defaults

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

src/compressed_tensors/quantization/quant_args.py

@@ -263,8 +263,6 @@ def validate_model_after(model: "QuantizationArgs") -> "QuantizationArgs":
         block_structure = model.block_structure
         actorder = model.actorder
         dynamic = model.dynamic
-        observer = model.observer
-        dynamic = model.dynamic
 
         # infer strategy
         if strategy is None:
@@ -316,45 +314,8 @@ def validate_model_after(model: "QuantizationArgs") -> "QuantizationArgs":
                 "activation ordering"
             )
 
-        # infer observer w.r.t. dynamic
-        if dynamic:
-            supported_strategies = (
-                QuantizationStrategy.TOKEN,
-                QuantizationStrategy.TENSOR,
-                QuantizationStrategy.TENSOR_GROUP,
-                QuantizationStrategy.GROUP,
-            )
-            if strategy not in supported_strategies:
-                raise ValueError(
-                    f"One of {supported_strategies} must be used for dynamic quant."
-                )
-
-            if (
-                dynamic == DynamicType.LOCAL
-                and strategy != QuantizationStrategy.TENSOR_GROUP
-            ):
-                raise ValueError("local is only supported for strategy tensor_group")
-
-            if observer is not None:
-                if dynamic is True:  # checking if dynamic is True, not "local"
-                    if (
-                        observer != "memoryless"
-                    ):  # avoid annoying users with old configs
-                        warnings.warn(
-                            "No observer is used for dynamic quant., setting to None"
-                        )
-                    observer = None
-            else:
-                if dynamic == DynamicType.LOCAL:
-                    observer = "minmax"
-
-        elif observer is None:
-            # default to minmax for non-dynamic cases
-            observer = "minmax"
-
         # write back modified values
         model.strategy = strategy
-        model.observer = observer
         return model
 
     def pytorch_dtype(self) -> torch.dtype:
@@ -373,10 +334,6 @@ def pytorch_dtype(self) -> torch.dtype:
         else:
             raise ValueError(f"Invalid quantization type {self.type}")
 
-    @deprecated("QuantizationArgs.observer")
-    def get_observer(self) -> str:
-        return self.observer
-
     model_config = ConfigDict(extra="forbid")
 
 

src/compressed_tensors/quantization/quant_scheme.py

@@ -59,6 +59,7 @@ def validate_model_after(model: "QuantizationScheme") -> "QuantizationScheme":
         weights = model.weights
         format = model.format
 
+        # validate input args
         if inputs is not None:
             if inputs.strategy not in (
                 QuantizationStrategy.TOKEN,
@@ -84,15 +85,21 @@ def validate_model_after(model: "QuantizationScheme") -> "QuantizationScheme":
             if inputs.actorder is not None:
                 raise ValueError("Cannot apply actorder to input activations")
 
+            if inputs.observer is None:
+                inputs.observer
+
+        # validate output args
         if outputs is not None:
             if outputs.actorder is not None:
                 raise ValueError("Cannot apply actorder to output activations")
 
+        # validate format
         if format == CompressionFormat.mixed_precision.value:
             raise ValueError(
                 "mixed-precision cannot be set as a format for a QuantizationScheme"
             )
 
+        # validate matching group sizes
         if (
             inputs
             and weights
@@ -110,8 +117,35 @@ def validate_model_after(model: "QuantizationScheme") -> "QuantizationScheme":
                 stacklevel=2,
             )
 
+        # set observer defaults
+        model._validate_observers()
+
         return model
 
+    def _validate_observers(self):
+        inputs = self.input_activations
+        weights = self.weights
+        outputs = self.output_activations
+
+        if inputs is not None and inputs.observer is None:
+            if inputs.dynamic:
+                inputs.observer = "memoryless-minmax"
+            else:
+                inputs.observer = "static-minmax"
+
+        if weights is not None and weights.observer is None:
+            weights.observer = "memoryless-minmax"
+
+        if outputs is not None and outputs.observer is None:
+            if outputs.dynamic:
+                outputs.observer = "memoryless-minmax"
+            else:
+                outputs.observer = "static-minmax"
+
+        self.input_activations = inputs
+        self.weights = weights
+        self.output_activations = outputs
+
     model_config = ConfigDict(extra="forbid")