fix transformers v5 & ministral compatibility

setup.py

@@ -122,11 +122,7 @@ def localversion_func(version: ScmVersion) -> str:
         ),
         ("tqdm>=4.66.3,<=4.67.1" if BUILD_TYPE == "release" else "tqdm>=4.66.3"),
         ("torch>=2.9.0,<=2.10.0" if BUILD_TYPE == "release" else "torch>=2.9.0"),
-        (
-            "transformers>=4.56.1,<=4.57.6"
-            if BUILD_TYPE == "release"
-            else "transformers>=4.56.1,<=4.57.6"
-        ),
+        "transformers>=4.56.1",
         ("datasets>=4.0.0,<=4.4.1" if BUILD_TYPE == "release" else "datasets>=4.0.0"),
         (
             # auto-round 0.9.1 cannot work with accelerate <1.10.0
@@ -148,7 +144,7 @@ def localversion_func(version: ScmVersion) -> str:
         (
             "compressed-tensors==0.13.0"
             if BUILD_TYPE == "release"
-            else "compressed-tensors>=0.13.1a2"
+            else "compressed-tensors==0.14.1a20260225"
         ),
     ],
     extras_require={

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

@@ -212,11 +212,8 @@ def _resolve_mappings(self, model: Module) -> list[SmoothQuantMapping]:
             )
 
         for mapping in self.mappings:
-            # we deliberately don't use the ignore list when matching mappings
-            # so that we can handle layers that need smoothing but not all operations
-            # we only skip if no layers in mapping would be smoothed.
             for *nested_balance_layers, smooth_layers in match_modules_set(
-                model, tree_leaves(mapping)
+                model, tree_leaves(mapping), ignore=self.ignore
             ):
                 if len(smooth_layers) > 1:
                     raise ValueError(

src/llmcompressor/transformers/compression/compressed_tensors_utils.py

@@ -93,6 +93,16 @@ def save_pretrained_wrapper(
                 # convert to accelerate offloaded for optimal saving with transformers
                 to_accelerate(model)
 
+                # Remove hf_device_map before saving to avoid a bug where
+                # transformers builds module_map with HF model keys but then applies
+                # revert_weight_conversion (which renames keys to checkpoint format)
+                # causing a KeyError when looking up renamed keys in module_map.
+                # After to_accelerate(), all parameters are real tensors (not meta,
+                # because init_hook is skipped), so model.state_dict() can be used
+                # directly without the module_map offload path.
+                if hasattr(model, "hf_device_map"):
+                    delattr(model, "hf_device_map")
+
                 # save (compressed) model structure
                 original_save_pretrained.__get__(model, model_class)(
                     save_directory,

src/llmcompressor/utils/dev.py

@@ -12,7 +12,10 @@
 from loguru import logger
 from safetensors.torch import save_file
 from transformers import AutoModelForCausalLM, PreTrainedModel
-from transformers.modeling_utils import TORCH_INIT_FUNCTIONS
+try:
+    from transformers.modeling_utils import TORCH_INIT_FUNCTIONS
+except ImportError:
+    from transformers.initialization import TORCH_INIT_FUNCTIONS
 from transformers.utils import SAFE_WEIGHTS_INDEX_NAME, WEIGHTS_INDEX_NAME
 
 __all__ = [