standardize, make modifier serializable

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

Author: kylesayrs

examples/transform/llama3_example.py

@@ -58,7 +58,7 @@ def tokenize(sample):
 #   * apply spinquant transforms to model in order to make quantization easier
 #   * quantize the weights to 4 bit with GPTQ with a group size 128
 recipe = [
-    SpinQuantModifier(rotations=["R1", "R2"], transform_type="random-hadamard"),
+    SpinQuantModifier(rotations=["R1", "R2"], transform_type="hadamard"),
     QuantizationModifier(targets="Linear", scheme="W4A16", ignore=["lm_head"]),
 ]
 

src/llmcompressor/modifiers/transform/quip/base.py

@@ -19,15 +19,15 @@
 from .norm_mappings import NORM_MAPPING_REGISTRY, NormMapping
 
 
-class SpinquantRotation(Enum):
+class SpinquantRotation(str, Enum):
     R1 = "R1"
     R2 = "R2"
     R3 = "R3"
     R4 = "R4"
 
 
-class SpinQuantModifier(Modifier):
-    rotations: Iterable[SpinquantRotation] = ("R1", "R2")
+class SpinQuantModifier(Modifier, use_enum_values=True):
+    rotations: List[SpinquantRotation] = Field(default_factory=lambda: ["R1", "R2"])
     transform_type: Literal["hadamard", "random-hadamard", "random-matrix"] = Field(
         default="hadamard"
     )
@@ -38,11 +38,12 @@ class SpinQuantModifier(Modifier):
     # override spinquant mappings with transform_config without overriding norms
     # we can combine these mappings, but it requires some more validation logic
     # maybe there's a reason to keep if other modifiers want norm fusing, idk
-    mappings: Optional[SpinQuantMappings] = None
-    norm_mappings: Optional[List[NormMapping]] = None
+    mappings: Optional[SpinQuantMappings] = Field(default=None, exclude=True)
+    norm_mappings: Optional[List[NormMapping]] = Field(default=None, exclude=True)
 
     # optional override for more fine-grained control
-    transform_config: Optional[TransformConfig] = None
+    # also included in recipe serialization
+    transform_config: Optional[TransformConfig] = Field(default=None)
 
     @field_validator("rotations", mode="before")
     def validate_rotations(cls, value):

src/llmcompressor/modifiers/transform/quip/template.py

@@ -8,6 +8,7 @@
 
 from llmcompressor.modifiers import Modifier
 from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.modifiers.transform import SpinQuantModifier
 
 if TYPE_CHECKING:
     from llmcompressor.args.dataset_arguments import DatasetArguments
@@ -60,5 +61,9 @@ def _infer_pipeline(modifiers: List[Modifier]) -> str:
             config = modifiers[0].resolve_quantization_config()
             if not config.requires_calibration_data():
                 return "datafree"
+            
+        # TODO: Remove hardcode
+        if len(modifiers) == 1 and isinstance(modifiers[0], SpinQuantModifier):
+            return "datafree"
 
         return "sequential"