fix: mobilenet_v2 tests

Author: JingyaHuang

optimum/exporters/neuron/__main__.py

@@ -138,7 +138,7 @@ def get_input_shapes(task: str, args: argparse.Namespace) -> dict[str, int]:
 def get_neuron_config_class(task: str, model_id: str) -> NeuronExportConfig:
     config = AutoConfig.from_pretrained(model_id)
 
-    model_type = config.model_type.replace("_", "-")
+    model_type = config.model_type
     if config.is_encoder_decoder:
         model_type = model_type + "-encoder"
 

optimum/exporters/neuron/model_configs.py

@@ -452,11 +452,9 @@ class LevitNeuronConfig(ViTNeuronConfig):
     pass
 
 
-@register_in_tasks_manager(
-    "mobilenet-v2", *["feature-extraction", "image-classification", "semantic-segmentation", "image-segmentation"]
-)
+@register_in_tasks_manager("mobilenet_v2", *["feature-extraction", "image-classification", "semantic-segmentation"])
 class MobileNetV2NeuronConfig(ViTNeuronConfig):
-    MODEL_TYPE = "mobilenet-v2"
+    MODEL_TYPE = "mobilenet_v2"
     pass
 
 

optimum/neuron/modeling_traced.py

@@ -492,7 +492,6 @@ def _neuron_config_init(cls, config: "PretrainedConfig") -> "NeuronDefaultConfig
         task = neuron_config.get("task", None) or TasksManager.infer_task_from_model(cls.auto_model_class)
         task = TasksManager.map_from_synonym(task)
         model_type = neuron_config.get("model_type", None) or config.model_type
-        model_type = model_type.replace("_", "-")
         neuron_config_constructor = TasksManager.get_exporter_config_constructor(
             model_type=model_type,
             exporter="neuron",

optimum/neuron/utils/argument_utils.py

@@ -302,7 +302,7 @@ def store_compilation_config(
     original_model_type = getattr(config, "export_model_type", None) or getattr(
         config, "model_type", None
     )  # prioritize sentence_transformers to transformers
-    neuron_model_type = str(model_type).replace("_", "-") if model_type is not None else model_type
+    neuron_model_type = str(model_type) if model_type is not None else model_type
     if original_model_type is None:
         update_func(
             "model_type", neuron_model_type

tests/exporters/exporters_utils.py

@@ -50,7 +50,7 @@
     "hubert": "hf-internal-testing/tiny-random-HubertModel",
     "levit": "hf-internal-testing/tiny-random-LevitModel",
     "mobilebert": "hf-internal-testing/tiny-random-MobileBertModel",
-    "mobilenet-v2": "hf-internal-testing/tiny-random-MobileNetV2Model",
+    "mobilenet_v2": "hf-internal-testing/tiny-random-MobileNetV2Model",
     # "mobilevit": "hf-internal-testing/tiny-random-mobilevit",  # blocked since neuron sdk 2.23: timeout
     "modernbert": "hf-internal-testing/tiny-random-ModernBertModel",
     "mpnet": "hf-internal-testing/tiny-random-MPNetModel",
@@ -80,7 +80,7 @@
     "clip": "sentence-transformers/clip-ViT-B-32",
 }
 
-WEIGHTS_NEFF_SEPARATION_UNSUPPORTED_ARCH = ["camembert", "roberta", "mobilenet-v2"]
+WEIGHTS_NEFF_SEPARATION_UNSUPPORTED_ARCH = ["camembert", "roberta", "mobilenet_v2"]
 
 # Diffusers
 

tests/inference/inference_utils.py

@@ -48,7 +48,7 @@
     "gpt2": "hf-internal-testing/tiny-random-gpt2",
     "levit": "hf-internal-testing/tiny-random-LevitModel",
     "mobilebert": "hf-internal-testing/tiny-random-MobileBertModel",
-    "mobilenet-v2": "hf-internal-testing/tiny-random-MobileNetV2Model",
+    "mobilenet_v2": "hf-internal-testing/tiny-random-MobileNetV2Model",
     "mobilevit": "hf-internal-testing/tiny-random-mobilevit",
     "modernbert": "hf-internal-testing/tiny-random-ModernBertModel",
     "mpnet": "hf-internal-testing/tiny-random-MPNetModel",

tests/inference/transformers/test_modeling.py

@@ -483,14 +483,6 @@ def _validate_outputs(self, model_arch, suffix, batch_size):
                 f"Inference results between pytorch model and neuron model of {model_arch} not close enough."
             )
 
-    def test_load_vanilla_transformers_which_is_not_supported(self):
-        with self.assertRaises(Exception) as context:
-            _ = NeuronModelForMaskedLM.from_pretrained(
-                "hf-internal-testing/tiny-random-t5", from_transformers=True, **self.STATIC_INPUTS_SHAPES
-            )
-
-        self.assertIn("Unrecognized configuration class", str(context.exception))
-
     @parameterized.expand(SUPPORTED_ARCHITECTURES, skip_on_empty=True)
     def test_compare_to_transformers_non_dyn_bs(self, model_arch):
         model_args = {
@@ -603,14 +595,6 @@ def _validate_outputs(self, model_arch, suffix, batch_size):
         if not result_close_end_logits:
             warnings.warn(f"End logits between pytorch model and neuron model of {model_arch} not close enough.")
 
-    def test_load_vanilla_transformers_which_is_not_supported(self):
-        with self.assertRaises(Exception) as context:
-            _ = NeuronModelForQuestionAnswering.from_pretrained(
-                "hf-internal-testing/tiny-random-t5", from_transformers=True, **self.STATIC_INPUTS_SHAPES
-            )
-
-        assert ("doesn't support" in str(context.exception)) or ("is not supported" in str(context.exception))
-
     def test_compare_to_transformers_dyn_bs(self):
         model_arch = "albert"
         # Neuron model with dynamic batching
@@ -719,14 +703,6 @@ def _validate_outputs(self, model_arch, suffix, batch_size):
                 f"Inference results between pytorch model and neuron model of {model_arch} not close enough."
             )
 
-    def test_load_vanilla_transformers_which_is_not_supported(self):
-        with self.assertRaises(Exception) as context:
-            _ = NeuronModelForSequenceClassification.from_pretrained(
-                "hf-internal-testing/tiny-random-t5", from_transformers=True, **self.STATIC_INPUTS_SHAPES
-            )
-
-        assert ("doesn't support" in str(context.exception)) or ("is not supported" in str(context.exception))
-
     @parameterized.expand(SUPPORTED_ARCHITECTURES, skip_on_empty=True)
     def test_compare_to_transformers_non_dyn_bs(self, model_arch):
         model_args = {
@@ -831,14 +807,6 @@ def _validate_outputs(self, model_arch, suffix, batch_size):
                 f"Inference results between pytorch model and neuron model of {model_arch} not close enough."
             )
 
-    def test_load_vanilla_transformers_which_is_not_supported(self):
-        with self.assertRaises(Exception) as context:
-            _ = NeuronModelForTokenClassification.from_pretrained(
-                "hf-internal-testing/tiny-random-t5", from_transformers=True, **self.STATIC_INPUTS_SHAPES
-            )
-
-        assert ("doesn't support" in str(context.exception)) or ("is not supported" in str(context.exception))
-
     @parameterized.expand(SUPPORTED_ARCHITECTURES, skip_on_empty=True)
     def test_compare_to_transformers_non_dyn_bs(self, model_arch):
         model_args = {
@@ -987,7 +955,7 @@ class NeuronModelForImageClassificationIntegrationTest(NeuronModelTestMixin):
         "cvt",
         "deit",
         "levit",
-        "mobilenet-v2",
+        "mobilenet_v2",
         "mobilevit",
         "swin",
         "vit",
@@ -1046,11 +1014,6 @@ def test_compare_to_transformers_non_dyn_bs(self, model_arch):
             "model_arch": model_arch,
             "dynamic_batch_size": False,
         }
-        # REMOVEME: convnextv2 contains a bug in the GRN layer, which is used in the convnextv2 model, but the bug has
-        # been fixed in the transformers library on newer versions. For more info see:
-        # https://github.com/huggingface/transformers/issues/38015
-        if model_arch == "convnextv2" and transformers_version.startswith("4.51"):
-            self.skipTest("convnextv2 contains a bug in this version of transformers.")
         self._setup(model_args)
         self._validate_outputs(model_arch, "_dyn_bs_false", batch_size=1)