Remove test_initialization (#41261)

remove it

Author: Cyrilvallez

conftest.py

@@ -54,7 +54,6 @@
     "test_gradient_checkpointing_backward_compatibility",
     "test_gradient_checkpointing_enable_disable",
     "test_torch_save_load",
-    "test_initialization",
     "test_forward_signature",
     "test_model_get_set_embeddings",
     "test_model_main_input_name",

tests/models/aimv2/test_modeling_aimv2.py

@@ -39,7 +39,6 @@
 from ...test_modeling_common import (
     TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION,
     ModelTesterMixin,
-    _config_zero_init,
     _test_eager_matches_sdpa_inference,
     floats_tensor,
     ids_tensor,
@@ -430,30 +429,6 @@ def test_model_get_set_embeddings(self):
     def test_multi_gpu_data_parallel_forward(self):
         pass
 
-    # Override as the `logit_scale` parameter initialization is different for Aimv2
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `logit_scale` is initialized as per the original implementation
-                    if name == "logit_scale":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            np.log(1 / 0.07),
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def test_load_vision_text_config(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 

tests/models/align/test_modeling_align.py

@@ -469,7 +469,7 @@ def test_batching_equivalence(self, atol=3e-4, rtol=3e-4):
 
     @unittest.skip(reason="Start to fail after using torch `cu118`.")
     def test_multi_gpu_data_parallel_forward(self):
-        super().test_multi_gpu_data_parallel_forward()
+        pass
 
     @unittest.skip(reason="Hidden_states is tested in individual model tests")
     def test_hidden_states_output(self):
@@ -491,35 +491,6 @@ def test_retain_grad_hidden_states_attentions(self):
     def test_model_get_set_embeddings(self):
         pass
 
-    # override as the `temperature` parameter initialization is different for ALIGN
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `temperature` is initialized as per the original implementation
-                    if name == "temperature":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            1.0,
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    elif name == "text_projection.weight":
-                        self.assertTrue(
-                            -1.0 <= ((param.data.mean() * 1e9).round() / 1e9).item() <= 1.0,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def _create_and_check_torchscript(self, config, inputs_dict):
         if not self.test_torchscript:
             self.skipTest(reason="test_torchscript is set to False")

tests/models/altclip/test_modeling_altclip.py

@@ -467,29 +467,6 @@ def test_retain_grad_hidden_states_attentions(self):
     def test_model_get_set_embeddings(self):
         pass
 
-    # override as the `logit_scale` parameter initialization is different for AltCLIP
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `logit_scale` is initialized as per the original implementation
-                    if name == "logit_scale":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            np.log(1 / 0.07),
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def _create_and_check_torchscript(self, config, inputs_dict):
         if not self.test_torchscript:
             self.skipTest(reason="test_torchscript is set to False")

tests/models/aria/test_modeling_aria.py

@@ -199,10 +199,6 @@ def setUp(self):
         self.model_tester = AriaVisionText2TextModelTester(self)
         self.config_tester = ConfigTester(self, config_class=AriaConfig, has_text_modality=False)
 
-    @unittest.skip(reason="Unstable test")
-    def test_initialization(self):
-        pass
-
 
 SKIP = False
 torch_accelerator_module = getattr(torch, torch_device)

tests/models/aya_vision/test_modeling_aya_vision.py

@@ -198,10 +198,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
     def test_training_gradient_checkpointing_use_reentrant_false(self):
         pass
 
-    @unittest.skip(reason="Siglip uses the same initialization scheme as the Flax original implementation")
-    def test_initialization(self):
-        pass
-
     @unittest.skip(reason="Compile not yet supported because in LLava models")
     @pytest.mark.torch_compile_test
     def test_sdpa_can_compile_dynamic(self):

tests/models/bamba/test_modeling_bamba.py

@@ -41,7 +41,7 @@
 
 from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, _config_zero_init, ids_tensor
+from ...test_modeling_common import ModelTesterMixin, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
 
 
@@ -336,30 +336,6 @@ def test_decoder_model_past_with_large_inputs(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
 
-    def test_initialization(self):
-        r"""
-        Overriding the test_initialization test as the A_log and D params of the Bamba mixer are initialized differently
-        """
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    if "A_log" in name:
-                        A = torch.arange(1, config.mamba_n_heads + 1, dtype=torch.float32)
-                        torch.testing.assert_close(param.data, torch.log(A), rtol=1e-5, atol=1e-5)
-                    elif "D" in name:
-                        D = torch.ones(config.mamba_n_heads, dtype=torch.float32)
-                        torch.testing.assert_close(param.data, D, rtol=1e-5, atol=1e-5)
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def test_attention_outputs(self):
         r"""
         Overriding the test_attention_outputs test as the Bamba model outputs attention only for its attention layers

tests/models/beit/test_modeling_beit.py

@@ -34,7 +34,7 @@
 
 from ...test_backbone_common import BackboneTesterMixin
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
 
 
@@ -383,24 +383,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
     def test_training_gradient_checkpointing_use_reentrant_false(self):
         pass
 
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                # we skip lambda parameters as these require special initial values
-                # determined by config.layer_scale_init_value
-                if "lambda" in name:
-                    continue
-                if param.requires_grad:
-                    self.assertIn(
-                        ((param.data.mean() * 1e9).round() / 1e9).item(),
-                        [0.0, 1.0],
-                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                    )
-
     @slow
     def test_model_from_pretrained(self):
         model_name = "microsoft/beit-base-patch16-224"

tests/models/bit/test_modeling_bit.py

@@ -28,7 +28,6 @@
 
 if is_torch_available():
     import torch
-    from torch import nn
 
     from transformers import BitBackbone, BitForImageClassification, BitImageProcessor, BitModel
 
@@ -201,22 +200,6 @@ def test_backbone(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_backbone(*config_and_inputs)
 
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        for model_class in self.all_model_classes:
-            model = model_class(config=config)
-            for name, module in model.named_modules():
-                if isinstance(module, (nn.BatchNorm2d, nn.GroupNorm)):
-                    self.assertTrue(
-                        torch.all(module.weight == 1),
-                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                    )
-                    self.assertTrue(
-                        torch.all(module.bias == 0),
-                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                    )
-
     def test_hidden_states_output(self):
         def check_hidden_states_output(inputs_dict, config, model_class):
             model = model_class(config)

tests/models/blip/test_modeling_blip.py

@@ -459,41 +459,6 @@ def test_retain_grad_hidden_states_attentions(self):
     def test_model_get_set_embeddings(self):
         pass
 
-    # override as the `logit_scale` parameter initialization is different for Blip
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `logit_scale` is initialized as per the original implementation
-                    if name == "logit_scale":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            np.log(1 / 0.07),
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        # See PR #38607 (to avoid flakiness)
-                        data = torch.flatten(param.data)
-                        n_elements = torch.numel(data)
-                        # skip 2.5% of elements on each side to avoid issues caused by `nn.init.trunc_normal_` described in
-                        # https://github.com/huggingface/transformers/pull/27906#issuecomment-1846951332
-                        n_elements_to_skip_on_each_side = int(n_elements * 0.025)
-                        data_to_check = torch.sort(data).values
-                        if n_elements_to_skip_on_each_side > 0:
-                            data_to_check = data_to_check[
-                                n_elements_to_skip_on_each_side:-n_elements_to_skip_on_each_side
-                            ]
-                        self.assertIn(
-                            ((data_to_check.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def _create_and_check_torchscript(self, config, inputs_dict):
         if not self.test_torchscript:
             self.skipTest(reason="test_torchscript is set to False")
@@ -990,30 +955,6 @@ def test_training_gradient_checkpointing_use_reentrant(self):
     def test_training_gradient_checkpointing_use_reentrant_false(self):
         pass
 
-    # override as the `logit_scale` parameter initialization is different for Blip
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `logit_scale` is initialized as per the original implementation
-                    if name == "logit_scale":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            np.log(1 / 0.07),
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def _create_and_check_torchscript(self, config, inputs_dict):
         if not self.test_torchscript:
             self.skipTest(reason="test_torchscript is set to False")
@@ -1208,30 +1149,6 @@ def test_training_gradient_checkpointing(self):
             loss = model(**inputs).loss
             loss.backward()
 
-    # override as the `logit_scale` parameter initialization is different for Blip
-    def test_initialization(self):
-        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
-
-        configs_no_init = _config_zero_init(config)
-        for model_class in self.all_model_classes:
-            model = model_class(config=configs_no_init)
-            for name, param in model.named_parameters():
-                if param.requires_grad:
-                    # check if `logit_scale` is initialized as per the original implementation
-                    if name == "logit_scale":
-                        self.assertAlmostEqual(
-                            param.data.item(),
-                            np.log(1 / 0.07),
-                            delta=1e-3,
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-                    else:
-                        self.assertIn(
-                            ((param.data.mean() * 1e9).round() / 1e9).item(),
-                            [0.0, 1.0],
-                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
-                        )
-
     def _create_and_check_torchscript(self, config, inputs_dict):
         if not self.test_torchscript:
             self.skipTest(reason="test_torchscript is set to False")