Limit openvino version (#907)

* Limit openvino version

* updated openvino-tokenizers

* fix

Author: echarlaix

setup.py

@@ -60,7 +60,7 @@
 
 EXTRAS_REQUIRE = {
     "neural-compressor": ["neural-compressor[pt]>3.0", "accelerate", "transformers<=4.43.2"],
-    "openvino": ["openvino>=2023.3", "nncf>=2.11.0", "openvino-tokenizers[transformers]"],
+    "openvino": ["openvino>=2023.3,<2024.4", "nncf>=2.11.0", "openvino-tokenizers[transformers]<2024.4"],
     "nncf": ["nncf>=2.11.0"],
     "ipex": ["intel-extension-for-pytorch", "transformers>=4.39,<4.45"],
     "diffusers": ["diffusers"],

tests/openvino/test_quantization.py

@@ -732,7 +732,7 @@ def preprocess_function(examples, tokenizer):
 
 
 class OVTrainerTest(unittest.TestCase):
-    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 65, 39),)
+    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 64, 39),)
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS)
     def test_aware_training_quantization(self, model_name, expected_fake_quantize, expected_int8):