compressed-tensors 0.10.3a20250811__tar.gz → 0.10.3a20250812__tar.gz

{compressed_tensors-0.10.3a20250811/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250812}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250811
+Version: 0.10.3a20250812
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.3a20250811 → compressed_tensors-0.10.3a20250812}/src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -207,7 +207,7 @@ class ModelCompressor:
 
     @staticmethod
     def parse_sparsity_config(
-        compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"]
+        compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"],
     ) -> Union[Dict[str, Any], None]:
         """
         Parse sparsity config from quantization/compression config. Sparsity
@@ -227,7 +227,7 @@ class ModelCompressor:
 
     @staticmethod
     def parse_quantization_config(
-        compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"]
+        compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"],
     ) -> Union[Dict[str, Any], None]:
         """
         Parse quantization config from quantization/compression config. The
@@ -246,6 +246,7 @@ class ModelCompressor:
 
         quantization_config = deepcopy(compression_config)
         quantization_config.pop(SPARSITY_CONFIG_NAME, None)
+        quantization_config.pop(TRANSFORM_CONFIG_NAME, None)
 
         # some fields are required, even if a qconfig is not present
         # pop them off and if nothing remains, then there is no qconfig
@@ -674,7 +675,7 @@ class ModelCompressor:
 
         # serialize configs into json
         qconfig_data = (
-            self.quantization_config.model_dump(exclude=["quant_method", "format"])
+            self.quantization_config.model_dump(exclude=["quant_method"])
             if self.quantization_config is not None
             else {}
         )

{compressed_tensors-0.10.3a20250811 → compressed_tensors-0.10.3a20250812}/src/compressed_tensors/utils/offload.py

@@ -86,6 +86,7 @@ __all__ = [
     "offloaded_dispatch",
     "disable_offloading",
     "remove_dispatch",
+    "cast_to_device",
 ]
 
 
@@ -169,6 +170,19 @@ def update_parameter_data(
 """ Candidates for Upstreaming """
 
 
+def cast_to_device(device_spec: Union[int, torch.device]) -> torch.device:
+    """
+    Convert an integer device index or torch.device into a torch.device object.
+
+    :param device_spec: Device index (int) or torch.device object.
+                        Negative integers map to CPU.
+    :return: torch.device corresponding to the given device specification.
+    """
+    if isinstance(device_spec, int):
+        return torch.device(f"cuda:{device_spec}" if device_spec >= 0 else "cpu")
+    return device_spec
+
+
 def get_execution_device(module: torch.nn.Module) -> torch.device:
     """
     Get the device which inputs should be moved to before module execution.
@@ -179,7 +193,7 @@ def get_execution_device(module: torch.nn.Module) -> torch.device:
     """
     for submodule in module.modules():
         if has_offloaded_params(submodule):
-            return submodule._hf_hook.execution_device
+            return cast_to_device(submodule._hf_hook.execution_device)
 
         param = next(submodule.parameters(recurse=False), None)
         if param is not None:

{compressed_tensors-0.10.3a20250811 → compressed_tensors-0.10.3a20250812}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250811'
+__version__ = version = '0.10.3.a20250812'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250811 → compressed_tensors-0.10.3a20250812/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250811
+Version: 0.10.3a20250812
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.