbolic_shape_infer import SymbolicShapeInference
 from onnxruntime.training import ortmodule
 
 from . import _are_deterministic_algorithms_enabled, _io, _logger, _onnx_models, _utils
-from ._custom_autograd_function import custom_autograd_function_enabler
 from ._custom_autograd_function_exporter import _post_process_after_export
 from ._fallback import (
     ORTModuleDeviceException,
@@ -147,6 +146,8 @@ class GraphExecutionManager(GraphExecutionInterface):
         self._run_symbolic_shape_infer = True
 
         # PyTorch custom Autograd function support
+        from ._custom_autograd_function import custom_autograd_function_enabler
+
         self._enable_custom_autograd_function = custom_autograd_function_enabler.state
 
         self._input_info = None

orttraining/orttraining/test/python/_test_helpers.py

@@ -1,17 +1,17 @@
 import copy
-import numpy as np
 import os
-import torch
 
+import numpy as np
+import torch
 from numpy.testing import assert_allclose
+
 from onnxruntime.capi.ort_trainer import ORTTrainer as Legacy_ORTTrainer
 from onnxruntime.training import orttrainer
 
 try:
     from onnxruntime.training.ortmodule import ORTModule
-    from onnxruntime.training.ortmodule._custom_autograd_function import enable_custom_autograd_support
-    from onnxruntime.training.ortmodule._graph_execution_manager_factory import GraphExecutionManagerFactory
     from onnxruntime.training.ortmodule._fallback import ORTModuleInitException
+    from onnxruntime.training.ortmodule._graph_execution_manager_factory import GraphExecutionManagerFactory
 except ImportError:
     # Some pipelines do not contain ORTModule
     pass
@@ -231,10 +231,6 @@ def assert_values_are_close(input, other, rtol=1e-05, atol=1e-06):
         assert False, err_msg
 
 
-def enable_custom_autograd_function(module):
-    enable_custom_autograd_support()
-
-
 def _run_model_on_device(device, model, input_list, label_input, is_eval_mode=False, run_forward_twice=False):
     if is_eval_mode:
         model.eval()
@@ -287,7 +283,6 @@ def run_with_ort_on_device(device, model, input_list, label_input, is_eval_mode=
     with torch.no_grad():
         model = copy.deepcopy(model)
         model.to(device)
-    enable_custom_autograd_function(model)
     model = ORTModule(model)
 
     return _run_model_on_device(device, model, input_list, label_input, is_eval_mode, run_forward_twice)

orttraining/orttraining/test/python/orttraining_test_ortmodule_autograd.py

@@ -1,6 +1,9 @@
 # Copyright (c) Microsoft Corporation. All rights reserved.
 # Licensed under the MIT License.
 
+# pylint: disable=missing-docstring
+# pylint: disable=C0103
+
 from distutils.version import LooseVersion
 
 import pytest
@@ -22,7 +25,19 @@ def torch_version_lower_than(v):
     return LooseVersion(torch.__version__) < LooseVersion(v)
 
 
-def test_GeLU():
+@pytest.fixture(scope="session", autouse=True)
+def run_before_test_session(request):
+    def insert_disable_fallback_in_env():
+        os.environ["ORTMODULE_FALLBACK_POLICY"] = "FALLBACK_DISABLE"
+
+    def remove_disable_fallback_from_env():
+        del os.environ["ORTMODULE_FALLBACK_POLICY"]
+
+    insert_disable_fallback_in_env()
+    request.addfinalizer(remove_disable_fallback_from_env)
+
+
+def test_gelu():
     @torch.jit.script
     def bias_gelu(bias, y):
         x = bias + y
@@ -741,7 +756,6 @@ def test_InnerModuleCall():
             ctx.device = device
             ctx.inner = InnerModel(dim, device).to(device)
             if use_ort:
-                enable_custom_autograd_function(ctx.inner)
                 ctx.inner = ORTModule(ctx.inner)
             z = ctx.inner(x)
             return z
@@ -1076,9 +1090,6 @@ def test_non_differentiable_autograd_function():
         print("Ref:")
         print(y_ref)
 
-        from onnxruntime.training.ortmodule._custom_autograd_function import enable_custom_autograd_support
-
-        enable_custom_autograd_support()
         m = ORTModule(m)
 
         # Inferene mode.

orttraining/orttraining/test/python/orttraining_test_ortmodule_fallback.py

@@ -4,23 +4,24 @@
 
 import copy
 import itertools
-import os
 import math
-import numpy as np
-import torch
-import pytest
+import os
 import warnings
 
-from onnxruntime.training.ortmodule import ORTModule, _fallback, ORTMODULE_TORCH_CPP_DIR
-from onnxruntime.training.ortmodule.torch_cpp_extensions import is_installed as is_torch_cpp_extensions_installed
 import _test_helpers
+import numpy as np
+import pytest
+import torch
 from _orttraining_ortmodule_models import (
-    NeuralNetSinglePositionalArgument,
-    NeuralNetCustomClassOutput,
     MyCustomClassInputNet,
     MyCustomFunctionReluModel,
+    NeuralNetCustomClassOutput,
+    NeuralNetSinglePositionalArgument,
 )
 
+from onnxruntime.training.ortmodule import ORTMODULE_TORCH_CPP_DIR, ORTModule, _fallback
+from onnxruntime.training.ortmodule.torch_cpp_extensions import is_installed as is_torch_cpp_extensions_installed
+
 # PyTorch model definitions for tests
 
 
@@ -373,6 +374,7 @@ def test_ortmodule_fallback_init__torch_version(is_training, fallback_enabled, m
     #   Otherwise, an incorrect policy (FALLBACK_UNSUPPORTED_DEVICE) is used to verify that the fallback does not happen
 
     from packaging import version
+
     from onnxruntime.training.ortmodule import MINIMUM_RUNTIME_PYTORCH_VERSION_STR
 
     runtime_pytorch_version = version.parse(torch.__version__.split("+")[0])
@@ -486,6 +488,15 @@ def test_ortmodule_fallback_init__missing_cpp_extensions(
 def test_ortmodule_fallback_onnx_model__custom_autograd(
     is_training, fallback_enabled, matching_policy, persist_fallback
 ):
+    from onnxruntime.training.ortmodule._custom_autograd_function import (
+        custom_autograd_function_enabler,
+        enable_custom_autograd_support,
+    )
+
+    # Disable the autograd support to test the fallback.
+    old_state = custom_autograd_function_enabler.state
+    enable_custom_autograd_support(False)
+
     # is_training: True for torch.nn.Module training model, eval mode otherwise
     # fallback_enabled: True PyTorch executes the forward graph instead of ORT backend
     # matching_policy: True matches FALLBACK_UNSUPPORTED_ONNX_MODEL policy to ORTModuleDeviceException exception.
@@ -536,6 +547,9 @@ def test_ortmodule_fallback_onnx_model__custom_autograd(
                 _ = ort_model(x.mm(w1)).mm(w2)
             assert "There was an error while exporting the PyTorch model to ONNX" in str(ex_info.value)
 
+    # Restore the autograd support state.
+    enable_custom_autograd_support(old_state)
+
 
 @pytest.mark.parametrize(
     "is_training,fallback_enabled,matching_policy,persist_fallback", list(itertools.product([True, False], repeat=4))