[GPTQ] Change actorder default to "static" (#1425)

## Purpose ##
* Use best defaults for GPTQ quantization

## Prerequisites ##
* #1453
* #1468

## Changes ##
* Set gptq actorder default to "static"

## Testing ##
* Ran llama w4a16 example to completion and validated the correct
activation ordering

---------

Signed-off-by: Kyle Sayers <[email protected]>
Co-authored-by: Dipika Sikka <[email protected]>

Author: kylesayrs

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

@@ -72,8 +72,9 @@ class GPTQModifier(Modifier, QuantizationMixin):
     :param block_size: Used to determine number of columns to compress in one pass
     :param dampening_frac: Amount of dampening to apply to H, as a fraction of the
         diagonal norm
-    :param actorder: order in which weight columns are quantized. For more information,
-        on actorder options, see https://github.com/vllm-project/vllm/pull/8135
+    :param actorder: order in which weight columns are quantized. Defaults to "static"
+        activation ordering, which achieves best accuracy recovery with no runtime cost.
+        For more information, see https://github.com/vllm-project/vllm/pull/8135
     :param offload_hessians: Set to True for decreased memory usage but increased
         runtime.
 
@@ -106,7 +107,7 @@ class GPTQModifier(Modifier, QuantizationMixin):
     sequential_targets: Union[str, List[str], None] = None
     block_size: int = 128
     dampening_frac: Optional[float] = 0.01
-    actorder: Optional[Union[ActivationOrdering, Sentinel]] = None
+    actorder: Optional[Union[ActivationOrdering, Sentinel]] = Sentinel("static")
     offload_hessians: bool = False
 
     # private variables
@@ -134,18 +135,17 @@ def resolve_actorder(existing):
                 return ActivationOrdering.STATIC if existing is None else existing
 
             # user-provided value always attempts to override
-            if self.actorder is not None:
-                if existing is None or self.actorder == existing:
-                    return self.actorder
-                raise ValueError(
-                    "Cannot resolve activation ordering when both "
-                    "`GPTQModifier.actorder` and `QuantizationScheme.actorder` "
-                    "are provided and differ. Either set `GPTQModifier.actorder = "
-                    "None` or remove `actorder` from config groups."
-                )
+            if existing is None or self.actorder == existing:
+                return self.actorder
 
-            # setting `GPTQModifier.actorder = None` does nothing
-            return existing
+            # if existing provided and conflicts
+            raise ValueError(
+                "Cannot resolve activation ordering when both "
+                "`GPTQModifier.actorder` and `QuantizationScheme.actorder` "
+                f"are provided and differ ({self.actorder}, {existing}). "
+                "Either unset `GPTQModifier.actorder` or "
+                "remove `actorder` from config groups."
+            )
 
         for scheme in config.config_groups.values():
             assert isinstance(scheme, QuantizationScheme)

tests/e2e/e2e_utils.py

@@ -70,7 +70,10 @@ def data_collator(batch):
         # a compatible preset sceme
         if quant_type == "GPTQ":
             oneshot_kwargs["recipe"] = GPTQModifier(
-                targets="Linear", scheme=scheme, ignore=["lm_head"]
+                targets="Linear",
+                scheme=scheme,
+                actorder=None,  # added for consistency with past testing configs
+                ignore=["lm_head"],
             )
         else:
             oneshot_kwargs["recipe"] = QuantizationModifier(

tests/e2e/vLLM/configs/w4a16_actorder_none.yaml

@@ -0,0 +1,9 @@
+cadence: "nightly"
+test_type: "regression"
+model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+recipe: tests/e2e/vLLM/recipes/actorder/recipe_w4a16_actorder_none.yaml
+dataset_id: openai/gsm8k
+dataset_config: main
+dataset_split: train
+scheme: W4A16_actorder_none
+save_dir: TinyLlama-1.1B-Chat-v1.0-actorder-group

tests/e2e/vLLM/configs/w4a16_actorder_none_qwen.yaml

@@ -0,0 +1,8 @@
+cadence: "nightly"
+test_type: "regression"
+model: Qwen/Qwen2.5-0.5B
+recipe: tests/e2e/vLLM/recipes/actorder/recipe_w4a16_actorder_none.yaml
+dataset_id: neuralmagic/LLM_compression_calibration
+dataset_split: train
+scheme: W4A16_actorder_none
+save_dir: Qwen2.5-0.5B-actorder-none

tests/e2e/vLLM/recipes/INT8/recipe_int8_channel_weight_dynamic_per_token.yaml

@@ -4,6 +4,7 @@ quant_stage:
       smoothing_strength: 0.8
     GPTQModifier:
       ignore: ["lm_head", "re:vision_tower.*", "re:multi_modal_projector.*", "re:visual.*", "re:vision_model.*", "re:model.visual.*"]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 8, type: int, symmetric: true, strategy: channel}

tests/e2e/vLLM/recipes/INT8/recipe_int8_channel_weight_static_per_tensor_act.yaml

@@ -4,6 +4,7 @@ quant_stage:
       smoothing_strength: 0.8
     GPTQModifier:
       ignore: [lm_head]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 8, type: int, symmetric: true, strategy: channel}

tests/e2e/vLLM/recipes/INT8/recipe_w8a8_dynamic_asym.yaml

@@ -9,6 +9,7 @@ quant_stage:
         - re:.*post_attention_layernorm
     GPTQModifier:
       ignore: [lm_head]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 8, type: int, symmetric: true, strategy: channel}

tests/e2e/vLLM/recipes/INT8/recipe_w8a8_static_asym.yaml

@@ -4,6 +4,7 @@ quant_stage:
       smoothing_strength: 0.8
     GPTQModifier:
       ignore: [lm_head]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 8, type: int, symmetric: true, strategy: channel}

tests/e2e/vLLM/recipes/WNA16/recipe_w4a16_channel_quant.yaml

@@ -2,6 +2,7 @@ quant_stage:
   quant_modifiers:
     GPTQModifier:
       ignore: [lm_head]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 4, type: int, symmetric: true, strategy: channel, dynamic: false}

tests/e2e/vLLM/recipes/WNA16/recipe_w4a16_group_quant_asym_awq.yaml

@@ -2,6 +2,7 @@ quant_stage:
   quant_modifiers:
     AWQModifier:
       ignore: [lm_head]
+      actorder: null
       config_groups:
         group_0:
           weights: {num_bits: 4, type: int, symmetric: false, strategy: "group", group_size: 128}