Small OpenVINO UX improvements (huggingface#629)

* Update ov_config, change warning in .to() to debug

- set PERFORMANCE_HINT to LATENCY if not specified in ov_config
- replace warning log in .to() about devices with debug log (to prevent
  confusing users who create a pipeline(), which always shows this warning)

* Set seq2seq ov_config in base model

optimum/intel/openvino/modeling.py

@@ -137,7 +137,7 @@ def to(self, device: str):
             self._device = device.upper()
             self.request = None
         else:
-            logger.warning(f"device must be of type {str} but got {type(device)} instead")
+            logger.debug(f"device must be of type {str} but got {type(device)} instead")
 
         return self
 

optimum/intel/openvino/modeling_base.py

@@ -64,7 +64,10 @@ def __init__(
         self.model_save_dir = model_save_dir
         self._device = device.upper()
         self.is_dynamic = dynamic_shapes
-        self.ov_config = ov_config if ov_config is not None else {"PERFORMANCE_HINT": "LATENCY"}
+        self.ov_config = ov_config if ov_config is not None else {}
+        if self.ov_config.get("PERFORMANCE_HINT") is None:
+            self.ov_config["PERFORMANCE_HINT"] = "LATENCY"
+
         self.preprocessors = kwargs.get("preprocessors", [])
         enable_compilation = kwargs.get("compile", True)
 

optimum/intel/openvino/modeling_base_seq2seq.py

@@ -67,6 +67,10 @@ def __init__(
         self._device = device.upper()
         self.is_dynamic = dynamic_shapes
         self.ov_config = ov_config if ov_config is not None else {}
+
+        if self.ov_config.get("PERFORMANCE_HINT") is None:
+            self.ov_config["PERFORMANCE_HINT"] = "LATENCY"
+
         self.preprocessors = kwargs.get("preprocessors", [])
 
         if self.is_dynamic:

optimum/intel/openvino/modeling_diffusion.py

@@ -101,6 +101,8 @@ def __init__(
         self._device = device.upper()
         self.is_dynamic = dynamic_shapes
         self.ov_config = ov_config if ov_config is not None else {}
+        if self.ov_config.get("PERFORMANCE_HINT") is None:
+            self.ov_config["PERFORMANCE_HINT"] = "LATENCY"
 
         # This attribute is needed to keep one reference on the temporary directory, since garbage collecting
         # would end-up removing the directory containing the underlying OpenVINO model
@@ -456,7 +458,7 @@ def to(self, device: str):
             self._device = device.upper()
             self.clear_requests()
         else:
-            logger.warning(f"device must be of type {str} but got {type(device)} instead")
+            logger.debug(f"device must be of type {str} but got {type(device)} instead")
 
         return self
 

optimum/intel/openvino/modeling_seq2seq.py

@@ -285,7 +285,7 @@ def to(self, device: str):
                 self.decoder_with_past._device = self._device
             self.clear_requests()
         else:
-            logger.warning(f"device must be of type {str} but got {type(device)} instead")
+            logger.debug(f"device must be of type {str} but got {type(device)} instead")
 
         return self
 

tests/openvino/test_modeling.py

@@ -116,6 +116,9 @@ def test_load_from_hub_and_save_model(self):
         tokens = tokenizer("This is a sample input", return_tensors="pt")
         loaded_model = OVModelForSequenceClassification.from_pretrained(self.OV_MODEL_ID)
         self.assertIsInstance(loaded_model.config, PretrainedConfig)
+        # Test that PERFORMANCE_HINT is set to LATENCY by default
+        self.assertEqual(loaded_model.ov_config.get("PERFORMANCE_HINT"), "LATENCY")
+        self.assertEqual(loaded_model.request.get_property("PERFORMANCE_HINT"), "LATENCY")
         loaded_model_outputs = loaded_model(**tokens)
 
         # Test specifying ov_config with throughput hint and manual cache dir
@@ -134,7 +137,10 @@ def test_load_from_hub_and_save_model(self):
             folder_contents = os.listdir(tmpdirname)
             self.assertTrue(OV_XML_FILE_NAME in folder_contents)
             self.assertTrue(OV_XML_FILE_NAME.replace(".xml", ".bin") in folder_contents)
-            model = OVModelForSequenceClassification.from_pretrained(tmpdirname)
+            model = OVModelForSequenceClassification.from_pretrained(tmpdirname, ov_config={"NUM_STREAMS": 2})
+            # Test that PERFORMANCE_HINT is set to LATENCY by default even with ov_config provided
+            self.assertEqual(model.ov_config.get("PERFORMANCE_HINT"), "LATENCY")
+            self.assertEqual(model.request.get_property("PERFORMANCE_HINT"), "LATENCY")
 
         outputs = model(**tokens)
         self.assertTrue(torch.equal(loaded_model_outputs.logits, outputs.logits))
@@ -150,6 +156,9 @@ def test_load_from_hub_and_save_decoder_model(self, use_cache):
         tokens = tokenizer("This is a sample input", return_tensors="pt")
         loaded_model = OVModelForCausalLM.from_pretrained(model_id, use_cache=use_cache)
         self.assertIsInstance(loaded_model.config, PretrainedConfig)
+        # Test that PERFORMANCE_HINT is set to LATENCY by default
+        self.assertEqual(loaded_model.ov_config.get("PERFORMANCE_HINT"), "LATENCY")
+        self.assertEqual(loaded_model.request.get_compiled_model().get_property("PERFORMANCE_HINT"), "LATENCY")
         loaded_model_outputs = loaded_model(**tokens)
 
         with tempfile.TemporaryDirectory() as tmpdirname:
@@ -172,6 +181,11 @@ def test_load_from_hub_and_save_seq2seq_model(self):
         loaded_model = OVModelForSeq2SeqLM.from_pretrained(self.OV_SEQ2SEQ_MODEL_ID, compile=False)
         self.assertIsInstance(loaded_model.config, PretrainedConfig)
         loaded_model.to("cpu")
+        loaded_model.compile()
+        # Test that PERFORMANCE_HINT is set to LATENCY by default
+        self.assertEqual(loaded_model.ov_config.get("PERFORMANCE_HINT"), "LATENCY")
+        self.assertEqual(loaded_model.decoder.request.get_compiled_model().get_property("PERFORMANCE_HINT"), "LATENCY")
+
         loaded_model_outputs = loaded_model.generate(**tokens)
 
         with tempfile.TemporaryDirectory() as tmpdirname:
@@ -192,6 +206,10 @@ def test_load_from_hub_and_save_seq2seq_model(self):
     def test_load_from_hub_and_save_stable_diffusion_model(self):
         loaded_pipeline = OVStableDiffusionPipeline.from_pretrained(self.OV_DIFFUSION_MODEL_ID, compile=False)
         self.assertIsInstance(loaded_pipeline.config, Dict)
+        # Test that PERFORMANCE_HINT is set to LATENCY by default
+        self.assertEqual(loaded_pipeline.ov_config.get("PERFORMANCE_HINT"), "LATENCY")
+        loaded_pipeline.compile()
+        self.assertEqual(loaded_pipeline.unet.request.get_property("PERFORMANCE_HINT"), "LATENCY")
         batch_size, height, width = 2, 16, 16
         np.random.seed(0)
         inputs = {