Allow loading model models in a subfolder (with config in root) (#933)

Author: tomaarsen

optimum/intel/openvino/modeling_base.py

@@ -512,7 +512,7 @@ def _cached_file(
         # locates a file in a local folder and repo, downloads and cache it if necessary.
         model_path = Path(model_path)
         if model_path.is_dir():
-            model_cache_path = model_path / file_name
+            model_cache_path = model_path / subfolder / file_name
         else:
             file_name = Path(file_name)
             if file_name.suffix != ".onnx":

tests/openvino/test_modeling.py

@@ -324,6 +324,20 @@ def test_load_model_from_hub_private_with_token(self):
         self.assertIsInstance(model.config, PretrainedConfig)
         self.assertTrue(model.stateful)
 
+    @parameterized.expand(("", "openvino"))
+    def test_loading_with_config_in_root(self, subfolder):
+        # config.json file in the root directory and not in the subfolder
+        model_id = "sentence-transformers-testing/stsb-bert-tiny-openvino"
+        export = subfolder == ""
+        # hub model
+        OVModelForFeatureExtraction.from_pretrained(model_id, subfolder=subfolder, export=export)
+        # local model
+        api = HfApi()
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            local_dir = Path(tmpdirname) / "model"
+            api.snapshot_download(repo_id=model_id, local_dir=local_dir)
+            OVModelForFeatureExtraction.from_pretrained(local_dir, subfolder=subfolder, export=export)
+
     def test_infer_export_when_loading(self):
         model_id = MODEL_NAMES["phi"]
         model = AutoModelForCausalLM.from_pretrained(model_id)