fix position ids and add tests

Author: eaidova

optimum/intel/openvino/modeling_decoder.py

@@ -428,7 +428,7 @@ def prepare_inputs(
                 position_ids = np.cumsum(attention_mask, axis=1) - 1
                 position_ids[attention_mask == 0] = 1
                 if past_key_values:
-                    position_ids = np.expand_dims(position_ids[:, -input_ids.shape[1] :], axis=-1)
+                    position_ids = position_ids[:, -input_ids.shape[1] :]
 
             inputs["position_ids"] = position_ids
 

tests/openvino/test_modeling.py

@@ -632,6 +632,11 @@ def test_multiple_inputs(self, model_arch):
         outputs = model.generate(**tokens, generation_config=generation_config)
         self.assertIsInstance(outputs, torch.Tensor)
         self.assertEqual(outputs.shape[0], 3)
+        # test that generation result is reproducible
+        outputs2 = model.generate(**tokens, generation_config=generation_config)
+        self.assertIsInstance(outputs2, torch.Tensor)
+        self.assertEqual(outputs2.shape[0], 3)
+        self.assertTrue(torch.allclose(outputs2, outputs))
         del model
         gc.collect()