update quantization tests

huggingface · Jul 4, 2024 · 5a717fc · 5a717fc
1 parent a7058e3
commit 5a717fc
Show file tree

Hide file tree

Showing 2 changed files with 14 additions and 4 deletions.
diff --git a/optimum/intel/openvino/modeling_base_seq2seq.py b/optimum/intel/openvino/modeling_base_seq2seq.py
@@ -239,6 +239,15 @@ def _from_pretrained(
                     force_download=force_download,
                     local_files_only=local_files_only,
                 )
+                _ = hf_hub_download(
+                    repo_id=model_id,
+                    filename=decoder_with_past_file_name.replace(".xml", ".bin"),
+                    token=token,
+                    revision=revision,
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    local_files_only=local_files_only,
+                )
 
                 decoder_with_past = cls.load_model(decoder_with_past_path, quantization_config)
 

diff --git a/tests/openvino/test_quantization.py b/tests/openvino/test_quantization.py
@@ -379,7 +379,7 @@ def test_ovmodel_load_with_compressed_weights(self, model_cls, model_type):
         self.assertEqual(model._openvino_config.dtype, "int8")
 
         if model.export_feature.startswith("text2text-generation"):
-            models = [model.encoder, model.decoder, model.decoder_with_past]
+            models = [model.encoder, model.decoder]
         elif model.export_feature.startswith("stable-diffusion"):
             models = [model.unet, model.vae_encoder, model.vae_decoder]
             models.append(model.text_encoder if model.export_feature == "stable-diffusion" else model.text_encoder_2)
@@ -488,7 +488,7 @@ def test_ovmodel_load_with_uncompressed_weights(self, model_cls, model_type):
         model = model_cls.from_pretrained(MODEL_NAMES[model_type], export=True, load_in_8bit=False)
 
         if model.export_feature.startswith("text2text-generation"):
-            models = [model.encoder, model.decoder, model.decoder_with_past]
+            models = [model.encoder, model.decoder]
         elif model.export_feature.startswith("stable-diffusion"):
             models = [model.unet, model.vae_encoder, model.vae_decoder]
             models.append(model.text_encoder if model.export_feature == "stable-diffusion" else model.text_encoder_2)
@@ -855,11 +855,12 @@ def _generate_random_audio_data(processor):
     @parameterized.expand(itertools.product(MODEL_ID, APPLY_CACHING))
     def test_calibration_data_uniqueness(self, model_id, apply_caching):
         ov_model = OVModelForSpeechSeq2Seq.from_pretrained(model_id, export=True, compile=True)
+        self.assertTrue(ov_model.decoder_with_past is None)
         processor = AutoProcessor.from_pretrained(model_id)
 
         calibration_data = []
-        ov_model.decoder_with_past.request = InferRequestWrapper(
-            ov_model.decoder_with_past.request, calibration_data, apply_caching=apply_caching
+        ov_model.decoder.request = InferRequestWrapper(
+            ov_model.decoder.request, calibration_data, apply_caching=apply_caching
         )
         for _ in range(2):
             input_features = self._generate_random_audio_data(processor)