Implement `apply_clip` argument to `quantize()` (#427)

awq/models/base.py

@@ -136,6 +136,12 @@ class BaseAWQForCausalLM(nn.Module):
                 "This argument avoids real quantization by only applying the scales without quantizing down to FP16."
             ),
         ] = False,
+        apply_clip: Annotated[
+            bool,
+            Doc(
+                "Whether to apply clipping to the model during quantization. Some models may perform better with this set to False."
+            ),
+        ] = True,
     ):
         """
         The main quantization function that you can use to quantize your model.
@@ -173,6 +179,7 @@ class BaseAWQForCausalLM(nn.Module):
             duo_scaling,
             modules_to_not_convert=self.quant_config.modules_to_not_convert,
             export_compatible=export_compatible,
+            apply_clip=apply_clip,
         )
         self.quantizer.quantize()
 

awq/quantize/quantizer.py

@@ -40,6 +40,7 @@ class AwqQuantizer:
         duo_scaling,
         modules_to_not_convert=None,
         export_compatible=False,
+        apply_clip=True,
     ) -> None:
         self.awq_model = awq_model
         self.model = model
@@ -53,6 +54,7 @@ class AwqQuantizer:
         self.text_column = text_column
         self.duo_scaling = duo_scaling
         self.export_compatible = export_compatible
+        self.apply_clip = apply_clip
         self.modules_to_not_convert = (
             modules_to_not_convert if modules_to_not_convert is not None else []
         )
@@ -161,6 +163,7 @@ class AwqQuantizer:
             )
 
             # [STEP 3]: Compute and apply clipping list
+            if self.apply_clip:
                 clip_list = self._search_best_clip(
                     self.modules[i], named_linears, input_feat
                 )