diff --git a/optimum/intel/openvino/modeling_seq2seq.py b/optimum/intel/openvino/modeling_seq2seq.py
index 7f77815f4..68d19cfef 100644
--- a/optimum/intel/openvino/modeling_seq2seq.py
+++ b/optimum/intel/openvino/modeling_seq2seq.py
@@ -652,7 +652,8 @@ def forward(
 
         if "cache_position" in self.input_names:
             if cache_position is None:
-                cache_position = np.arange(self._past_len, self._past_len + input_ids.shape[1])
+                past_len = self._get_past_length(past_key_values)
+                cache_position = np.arange(past_len, past_len + input_ids.shape[1])
             inputs["cache_position"] = cache_position
 
         if "beam_idx" in self.input_names:
diff --git a/tests/openvino/test_quantization.py b/tests/openvino/test_quantization.py
index d02dea3f1..72d4a4801 100644
--- a/tests/openvino/test_quantization.py
+++ b/tests/openvino/test_quantization.py
@@ -208,8 +208,12 @@ def test_ov_model_static_quantization_with_auto_dataset(
             ov_model.save_pretrained(tmp_dir)
 
             if model_cls == OVModelForSpeechSeq2Seq:
+                models = [ov_model.encoder.model, ov_model.decoder.model]
+
+                if ov_model.decoder_with_past is not None:
+                    models.append(ov_model.decoder_with_past.model)
                 for model, expected_fq, expected_i8 in zip(
-                    (ov_model.encoder.model, ov_model.decoder.model, ov_model.decoder_with_past.model),
+                    models,
                     expected_fake_quantize,
                     expected_int8,
                 ):
@@ -629,7 +633,9 @@ def test_ovmodel_load_with_compressed_weights(self, model_cls, model_type, trust
             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]
+            if model.decoder_with_past is not None:
+                models.append(model.decoder_with_past)
         elif model.export_feature == "text-to-image":
             models = [model.unet, model.vae_encoder, model.vae_decoder]
             models.append(model.text_encoder if model_type == "stable-diffusion" else model.text_encoder_2)
@@ -772,7 +778,9 @@ def test_ovmodel_load_with_uncompressed_weights(self, model_cls, model_type, tru
             MODEL_NAMES[model_type], export=True, load_in_8bit=False, trust_remote_code=trust_remote_code
         )
         if model.export_feature.startswith("text2text-generation"):
-            models = [model.encoder, model.decoder, model.decoder_with_past]
+            models = [model.encoder, model.decoder]
+            if model.decoder_with_past is not None:
+                models.append(model.decoder_with_past)
         elif model.export_feature == "text-to-image":
             models = [model.unet, model.vae_encoder, model.vae_decoder]
             models.append(model.text_encoder if model_type == "stable-diffusion" else model.text_encoder_2)
@@ -1205,9 +1213,14 @@ def test_calibration_data_uniqueness(self, model_name, apply_caching):
         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
-        )
+        if not ov_model.stateful:
+            ov_model.decoder_with_past.request = InferRequestWrapper(
+                ov_model.decoder_with_past.request, calibration_data, apply_caching=apply_caching
+            )
+        else:
+            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)
             ov_model.generate(input_features, max_new_tokens=10, min_new_tokens=10)