FIX DeepSpeed recursion error (#1892)

Happened when accessing attribute before init.

examples/feature_extraction/peft_lora_embedding_semantic_search.py

@@ -194,6 +194,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
 

src/peft/mixed_model.py

@@ -191,6 +191,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "base_model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.base_model, name)
 
     def forward(self, *args: Any, **kwargs: Any):

src/peft/peft_model.py

@@ -680,6 +680,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "base_model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.base_model, name)
 
     @contextmanager

src/peft/tuners/adalora/model.py

@@ -229,6 +229,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def forward(self, *args, **kwargs):

src/peft/tuners/adaption_prompt/model.py

@@ -158,4 +158,6 @@ def __getattr__(self, name: str):
         except AttributeError:
             # This is necessary as e.g. causal models have various methods that we
             # don't want to re-implement here.
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)

src/peft/tuners/boft/model.py

@@ -207,6 +207,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def get_peft_config_as_dict(self, inference: bool = False):

src/peft/tuners/ia3/model.py

@@ -227,6 +227,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def get_peft_config_as_dict(self, inference: bool = False):

src/peft/tuners/ln_tuning/model.py

@@ -72,6 +72,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     # TODO: here need to handle the modules_to_save rather than the target_modules

src/peft/tuners/lora/model.py

@@ -355,6 +355,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def get_peft_config_as_dict(self, inference: bool = False):

src/peft/tuners/lycoris_utils.py

@@ -200,6 +200,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     @staticmethod

src/peft/tuners/mixed/model.py

@@ -183,6 +183,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def _set_adapter_layers(self, enabled=True):

src/peft/tuners/poly/model.py

@@ -114,6 +114,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def get_peft_config_as_dict(self, inference: bool = False):

src/peft/tuners/vera/model.py

@@ -329,6 +329,8 @@ def __getattr__(self, name: str):
         try:
             return super().__getattr__(name)  # defer to nn.Module's logic
         except AttributeError:
+            if name == "model":  # see #1892: prevent infinite recursion if class is not initialized
+                raise
             return getattr(self.model, name)
 
     def get_peft_config_as_dict(self, inference: bool = False):

tests/test_initialization.py

@@ -20,7 +20,21 @@
 from scipy import stats
 from torch import nn
 
-from peft import AdaLoraConfig, LoraConfig, PeftModel, PromptTuningConfig, VeraConfig, get_peft_model
+from peft import (
+    AdaLoraConfig,
+    LoraConfig,
+    PeftMixedModel,
+    PeftModel,
+    PeftModelForCausalLM,
+    PeftModelForFeatureExtraction,
+    PeftModelForQuestionAnswering,
+    PeftModelForSeq2SeqLM,
+    PeftModelForSequenceClassification,
+    PeftModelForTokenClassification,
+    PromptTuningConfig,
+    VeraConfig,
+    get_peft_model,
+)
 from peft.utils import infer_device
 
 
@@ -601,3 +615,55 @@ def test_vera_mixing_save_projection_raises(self):
         )
         with pytest.raises(ValueError, match=msg):
             model.add_adapter("other", config1)
+
+
+class TestNoInfiniteRecursionDeepspeed:
+    # see #1892 for details
+    classes = [
+        PeftModel,
+        PeftMixedModel,
+        PeftModelForSequenceClassification,
+        PeftModelForQuestionAnswering,
+        PeftModelForTokenClassification,
+        PeftModelForCausalLM,
+        PeftModelForSeq2SeqLM,
+        PeftModelForFeatureExtraction,
+    ]
+
+    @pytest.fixture
+    def wrap_init(self):
+        # emulates the wrapper from DeepSpeed
+        import functools
+
+        def decorator(f):
+            @functools.wraps(f)
+            def wrapper(self, *args, **kwargs):
+                hasattr(self, "abc")  # any hasattr will do
+                f(self, *args, **kwargs)
+
+            return wrapper
+
+        return decorator
+
+    @pytest.fixture
+    def model(self):
+        class MyModule(nn.Module):
+            def __init__(self):
+                super().__init__()
+                self.linear = nn.Linear(10, 10)
+                # to emulate LMs:
+                self.prepare_inputs_for_generation = None
+                self._prepare_encoder_decoder_kwargs_for_generation = None
+
+        return MyModule()
+
+    @pytest.mark.parametrize("cls", classes)
+    def test_no_infinite_recursion(self, cls, model, wrap_init):
+        original_init = cls.__init__
+        try:
+            cls.__init__ = wrap_init(cls.__init__)
+            # this would trigger an infinite loop before the fix in 1892
+            cls(model, LoraConfig(target_modules=["linear"]))
+        finally:
+            # ensure there are no side effects of this test
+            cls.__init__ = original_init