Constant folding is enabled for OpenVINOQuantizer

daniil-lyakhov · Dec 2, 2024 · 500e549 · 500e549
1 parent d765170
commit 500e549
Show file tree

Hide file tree

Showing 6 changed files with 13,613 additions and 17,097 deletions.
diff --git a/nncf/experimental/common/quantization/algorithms/quantizer/fx_quantizer.py b/nncf/experimental/common/quantization/algorithms/quantizer/fx_quantizer.py
@@ -41,7 +41,8 @@ def __init__(self, quantizer: Quantizer):
     def get_quantization_setup(self, model: torch.fx.GraphModule, nncf_graph: NNCFGraph) -> SingleConfigQuantizerSetup:
         anotated_model = deepcopy(model)
 
-        self._quantizer.transform_for_annotation(anotated_model)
+        # self._quantizer.transform_for_annotation is called in the nncf quantize_pt2e method
+        # before the nncf_graph building.
         self._quantizer.annotate(anotated_model)
         self._quantizer.validate(anotated_model)
         return self.get_quantizer_config_from_anotated_model(anotated_model)

diff --git a/nncf/experimental/common/quantization/algorithms/quantizer/openvino_quantizer.py b/nncf/experimental/common/quantization/algorithms/quantizer/openvino_quantizer.py
@@ -29,6 +29,7 @@
 from nncf.experimental.common.quantization.algorithms.quantizer.base_quantizer import NNCFQuantizer
 from nncf.experimental.torch.fx.nncf_graph_builder import GraphConverter
 from nncf.experimental.torch.fx.node_utils import get_graph_node_by_name
+from nncf.experimental.torch.fx.transformations import fold_constant_except_qdq
 from nncf.parameters import ModelType
 from nncf.parameters import QuantizationMode
 from nncf.parameters import TargetDevice
@@ -164,3 +165,7 @@ def _convert_nncf_qspec_to_inductor_qspec(self, qspec: NNCFQuantizerConfig, is_w
 
     def validate(self, model: torch.fx.GraphModule) -> None:
         pass
+
+    def transform_for_annotation(self, model: torch.fx.GraphModule) -> torch.fx.GraphModule:
+        fold_constant_except_qdq(model)
+        return model