Stop storing references to bound methods via tf.function (#24146)

* Stop storing references to bound methods in tf.functions

* Remove the gc.collect calls now that we resolved the underlying problem

* Remove the default signature from model.serving entirely, big cleanup

* Remove _prune_signature as self.input_signature can prune itself

* Restore serving docstring

* Update int support test to check the input signature

* Make sure other tests also use model.input_signature and not serving.input_signature

* Restore _prune_signature

* Remove the doctest GC now it's no longer needed

* Correct core tests to use the pruned sig

* order lines correctly in core tests

* Add eager_serving back with a deprecation warning

src/transformers/modeling_tf_utils.py

@@ -1171,12 +1171,8 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
         self.config = config
         self.name_or_path = config.name_or_path
         self.generation_config = GenerationConfig.from_model_config(config) if self.can_generate() else None
-        if not hasattr(self, "serving"):  # Don't overwrite existing serving signatures
-            self.serving = tf.function(
-                self.eager_serving, input_signature=[self._prune_signature(self.input_signature)]
-            )
         # Set the serving spec quickly to ensure that Keras doesn't use the specific dummy input shapes as the spec
-        self._set_save_spec(self.serving.input_signature[0])
+        self._set_save_spec(self._prune_signature(self.input_signature))
 
     def get_config(self):
         return self.config.to_dict()
@@ -1226,15 +1222,31 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
         head_mask = tf.cast(head_mask, tf.float32)  # switch to float if need + fp16 compatibility
         return head_mask
 
+    @tf.function
+    def serving(self, inputs):
+        """
+        Args:
+        Method used for serving the model. Does not have a specific signature, but will be specialized as concrete
+        functions when saving with `save_pretrained`.
+            inputs (`Dict[str, tf.Tensor]`):
+                The input of the saved model as a dictionary of tensors.
+        """
+        output = self.call(inputs)
+
+        return self.serving_output(output)
+
     def eager_serving(self, inputs):
         """
-        Method used for serving the model. Intended not to be compiled with a tf.function decorator so that we can use
-        it to generate multiple signatures later.
+        Method used for serving the model. This method is deprecated, and will be removed.
 
         Args:
             inputs (`Dict[str, tf.Tensor]`):
                 The input of the saved model as a dictionary of tensors.
         """
+        warnings.warn(
+            "The function `eager_serving` is deprecated and will be removed in version 4.32.0 of Transformers",
+            FutureWarning,
+        )
         output = self.call(inputs)
 
         return self.serving_output(output)
@@ -2409,17 +2421,19 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
             if getattr(self.config, "torch_dtype", None) is not None and not isinstance(self.config.torch_dtype, str):
                 self.config.torch_dtype = str(self.config.torch_dtype).split(".")[1]
             if signatures is None:
-                if any(spec.dtype == tf.int32 for spec in self.serving.input_signature[0].values()):
+                sig = self._prune_signature(self.input_signature)
+                serving_default = self.serving.get_concrete_function(sig)
+                if any(spec.dtype == tf.int32 for spec in sig.values()):
                     int64_spec = {
                         key: tf.TensorSpec(
                             shape=spec.shape, dtype=tf.int64 if spec.dtype == tf.int32 else spec.dtype, name=spec.name
                         )
-                        for key, spec in self.serving.input_signature[0].items()
+                        for key, spec in sig.items()
                     }
-                    int64_serving = tf.function(self.eager_serving, input_signature=[int64_spec])
-                    signatures = {"serving_default": self.serving, "int64_serving": int64_serving}
+                    int64_serving = self.serving.get_concrete_function(int64_spec)
+                    signatures = {"serving_default": serving_default, "int64_serving": int64_serving}
                 else:
-                    signatures = self.serving
+                    signatures = serving_default
             saved_model_dir = os.path.join(save_directory, "saved_model", str(version))
             self.save(saved_model_dir, include_optimizer=False, signatures=signatures)
             logger.info(f"Saved model created in {saved_model_dir}")

src/transformers/testing_utils.py

@@ -1882,13 +1882,6 @@ def preprocess_string(string, skip_cuda_tests):
     if not is_cuda_found:
         modified_string = "".join(codeblocks)
 
-        if ">>>" in modified_string:
-            lines = modified_string.split("\n")
-            indent = len(lines[-1]) - len(lines[-1].lstrip())
-
-            cleanup = ">>> import gc; gc.collect()  # doctest: +IGNORE_RESULT"
-            modified_string += "\n" + " " * indent + cleanup
-
     return modified_string
 
 

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_tf_{{cookiecutter.lowercase_modelname}}.py

@@ -2676,7 +2676,7 @@ class TF{{cookiecutter.camelcase_modelname}}ForConditionalGeneration(TF{{cookiec
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
         self.model = TF{{cookiecutter.camelcase_modelname}}MainLayer(config, name="model")
-        self.model._set_save_spec(inputs=self.serving.input_signature)
+        self.model._set_save_spec(self._prune_signature(self.input_signature))
         self.use_cache = config.use_cache
         # final_bias_logits is registered as a buffer in pytorch, so not trainable for the sake of consistency.
         self.bias_layer = BiasLayer(

tests/models/rag/test_modeling_tf_rag.py

 from __future__ import annotations
 
-import gc
 import json
 import os
 import shutil
@@ -551,11 +550,6 @@ class TFRagDPRBartTest(TFRagTestMixin, unittest.TestCase):
 @require_sentencepiece
 @require_tokenizers
 class TFRagModelIntegrationTests(unittest.TestCase):
-    def tearDown(self):
-        super().tearDown()
-        # clean-up as much as possible GPU memory occupied by PyTorch
-        gc.collect()
-
     @cached_property
     def token_model(self):
         return TFRagTokenForGeneration.from_pretrained_question_encoder_generator(

tests/models/sam/test_modeling_tf_sam.py

@@ -17,7 +17,6 @@
 
 from __future__ import annotations
 
-import gc
 import inspect
 import unittest
 
@@ -431,11 +430,6 @@ def prepare_dog_img():
 @require_tf
 @slow
 class TFSamModelIntegrationTest(unittest.TestCase):
-    def tearDown(self):
-        super().tearDown()
-        # clean-up as much as possible GPU memory occupied by PyTorch
-        gc.collect()
-
     def test_inference_mask_generation_no_point(self):
         model = TFSamModel.from_pretrained("facebook/sam-vit-base")
         processor = SamProcessor.from_pretrained("facebook/sam-vit-base")

tests/models/xglm/test_modeling_tf_xglm.py

@@ -15,7 +15,6 @@
 
 from __future__ import annotations
 
-import gc
 import unittest
 
 from transformers import XGLMConfig, XGLMTokenizer, is_tf_available
@@ -173,11 +172,6 @@ class TFXGLMModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase
 
 @require_tf
 class TFXGLMModelLanguageGenerationTest(unittest.TestCase):
-    def tearDown(self):
-        super().tearDown()
-        # clean-up as much as possible GPU memory occupied by PyTorch
-        gc.collect()
-
     @slow
     def test_lm_generate_xglm(self, verify_outputs=True):
         model = TFXGLMForCausalLM.from_pretrained("facebook/xglm-564M")

tests/test_modeling_tf_common.py

@@ -1687,14 +1687,10 @@ class TFModelTesterMixin:
                 if tensor.dtype.is_integer:
                     self.assertTrue(tensor.dtype == tf.int32, "Integer dummy inputs should be tf.int32!")
 
-            # Also confirm that the serving sig uses int32
-            if hasattr(model, "serving"):
-                serving_sig = model.serving.input_signature
-                for key, tensor_spec in serving_sig[0].items():
+            # Also confirm that the input_signature uses int32
+            for key, tensor_spec in model.input_signature.items():
                 if tensor_spec.dtype.is_integer:
-                        self.assertTrue(
-                            tensor_spec.dtype == tf.int32, "Serving signatures should use tf.int32 for ints!"
-                        )
+                    self.assertTrue(tensor_spec.dtype == tf.int32, "Input signatures should use tf.int32 for ints!")
 
     def test_generate_with_headmasking(self):
         attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"]

tests/utils/test_modeling_tf_core.py

@@ -217,17 +217,18 @@ class TFCoreModelTesterMixin:
         for model_class in self.all_model_classes:
             class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
             model = model_class(config)
+            class_sig = model._prune_signature(model.input_signature)
             num_out = len(model(class_inputs_dict))
 
             for key in list(class_inputs_dict.keys()):
                 # Remove keys not in the serving signature, as the SavedModel will not be compiled to deal with them
-                if key not in model.serving.input_signature[0]:
+                if key not in class_sig:
                     del class_inputs_dict[key]
                 # Check it's a tensor, in case the inputs dict has some bools in it too
                 elif isinstance(class_inputs_dict[key], tf.Tensor) and class_inputs_dict[key].dtype.is_integer:
                     class_inputs_dict[key] = tf.cast(class_inputs_dict[key], tf.int32)
 
-            if set(class_inputs_dict.keys()) != set(model.serving.input_signature[0].keys()):
+            if set(class_inputs_dict.keys()) != set(class_sig.keys()):
                 continue  # Some models have inputs that the preparation functions don't create, we skip those
 
             with tempfile.TemporaryDirectory() as tmpdirname: