Full Scheme support

src/sparseml/modifiers/quantization/gptq/base.py

@@ -68,6 +68,10 @@ class GPTQModifier(Modifier):
         not be updated. Leave None to not disable observers during QAT. Default is None
     :param num_calibration_steps: Number of steps to run post training calibration for.
         When None, the entire calibration_dataloader is used
+    :param scheme: [Used, if a quantization modifier is not specified], the quantization
+        scheme to apply to the model, this is a dictionary that supports all keys from
+        QuantizationScheme except targets, which will be set to the targets parameter
+        set at the modifier level.
     """
 
     sequential_update: Optional[bool] = False
@@ -79,6 +83,7 @@ class GPTQModifier(Modifier):
     ignore: List[str] = Field(default_factory=list)
     disable_quantization_observer_epoch: Optional[float] = None
     num_calibration_steps: Optional[int] = None
+    scheme: Optional[Dict[str, Any]] = None
     compressible_layers_: Optional[List] = None
     quantization_modifier_: Any = None
 
@@ -156,6 +161,14 @@ def _build_quant_modifier(self, framework):
             if getattr(self, key, False)
         }
 
+        if self.scheme is not None:
+            # takes precedence over config_groups
+            targets = self.targets or ["Linear"]
+            config_group = QuantizationScheme.model_validate(
+                {"targets": targets, **self.scheme}
+            )
+            quant_args["config_groups"] = {"config_group_0": config_group}
+
         if "config_groups" not in quant_args:
             default_quant_scheme = QuantizationScheme.default_scheme(
                 targets=self.targets