compressed-tensors 0.9.5a20250520__py3-none-any.whl → 0.9.5a20250521__py3-none-any.whl

compressed_tensors/quantization/lifecycle/initialize.py

@@ -304,6 +304,9 @@ def update_fused_layer_weight_global_scales(model: torch.nn.Module):
     ):
 
         if _is_attention_module(submodule):
+            # already fused/treated as one layer
+            if hasattr(submodule, "qkv_proj"):
+                continue
 
             if not _valid_fp4_quant(
                 [submodule.q_proj, submodule.v_proj, submodule.k_proj]

compressed_tensors/quantization/quant_args.py

@@ -269,8 +269,8 @@ class QuantizationArgs(BaseModel, use_enum_values=True):
                 observer = None
 
         elif observer is None:
-            # default to minmax for non-dynamic cases
-            observer = "minmax"
+            # default to mse for non-dynamic cases
+            observer = "mse"
 
         # write back modified values
         model.strategy = strategy

compressed_tensors/registry/registry.py

@@ -19,7 +19,7 @@ of neuralmagic utilities
 
 import importlib
 from collections import defaultdict
-from typing import Any, Dict, List, Optional, Type, Union
+from typing import Any, Dict, List, Optional, TypeVar, Union
 
 
 __all__ = [
@@ -32,8 +32,9 @@ __all__ = [
 ]
 
 
-_ALIAS_REGISTRY: Dict[Type, Dict[str, str]] = defaultdict(dict)
-_REGISTRY: Dict[Type, Dict[str, Any]] = defaultdict(dict)
+_ALIAS_REGISTRY: Dict[type, Dict[str, str]] = defaultdict(dict)
+_REGISTRY: Dict[type, Dict[str, Any]] = defaultdict(dict)
+T = TypeVar("", bound="RegistryMixin")
 
 
 def standardize_lookup_name(name: str) -> str:
@@ -159,7 +160,7 @@ class RegistryMixin:
         )
 
     @classmethod
-    def load_from_registry(cls, name: str, **constructor_kwargs) -> object:
+    def load_from_registry(cls: type[T], name: str, **constructor_kwargs) -> T:
         """
         :param name: name of registered class to load
         :param constructor_kwargs: arguments to pass to the constructor retrieved
@@ -172,7 +173,7 @@ class RegistryMixin:
         return constructor(**constructor_kwargs)
 
     @classmethod
-    def get_value_from_registry(cls, name: str):
+    def get_value_from_registry(cls: type[T], name: str) -> T:
         """
         :param name: name to retrieve from the registry
         :return: value from retrieved the registry for the given name, raises
@@ -200,7 +201,7 @@ class RegistryMixin:
 
 
 def register(
-    parent_class: Type,
+    parent_class: type,
     value: Any,
     name: Optional[str] = None,
     alias: Union[List[str], str, None] = None,
@@ -240,7 +241,7 @@ def register(
 
 
 def get_from_registry(
-    parent_class: Type, name: str, require_subclass: bool = False
+    parent_class: type, name: str, require_subclass: bool = False
 ) -> Any:
     """
     :param parent_class: class that the name is registered under
@@ -276,7 +277,7 @@ def get_from_registry(
     return retrieved_value
 
 
-def registered_names(parent_class: Type) -> List[str]:
+def registered_names(parent_class: type) -> List[str]:
     """
     :param parent_class: class to look up the registry of
     :return: all names registered to the given class
@@ -284,7 +285,7 @@ def registered_names(parent_class: Type) -> List[str]:
     return list(_REGISTRY[parent_class].keys())
 
 
-def registered_aliases(parent_class: Type) -> List[str]:
+def registered_aliases(parent_class: type) -> List[str]:
     """
     :param parent_class: class to look up the registry of
     :return: all aliases registered to the given class
@@ -297,7 +298,7 @@ def registered_aliases(parent_class: Type) -> List[str]:
 
 
 def register_alias(
-    name: str, parent_class: Type, alias: Union[str, List[str], None] = None
+    name: str, parent_class: type, alias: Union[str, List[str], None] = None
 ):
     """
     Updates the mapping from the alias(es) to the given name.
@@ -352,7 +353,7 @@ def _import_and_get_value_from_module(module_path: str, value_name: str) -> Any:
     return value
 
 
-def _validate_subclass(parent_class: Type, child_class: Type):
+def _validate_subclass(parent_class: type, child_class: type):
     if not issubclass(child_class, parent_class):
         raise ValueError(
             f"class {child_class} is not a subclass of the class it is "

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.9.5.a20250520'
+__version__ = version = '0.9.5.a20250521'
 __version_tuple__ = version_tuple = (0, 9, 5)

{compressed_tensors-0.9.5a20250520.dist-info → compressed_tensors-0.9.5a20250521.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.5a20250520
+Version: 0.9.5a20250521
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.9.5a20250520.dist-info → compressed_tensors-0.9.5a20250521.dist-info}/RECORD

@@ -1,6 +1,6 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=l0AqsnHO2iPLgqPeQ4oLbCAjxwSlFILMGGvhroOTz2w,521
+compressed_tensors/version.py,sha256=FJ5OPohL511E88TFF_Jipl_3ikvZ6NgmdrYxPbi2vo8,521
 compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1tWZewqVnOb4lM,825
 compressed_tensors/compressors/base.py,sha256=nvWsv4xEw1Tkxkxth6TmHplDYXfBeP22xWxOsZERyDY,7204
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
@@ -26,7 +26,7 @@ compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5y
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/linear/compressed_linear.py,sha256=1yo9RyjA0aQ--iuIknFfcSorJn43Mn4CoV-q4JlTJ_o,4052
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
-compressed_tensors/quantization/quant_args.py,sha256=VOVItkelnPAnSNHKhgeScrdtn33QZLbfyo3P-2DwhS4,10508
+compressed_tensors/quantization/quant_args.py,sha256=5-mq43RmbI81z9Xl9pYNv4bqIP5AIT65FgT--4ERsE8,10502
 compressed_tensors/quantization/quant_config.py,sha256=MxSUcb5dOqMN6LFyD5K2h8X0TvEtcWIAoiUJqD2dHGE,10159
 compressed_tensors/quantization/quant_scheme.py,sha256=Fx7Ma4bDlFB6OWkHKhOB6_0AOVIOPRgNE_qTwmDLSbc,6586
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
@@ -34,11 +34,11 @@ compressed_tensors/quantization/lifecycle/apply.py,sha256=-OKZ-FFFfIIoeGTrho8lXx
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=WY-HY5kXY2Zs9HMpaq44bpolQUAQ1ELrNZC7GM5C4jw,14494
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
-compressed_tensors/quantization/lifecycle/initialize.py,sha256=9OyGoRmLTWDiW8XUmc2ajv4tRz5NBvJmwWEOR4mk5lk,11920
+compressed_tensors/quantization/lifecycle/initialize.py,sha256=dWXxjYLemjmtrSnb8vyuvNoNTSm8ywmUswze3soKY4o,12041
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
 compressed_tensors/quantization/utils/helpers.py,sha256=w3Ucpdog88b0MnZdJ37VzgtYi1fqrwJafYdfWPc0hTk,16852
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
-compressed_tensors/registry/registry.py,sha256=vRcjVB1ITfSbfYUaGndBBmqhip_5vsS62weorVg0iXo,11896
+compressed_tensors/registry/registry.py,sha256=0s15BxdGgzBv8RL4kUJCYcuDOFUh_KZYvNvLEeRqWTc,11956
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
 compressed_tensors/utils/helpers.py,sha256=RrNvzD08naEjEiXdU-FdZjQVda1nQywu1hA_GCDj0vg,10415
 compressed_tensors/utils/offload.py,sha256=JNQ66_6vhSsizhlUaMgyEdBuFolYxbgUuT1mAZrCfKY,15436
@@ -46,8 +46,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.5a20250520.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.5a20250520.dist-info/METADATA,sha256=WEVbaXET4YHRFp-IASXkfWPZ9RBLGdKJYI1BKg0Bk6w,7004
-compressed_tensors-0.9.5a20250520.dist-info/WHEEL,sha256=zaaOINJESkSfm_4HQVc5ssNzHCPXhJm0kEUakpsEHaU,91
-compressed_tensors-0.9.5a20250520.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.5a20250520.dist-info/RECORD,,
+compressed_tensors-0.9.5a20250521.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.5a20250521.dist-info/METADATA,sha256=Xl6EbYwMlKhFyy6VXtxD2x0TsiTDG36YszGdub5wLqM,7004
+compressed_tensors-0.9.5a20250521.dist-info/WHEEL,sha256=zaaOINJESkSfm_4HQVc5ssNzHCPXhJm0kEUakpsEHaU,91
+compressed_tensors-0.9.5a20250521.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.5a20250521.dist-info/RECORD,,