do not overwrite schemes

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

Author: kylesayrs

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

@@ -138,20 +138,29 @@ def on_initialize(self, state: State, **kwargs) -> bool:
         :param state: session state storing input model and calibration data
         """
         # apply config to model and prepare calibration hooks
-        QuantizationMixin.attach_scheme_and_observers(self, state.model)
-        QuantizationMixin.register_calibration_hooks(self, state.model)
+        if QuantizationMixin.has_config(self):
+            QuantizationMixin.attach_scheme_and_observers(self, state.model)
+            QuantizationMixin.register_calibration_hooks(self, state.model)
 
         # prepare module names
         self._module_names = {m: name for name, m in state.model.named_modules()}
 
         # register hooks
+        added_hook = False
         for module in state.model.modules():
             if getattr_chain(module, "quantization_scheme.weights", None) is not None:
                 # HACK: previously, embeddings were not quantized because they were not
                 # accessible by the layer compressor. For now, we manually ignore it,
                 # but in the FUTURE this should be ignored by the user
                 if not isinstance(module, torch.nn.Embedding):
                     self.register_hook(module, self.calibrate_module, "forward")
+                    added_hook = True
+
+        if not added_hook:
+            raise ValueError(
+                "GPTQModifier requires a quantization config be specified by this "
+                "modifier or a modifier preceding it"
+            )
 
         # prepare for calibration
         state.model.apply(apply_calibration_status)

src/llmcompressor/modifiers/quantization/quantization/base.py

@@ -59,6 +59,11 @@ def on_initialize(self, state: State, **kwargs) -> bool:
         Then, according to the module's quantization scheme, observers and calibration
         hooks are added. These hooks are disabled until the modifier starts.
         """
+        if not QuantizationMixin.has_config(self):
+            raise ValueError(
+                "QuantizationModifier requires that quantization fields to be specified"
+            )
+
         QuantizationMixin.attach_scheme_and_observers(self, state.model)
         state.model.apply(disable_quantization)  # disable quantization until start
 

src/llmcompressor/modifiers/quantization/quantization/mixin.py

@@ -97,6 +97,15 @@ def register_calibration_hooks(self, model: torch.nn.Module):
         """
         model.apply(self._initialize_hooks)
 
+    def has_config(self) -> bool:
+        return not (
+            self.config_groups is None
+            and self.targets == ["Linear"]
+            and self.ignore == []
+            and self.scheme is None
+            and self.kv_cache_scheme is None
+        )
+
     def resolve_quantization_config(self) -> QuantizationConfig:
         """
         Returns the quantization config specified by this modifier

tests/llmcompressor/transformers/compression/recipes/new_quant_actorder_weight.yaml

@@ -1,7 +1,6 @@
 test_stage:
     quant_modifiers:
-        GPTQModifier:
-            block_size: 128
+        QuantizationModifier:
             ignore: ["lm_head", "model.layers.0.mlp.down_proj"]
             config_groups:
                 group_0:
@@ -14,4 +13,6 @@ test_stage:
                         actorder: "weight"
                     input_activations: null
                     output_activations: null
-                    targets: ["Linear"]
+                    targets: ["Linear"]
+        GPTQModifier:
+            block_size: 128

tests/llmcompressor/transformers/compression/recipes/new_quant_channel.yaml

@@ -1,7 +1,6 @@
 test_stage:
     quant_modifiers:
-        GPTQModifier:
-            block_size: 128
+        QuantizationModifier:
             ignore: ["lm_head", "model.layers.0.mlp.down_proj"]
             config_groups:
                 group_0:
@@ -12,4 +11,6 @@ test_stage:
                         strategy: "channel"
                     input_activations: null
                     output_activations: null
-                    targets: ["Linear"]
+                    targets: ["Linear"]
+        GPTQModifier:
+            block_size: 128

tests/llmcompressor/transformers/compression/recipes/new_quant_group.yaml

@@ -1,7 +1,6 @@
 test_stage:
     quant_modifiers:
-        GPTQModifier:
-            block_size: 128
+        QuantizationModifier:
             ignore: ["lm_head", "model.layers.0.mlp.down_proj"]
             config_groups:
                 group_0:
@@ -13,4 +12,6 @@ test_stage:
                         group_size: 128
                     input_activations: null
                     output_activations: null
-                    targets: ["Linear"]
+                    targets: ["Linear"]
+        GPTQModifier:
+            block_size: 128

tests/llmcompressor/transformers/compression/recipes/new_quant_weight.yaml

@@ -1,8 +1,6 @@
 test_stage:
     quant_modifiers:
-        GPTQModifier:
-            block_size: 128
-            targets: ["re:model.layers.\\d+$"]
+        QuantizationModifier:
             ignore: ["lm_head", "model.layers.0.mlp.down_proj"]
             config_groups:
                 group_0:
@@ -13,4 +11,7 @@ test_stage:
                         strategy: "tensor"
                     input_activations: null
                     output_activations: null
-                    targets: ["Linear", "Embedding"]
+                    targets: ["Linear", "Embedding"]
+        GPTQModifier:
+            block_size: 128
+            targets: ["re:model.layers.\\d+$"]