Fixed issues. Applied comments.

Author: AlexKoff88

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -253,7 +253,7 @@ def _from_transformers(
 
         compression_option = None
         if load_in_8bit is not None:
-            compression_option = "int8" if load_in_8bit else "fp32"
+            compression_option = "fp32"
         main_export(
             model_name_or_path=model_id,
             output=save_dir_path,
@@ -270,7 +270,7 @@ def _from_transformers(
 
         config.save_pretrained(save_dir_path)
         return cls._from_pretrained(
-            model_id=save_dir_path, config=config, use_cache=use_cache, load_in_8bit=False, **kwargs
+            model_id=save_dir_path, config=config, use_cache=use_cache, load_in_8bit=load_in_8bit, **kwargs
         )
 
     def _reshape(self, model: openvino.runtime.Model, batch_size: int, sequence_length: int, is_decoder=True):

optimum/intel/openvino/modeling_decoder.py

@@ -289,7 +289,7 @@ def _from_transformers(
             model_id=save_dir_path,
             config=config,
             use_cache=use_cache,
-            load_in_8bit=False,
+            load_in_8bit=load_in_8bit,
             stateful=None,
             load_in_4bit=load_in_4bit,
             quantization_config=quantization_config,
@@ -360,7 +360,7 @@ class OVModelForCausalLM(OVBaseDecoderModel, GenerationMixin):
             checkpoint="gpt2",
         )
     )
-    def prepare_forward_inputs(
+    def prepare_inputs(
         self,
         input_ids: torch.LongTensor,
         attention_mask: Optional[torch.LongTensor] = None,

optimum/intel/openvino/weight_quantization.py

@@ -143,7 +143,7 @@ def compress_decoder_weights(model, quantization_config: Union[OVWeightQuantizat
 
             dataset = get_dataset(config.dataset, tokenizer, seqlen=32)
             dataset = prepare_dataset(dataset)
-            dataset = nncf.Dataset(dataset, lambda x: model.prepare_forward_inputs(**x))
+            dataset = nncf.Dataset(dataset, lambda x: model.prepare_inputs(**x))
 
         model.model = nncf.compress_weights(
             ov_model,

optimum/intel/utils/dummy_openvino_and_nncf_objects.py

@@ -57,3 +57,14 @@ def __init__(self, *args, **kwargs):
     @classmethod
     def from_pretrained(cls, *args, **kwargs):
         requires_backends(cls, ["openvino", "nncf"])
+
+
+class OVWeightQuantizationConfig(metaclass=DummyObject):
+    _backends = ["openvino", "nncf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["openvino", "nncf"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["openvino", "nncf"])

tests/openvino/test_quantization.py

@@ -303,13 +303,12 @@ def test_ovmodel_4bit_weight_compression(self, model_cls, model_name, expected_i
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES_STATEFUL_WITH_EXPECTED_8BIT_COMPRESSED_MATMULS)
     @unittest.skipIf(not IS_SUPPORT_STATEFUL, "Stateful models supported only in 2023.3 and above")
-    def test_ovmodel_8bit_weight_compression_stateful(self, model_cls, model_name, expected_pt_int8, expected_ov_int8):
+    def test_ovmodel_8bit_weight_compression_stateful(self, model_cls, model_id, expected_pt_int8, expected_ov_int8):
         task = model_cls.export_feature
 
         with tempfile.TemporaryDirectory() as tmp_dir:
-            model_id = MODEL_NAMES[model_name]
             transformers_model = model_cls.from_pretrained(model_id, export=True, stateful=True)
-            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            tokenizer = AutoTokenizer.from_pretrained(model_id)
             if tokenizer.pad_token is None:
                 tokenizer.pad_token = tokenizer.eos_token
 

tests/openvino/utils_tests.py

@@ -103,15 +103,15 @@
     "bert": (70,),
     "roberta": (68,),
     "albert": (84,),
-    "vit": (62,),
+    "vit": (64,),
     "blenderbot": (70,),
     "gpt2": (46,),
-    "wav2vec2": (30,),
+    "wav2vec2": (34,),
     "distilbert": (66,),
     "t5": (64, 104, 84),
-    "stable-diffusion": (148, 8, 8, 64),
-    "stable-diffusion-xl": (296, 8, 8, 66),
-    "stable-diffusion-xl-refiner": (296, 8, 8, 66),
+    "stable-diffusion": (242, 34, 42, 64),
+    "stable-diffusion-xl": (366, 34, 42, 66),
+    "stable-diffusion-xl-refiner": (366, 34, 42, 66),
 }
 
 _ARCHITECTURES_TO_EXPECTED_INT4_INT8 = {"opt125m": (64, 477)}