[Bugfix] Improve pipeline inference

src/llmcompressor/pipelines/registry.py

@@ -7,7 +7,10 @@
 from torch.utils.data.dataloader import DataLoader
 
 from llmcompressor.modifiers import Modifier
-from llmcompressor.modifiers.quantization import QuantizationModifier
+from llmcompressor.modifiers.awq import AWQModifier
+from llmcompressor.modifiers.pruning.sparsegpt.sgpt_base import SparsityModifierBase
+from llmcompressor.modifiers.quantization import GPTQModifier, QuantizationModifier
+from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
 
 if TYPE_CHECKING:
     from llmcompressor.args.dataset_arguments import DatasetArguments
@@ -55,10 +58,19 @@ def from_modifiers(
 
     @staticmethod
     def _infer_pipeline(modifiers: list[Modifier]) -> str:
-        # only in the case of weight-only qmod quantization can we skip calibration
-        if len(modifiers) == 1 and isinstance(modifiers[0], QuantizationModifier):
-            config = modifiers[0].resolve_quantization_config()
-            if not config.requires_calibration_data():
-                return "datafree"
+        def _modifier_requires_calibration(modifier: Modifier):
+            if isinstance(
+                modifier,
+                (SmoothQuantModifier, SparsityModifierBase, GPTQModifier, AWQModifier),
+            ):
+                return True
+            elif isinstance(modifier, QuantizationModifier):
+                config = modifier.resolve_quantization_config()
+                return config.requires_calibration_data()
+            else:
+                return False
 
-        return "sequential"
+        if any(_modifier_requires_calibration(modifier) for modifier in modifiers):
+            return "sequential"
+        else:
+            return "datafree"

tests/llmcompressor/pipelines/test_registry.py

@@ -0,0 +1,35 @@
+import pytest
+
+from llmcompressor.modifiers.awq import AWQModifier
+from llmcompressor.modifiers.pruning import SparseGPTModifier, WandaPruningModifier
+from llmcompressor.modifiers.quantization import GPTQModifier, QuantizationModifier
+from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
+from llmcompressor.modifiers.transform import QuIPModifier, SpinQuantModifier
+from llmcompressor.pipelines import (
+    CalibrationPipeline,
+    DataFreePipeline,
+    SequentialPipeline,
+)
+
+
+@pytest.mark.parametrize(
+    "modifiers,exp_pipeline",
+    [
+        ([QuantizationModifier(scheme="FP8")], SequentialPipeline),
+        ([QuantizationModifier(scheme="W4A16")], DataFreePipeline),
+        ([GPTQModifier(scheme="FP8")], SequentialPipeline),
+        ([GPTQModifier(scheme="W4A16")], SequentialPipeline),
+        ([SmoothQuantModifier(), GPTQModifier(scheme="W4A16")], SequentialPipeline),
+        ([AWQModifier(scheme="W4A16")], SequentialPipeline),
+        ([AWQModifier(scheme="FP8")], SequentialPipeline),
+        ([SparseGPTModifier(sparsity=1.0)], SequentialPipeline),
+        ([WandaPruningModifier(sparsity=1.0)], SequentialPipeline),
+        ([QuIPModifier()], DataFreePipeline),
+        ([SpinQuantModifier()], DataFreePipeline),
+        ([QuIPModifier(), QuantizationModifier(scheme="FP8")], SequentialPipeline),
+        ([QuIPModifier(), QuantizationModifier(scheme="W4A16")], DataFreePipeline),
+    ],
+)
+def test_infer_pipeline(modifiers, exp_pipeline):
+    pipeline = CalibrationPipeline.from_modifiers(modifiers)
+    assert isinstance(pipeline, exp_pipeline)