Transition to OV 2024.4 (#914)

* Add GroupNormalization to ignored scope; tweak minicpm test

* Check tests on OV nightly

* Make ignored scope non-strict

* Update to openvino 2024.4.1 dev

* Minor reorder

* Style

optimum/intel/openvino/quantization.py

@@ -415,13 +415,20 @@ def _quantize_ovbasemodel(
         if calibration_dataset is None:
             raise ValueError("Calibration dataset is required to run quantization.")
 
+        # TODO: remove after update to NNCF 2.14
+        model_type = nncf.ModelType(quantization_config.model_type)
+        ignored_scope = quantization_config.get_ignored_scope_instance()
+        if model_type == nncf.ModelType.TRANSFORMER:
+            ignored_scope.types += ["GroupNormalization"]
+            ignored_scope.validate = False
+
         # Actual model quantization
         quantized_model = nncf.quantize(
             self.model.model,
             calibration_dataset,
             subset_size=quantization_config.num_samples,
-            ignored_scope=quantization_config.get_ignored_scope_instance(),
-            model_type=nncf.ModelType(quantization_config.model_type),
+            ignored_scope=ignored_scope,
+            model_type=model_type,
             preset=nncf.QuantizationPreset.PERFORMANCE if quantization_config.sym else nncf.QuantizationPreset.MIXED,
             fast_bias_correction=quantization_config.fast_bias_correction,
             advanced_parameters=nncf.AdvancedQuantizationParameters(
@@ -916,6 +923,11 @@ def _hybrid_quantization(
 
     ptq_ignored_scope = quantization_config.get_ignored_scope_instance()
     ptq_ignored_scope.names += ops_to_compress
+
+    # TODO: remove after update to NNCF 2.14
+    ptq_ignored_scope.types += ["GroupNormalization"]
+    ptq_ignored_scope.validate = False
+
     subset_size = quantization_config.num_samples if quantization_config.num_samples else 200
     quantized_model = nncf.quantize(
         model=compressed_model,

setup.py

@@ -61,7 +61,11 @@
 
 EXTRAS_REQUIRE = {
     "neural-compressor": ["neural-compressor[pt]>3.0", "accelerate"],
-    "openvino": ["openvino>=2023.3,<2024.4", "nncf>=2.11.0", "openvino-tokenizers[transformers]<2024.4"],
+    "openvino": [
+        "openvino==2024.4.1.dev20240926",
+        "nncf>=2.11.0",
+        "openvino-tokenizers[transformers]==2024.4.1.0.dev20240926",
+    ],
     "nncf": ["nncf>=2.11.0"],
     "ipex": ["intel-extension-for-pytorch", "transformers>=4.39,<4.45"],
     "diffusers": ["diffusers"],

tests/openvino/test_modeling.py

@@ -842,7 +842,8 @@ def test_compare_to_transformers(self, model_arch):
             transformers_outputs = transformers_model(**tokens)
 
         # Compare tensor outputs
-        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, equal_nan=True, atol=1e-4))
+        atol = 1e-3 if model_arch == "minicpm" else 1e-4
+        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, equal_nan=True, atol=atol))
 
         # Qwen tokenizer does not support padding