Add module attribute to ORTModule to support HuggingFace Trainer save_model

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

@@ -193,3 +193,18 @@ def add_module(self, name: str, module: Optional['Module']) -> None:
         """Raises a NotImplementedError exception since ORTModule does not support adding modules to it"""
 
         raise NotImplementedError("ORTModule does not support adding modules to it.")
+
+    @property
+    def module(self):
+        """The original `torch.nn.Module` that this module wraps.
+
+        This property provides access to methods and properties on the original module.
+        """
+
+        # HuggingFace Trainer `save_model` method checks to see if the input model is a HuggingFace PreTrainedModel
+        # or if the model has an attribute called `module` which references a HuggingFace PreTrainedModel to save
+        # the entire context of the model so that it can be loaded using HuggingFace `from_pretrained` method.
+        # This `module` property enables HuggingFace Trainer to retrieve the underlying PreTrainedModel inside ORTModule
+        # to save and load a complete checkpoint
+
+        return self._module_metadata.original_module

orttraining/orttraining/test/python/orttraining_test_ortmodule_api.py

@@ -6,7 +6,7 @@
 import random
 import copy
 import torch
-from transformers import AutoConfig, BertForSequenceClassification
+from transformers import AutoConfig, BertForSequenceClassification, Trainer
 from transformers.modeling_outputs import SequenceClassifierOutput
 import pytest
 from time import sleep
@@ -15,6 +15,7 @@
 from collections import OrderedDict
 from collections import namedtuple
 from inspect import signature
+import tempfile
 
 from onnxruntime.training.ortmodule import ORTModule, _utils, _io
 import _test_helpers
@@ -2777,3 +2778,33 @@ def forward(self, x):
     for param_name, param_value in state_dict1.items():
         assert param_name in state_dict2
         assert torch.equal(param_value, state_dict2[param_name])
+
+def test_hf_save_pretrained():
+    device = 'cuda'
+
+    model1 = _get_bert_for_sequence_classification_model(device)
+    model1 = ORTModule(model1)
+    state_dict = model1.state_dict()
+    list(next(iter(state_dict.items())))[1] += 100
+    model1.load_state_dict(state_dict)
+
+    trainer = Trainer(model=model1)
+
+    # Assert that ORTModule has an attribute called module. This attribute is used
+    # for trainer.save_model to reference the underlying HuggingFace PreTrainedModel
+    assert hasattr(model1, "module")
+
+    # Create a temporary directory for the checkpoint from save_pretrained
+    with tempfile.TemporaryDirectory() as temporary_dir:
+        trainer.save_model(temporary_dir)
+
+        # Create a new model and compare all state dictionary values for equality
+        # to check if from_pretrained worked.
+        config = AutoConfig.from_pretrained(temporary_dir)
+        model2 = BertForSequenceClassification.from_pretrained(
+            temporary_dir, config=config,
+        ).to(device)
+        model2 = ORTModule(model2)
+
+        for p1, p2 in zip(model1.parameters(), model2.parameters()):
+            assert p1.data.ne(p2.data).sum() == 0

tools/ci_build/github/azure-pipelines/orttraining-linux-gpu-ci-pipeline.yml

@@ -23,7 +23,7 @@ jobs:
     DockerImageTag: 'onnxruntime_orttraining_ortmodule_tests_image'
     BuildConfig: $(buildConfig)
     ArtifactName: 'drop-linux'
-    TimeoutInMinutes: 120
+    TimeoutInMinutes: 140
     # Enable unreleased onnx opsets in CI builds
     # This facilitates testing the implementation for the new opsets
     AllowReleasedOpsetOnly: '0'

tools/ci_build/github/linux/docker/scripts/training/ortmodule/stage2/requirements.txt

@@ -1,7 +1,7 @@
 pandas
 sklearn
 numpy==1.19.5
-transformers==v4.3.2
+transformers==v4.4.2
 tensorboard>=2.2.0,<2.5.0
 h5py
 wget