replied to comments

Author: alexsu52

optimum/intel/openvino/quantization.py

@@ -30,6 +30,7 @@
 from nncf.torch.initialization import PTInitializingDataLoader
 from openvino._offline_transformations import compress_quantize_weights_transformation
 from openvino.runtime import Core, Tensor
+from torch.utils._pytree import tree_map
 from torch.utils.data import DataLoader, RandomSampler
 from transformers import DataCollator, PreTrainedModel, default_data_collator
 from transformers.pytorch_utils import Conv1D
@@ -350,9 +351,7 @@ def _quantize_ovcausallm(
             self.model.model,
             quantization_dataset,
             model_type=nncf.ModelType.TRANSFORMER if not kwargs.get("model_type") else kwargs.get("model_type"),
-            fast_bias_correction=(
-                True if not kwargs.get("fast_bias_correction") else kwargs.get("fast_bias_correction")
-            ),
+            fast_bias_correction=kwargs.get("fast_bias_correction", True),
             **kwargs,
         )
         self.model.model = quantized_model
@@ -410,8 +409,6 @@ def _quantize_torchmodel(
         stateful = ensure_stateful_is_available() and ensure_export_task_support_stateful(task)
 
         if weights_only:
-            from torch.utils._pytree import tree_map
-
             if stateful:
                 # patch model before weight compression
                 model = patch_model_with_bettertransformer(model)

tests/openvino/test_quantization.py

@@ -174,11 +174,6 @@ class OVWeightCompressionTest(unittest.TestCase):
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES_WITH_EXPECTED_8BIT_COMPRESSED_MATMULS)
     def test_automodel_weight_compression(self, model_cls, model_name, expected_pt_int8, expected_ov_int8):
-        import nncf
-
-        if nncf.__version__ == "2.8.0":
-            self.skipTest("https://github.com/openvinotoolkit/nncf/issues/2432")
-
         task = model_cls.export_feature
 
         with tempfile.TemporaryDirectory() as tmp_dir: