Don't have a sparsity_config if model is quantized

src/llmcompressor/transformers/compression/quantization_format.py

@@ -1,7 +1,6 @@
 from typing import Optional
 
 from compressed_tensors import CompressionFormat
-from compressed_tensors.config import SparsityCompressionConfig
 from compressed_tensors.quantization import QuantizationStrategy, QuantizationType
 from compressed_tensors.quantization.utils import (
     is_model_quantized,
@@ -16,7 +15,7 @@ def infer_quantization_format(
     model,
     quantization_format: Optional[str] = None,
     save_compressed: bool = False,
-    sparsity_config: Optional[SparsityCompressionConfig] = None,
+    sparsity_structure: Optional[str] = None,
 ) -> str:
     """
     Infers a quantization format based on model state and compression args
@@ -36,9 +35,7 @@ def infer_quantization_format(
 
     if save_compressed:
         weight_args, input_args = _get_unique_quant_args(model)
-        is_24_structure = (
-            sparsity_config and sparsity_config.sparsity_structure == "2:4"
-        )
+        is_24_structure = sparsity_structure == "2:4"
         is_weight_only = len(input_args) == 0 and len(weight_args) > 0
 
         if is_weight_only:  # w4a16 and w8a16

src/llmcompressor/transformers/compression/sparsity_config.py

@@ -84,6 +84,9 @@ def from_pretrained(
         :param compress: whether or not to compress the model on disk
         :return: compression config inferred from the model
         """
+        if is_model_quantized(model):
+            # compressing a sparse quantized model is not supported yet
+            return None
 
         global_sparsity = SparsityConfigMetadata.infer_global_sparsity(
             model, state_dict=state_dict
@@ -95,10 +98,8 @@ def from_pretrained(
         sparsity_structure = SparsityConfigMetadata.infer_sparsity_structure(
             model=model
         )
-        if is_model_quantized(model):
-            # compressing a sparse quantized model is not supported yet
-            format = CompressionFormat.dense.value
-        elif compress:
+
+        if compress:
             format = CompressionFormat.sparse_bitmask.value
         else:
             format = CompressionFormat.dense.value

src/llmcompressor/transformers/sparsification/compressed_tensors_utils.py

@@ -94,11 +94,16 @@ def save_pretrained_wrapper(
                     model, state_dict=state_dict, compress=False
                 )
 
+            sparsity_structure = (
+                sparsity_config.sparsity_structure
+                if sparsity_config is not None
+                else SparsityConfigMetadata.infer_sparsity_structure(model)
+            )
             quantization_format = infer_quantization_format(
                 model=model,
                 quantization_format=quantization_format,
                 save_compressed=save_compressed,
-                sparsity_config=sparsity_config,
+                sparsity_structure=sparsity_structure,
             )
             compressor = ModelCompressor.from_pretrained_model(
                 model,

tests/llmcompressor/transformers/compression/test_infer_quant_format.py

@@ -30,6 +30,8 @@ def test_infer_quant_format(preset, sparsity_structure, expected_format):
         module.quantization_scheme = quant_scheme
 
     inferred_format = infer_quantization_format(
-        dummy_model, save_compressed=True, sparsity_config=sparsity_config
+        dummy_model,
+        save_compressed=True,
+        sparsity_structure=sparsity_config.sparsity_structure,
     )
     assert inferred_format.value == expected_format