[core] support tensor-valued _extra_state values in from_pretrained (#38155)

Support tensor-valued _extra_state values

TransformerEngine uses the pytorch get/set_extra_state API to store FP8
layer config information as bytes Tensor in the _extra_state entry in
the state dict. With recent changes to from_pretrained, this
functionality has broken and loading a model that uses this API doesn't
appear to work. This PR fixes the save/load pretrained functions for
extra state entries that use a pytorch tensor, and adds a (currently
x-failing) test for a dictionary extra state.

Signed-off-by: Peter St. John <pstjohn@nvidia.com>

src/transformers/modeling_utils.py

@@ -5577,8 +5577,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, PushToHubMixin, PeftAdapterMi
     def get_parameter_or_buffer(self, target: str):
         """
         Return the parameter or buffer given by `target` if it exists, otherwise throw an error. This combines
-        `get_parameter()` and `get_buffer()` in a single handy function. Note that it only work if `target` is a
+        `get_parameter()` and `get_buffer()` in a single handy function. If the target is an `_extra_state` attribute,
-        leaf of the model.
+        it will return the extra state provided by the module. Note that it only work if `target` is a leaf of the model.
         """
         try:
             return self.get_parameter(target)
@@ -5588,7 +5588,15 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, PushToHubMixin, PeftAdapterMi
             return self.get_buffer(target)
         except AttributeError:
             pass
-        raise AttributeError(f"`{target}` is neither a parameter nor a buffer.")
+        module, param_name = get_module_from_name(self, target)
+        if (
+            param_name == "_extra_state"
+            and getattr(module.__class__, "get_extra_state", torch.nn.Module.get_extra_state)
+            is not torch.nn.Module.get_extra_state
+        ):
+            return module.get_extra_state()
+
+        raise AttributeError(f"`{target}` is neither a parameter, buffer, nor extra state.")
 
 
 PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)

tests/utils/test_modeling_utils.py

@@ -2815,3 +2815,86 @@ class TestTensorSharing(TestCasePlus):
         shared_names, identical_names = _find_identical([{"a", "b"}], state_dict)
         self.assertEqual(shared_names, [{"a", "b"}])
         self.assertEqual(identical_names, [])
+
+
+@require_torch
+class TestSaveAndLoadModelWithExtraState(TestCasePlus):
+    """
+    This test checks that a model can be saved and loaded that uses the torch extra state API.
+    https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.get_extra_state.
+
+    Currently, only tensor-valued extra_states are supported.
+    """
+
+    def test_save_and_load_model_with_tensor_extra_state(self):
+        class MyConfig(PretrainedConfig):
+            def __init__(self, **kwargs):
+                super().__init__(**kwargs)
+
+        class MyModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.some_counter = 0
+                self.linear = torch.nn.Linear(320, 320)
+
+            def get_extra_state(self):
+                return torch.tensor(self.some_counter)
+
+            def set_extra_state(self, state):
+                self.some_counter = state.item()
+
+        class MyModel(PreTrainedModel):
+            config_class = MyConfig
+
+            def __init__(self, config: MyConfig):
+                super().__init__(config)
+                self.my_layer = MyModule()
+
+            def forward(self, hidden_states, attention_mask):
+                return self.my_layer(hidden_states, attention_mask)
+
+        config = MyConfig()
+        model = MyModel(config)
+        model.my_layer.some_counter = 42
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            model.save_pretrained(tmpdirname)
+            model = MyModel.from_pretrained(tmpdirname)
+            self.assertEqual(model.my_layer.some_counter, 42)
+
+    @mark.xfail(reason="save and from_pretrained currently only supports tensor extra_state")
+    def test_save_and_load_model_with_dict_extra_state(self):
+        class MyConfig(PretrainedConfig):
+            def __init__(self, **kwargs):
+                super().__init__(**kwargs)
+
+        class MyModule(torch.nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.some_counter = 0
+                self.linear = torch.nn.Linear(320, 320)
+
+            def get_extra_state(self):
+                return {"some_counter": self.some_counter}
+
+            def set_extra_state(self, state):
+                self.some_counter = state["some_counter"]
+
+        class MyModel(PreTrainedModel):
+            config_class = MyConfig
+
+            def __init__(self, config: MyConfig):
+                super().__init__(config)
+                self.my_layer = MyModule()
+
+            def forward(self, hidden_states, attention_mask):
+                return self.my_layer(hidden_states, attention_mask)
+
+        config = MyConfig()
+        model = MyModel(config)
+        model.my_layer.some_counter = 42
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            model.save_pretrained(tmpdirname)
+            model = MyModel.from_pretrained(tmpdirname)
+            self.assertEqual(model.my_layer.some_counter, 42)