Update NNCF commit; tweak number of FQ for albert

Author: nikita-savelyevv

.github/workflows/test_openvino.yml

@@ -40,7 +40,6 @@ jobs:
           pip install .[openvino,openvino-tokenizers,tests,diffusers] onnxruntime
           pip install transformers==${{ matrix.transformers-version }}
           pip install -U --pre openvino==2024.5.0.dev20240923 openvino-tokenizers==2024.5.0.dev20240923 --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly/
-          pip install git+https://github.com/openvinotoolkit/nncf.git@2efda7dd16276c8d8f70fabfb9a94b4bcd7af08f
 
       - name: Test with Pytest
         env:

setup.py

@@ -61,8 +61,8 @@
 
 EXTRAS_REQUIRE = {
     "neural-compressor": ["neural-compressor[pt]>3.0", "accelerate", "transformers<=4.43.2"],
-    "openvino": ["openvino-nightly==2024.5.0.dev20240923", "nncf@git+https://github.com/openvinotoolkit/nncf.git@2efda7dd16276c8d8f70fabfb9a94b4bcd7af08f", "openvino-tokenizers[transformers]"],
-    "nncf": ["nncf@git+https://github.com/openvinotoolkit/nncf.git@2efda7dd16276c8d8f70fabfb9a94b4bcd7af08f"],
+    "openvino": ["openvino-nightly==2024.5.0.dev20240923", "nncf@git+https://github.com/nikita-savelyevv/nncf.git@release_2_13_1", "openvino-tokenizers[transformers]"],
+    "nncf": ["nncf@git+https://github.com/nikita-savelyevv/nncf.git@release_2_13_1"],
     "ipex": ["intel-extension-for-pytorch", "transformers>=4.39,<4.45"],
     "diffusers": ["diffusers"],
     "quality": QUALITY_REQUIRE,

tests/openvino/test_quantization.py

@@ -510,6 +510,8 @@ def test_ovmodel_4bit_auto_compression_with_config(
         with tempfile.TemporaryDirectory() as tmp_dir:
             quantization_config = OVWeightQuantizationConfig.from_dict(quantization_config)
             model = model_cls.from_pretrained(model_id, export=True, quantization_config=quantization_config)
+            import openvino
+            openvino.save_model(model.model, "test_opt/nncf_213/compressed.xml")
             if quantization_config.quant_method.lower() == "awq" or quantization_config.scale_estimation:
                 # TODO: Check that AWQ and SE was actually applied
                 pass
@@ -745,7 +747,7 @@ def preprocess_function(examples, tokenizer):
 
 
 class OVTrainerTest(unittest.TestCase):
-    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 64, 39),)
+    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 63, 39),)
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS)
     def test_aware_training_quantization(self, model_name, expected_fake_quantize, expected_int8):