Merge branch 'main' into feat/group-size-divisibility-check

   Signed-off-by: Avishek Goswami <avishek.goswami@ibm.com>

Author: Avishek Goswami

src/llmcompressor/modifiers/quantization/group_size_validation.py

@@ -95,12 +95,17 @@ def validate_group_size_divisibility(
     model: torch.nn.Module,
     resolved_targets: Set[str],
     ignore: list[str],
+    *,
+    bypass: bool = False,
 ) -> None:
     """
     Ensure targeted group/tensor_group layers have columns divisible by group_size.
 
     If any such layer has columns % group_size != 0, raises ValueError with layer FQNs.
+    When bypass is True, skips the check (e.g. for runtimes that support non-divisible).
     """
+    if bypass:
+        return
     indivisible = get_layers_indivisible_by_group_size(model, resolved_targets, ignore)
     if not indivisible:
         return

tests/llmcompressor/modifiers/quantization/test_group_size_validation.py

@@ -60,8 +60,7 @@ def test_get_layers_indivisible_by_group_size_finds_layer():
     from compressed_tensors.quantization import QuantizationScheme, QuantizationStrategy
     from compressed_tensors.quantization.quant_args import QuantizationArgs
 
-    # 1) Unit test: _layer_indivisible with a simple args object (no CT QuantizationArgs
-    #    attribute quirks; tests our logic in isolation).
+    # 1) Unit test: _layer_indivisible with strategy=GROUP (enum).
     # Linear(in_features, out_features) has weight.shape = (out_features, in_features);
     # we use shape[-1] (columns) for group divisibility, so use in_features=200.
     linear = torch.nn.Linear(