compressed-tensors 0.10.3a20250703__tar.gz → 0.10.3a20250707__tar.gz

{compressed_tensors-0.10.3a20250703/src/compressed_tensors.egg-info → compressed_tensors-0.10.3a20250707}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250703
+Version: 0.10.3a20250707
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
@@ -11,6 +11,7 @@ License-File: LICENSE
 Requires-Dist: torch>=1.7.0
 Requires-Dist: transformers
 Requires-Dist: pydantic>=2.0
+Requires-Dist: frozendict
 Provides-Extra: dev
 Requires-Dist: black==22.12.0; extra == "dev"
 Requires-Dist: isort==5.8.0; extra == "dev"

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/setup.py

@@ -88,7 +88,7 @@ def _setup_packages() -> List:
     )
 
 def _setup_install_requires() -> List:
-    return ["torch>=1.7.0", "transformers", "pydantic>=2.0"]
+    return ["torch>=1.7.0", "transformers", "pydantic>=2.0", "frozendict"]
 
 def _setup_extras() -> Dict:
     return {

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/transform/factory/hadamard.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Optional
+from typing import Optional, Union
 
 import torch
 from compressed_tensors.transform import TransformArgs, TransformScheme
@@ -22,7 +22,7 @@ from compressed_tensors.transform.utils.utils import (
     apply_transform_weight,
     get_matrix_size,
 )
-from compressed_tensors.utils import get_offloaded_device
+from compressed_tensors.utils import get_execution_device, get_offloaded_device
 from compressed_tensors.utils.helpers import ParameterizedDefaultDict
 from torch import Tensor, device, dtype
 from torch.nn import Linear, Module, Parameter
@@ -41,6 +41,7 @@ class HadamardFactory(TransformFactory):
     def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
         super().__init__(name, scheme, seed)
         self.weights = ParameterizedDefaultDict(self._create_weight)
+        self.perms = ParameterizedDefaultDict(self._create_permutation)
 
     def create_transform(self, module: Module, args: TransformArgs):
         """
@@ -54,26 +55,46 @@ class HadamardFactory(TransformFactory):
         size = get_matrix_size(module, args.location)
         dtype = module.weight.dtype
         device = get_offloaded_device(module)
+        exec_device = get_execution_device(module)
 
-        weight = self.weights[size, dtype, device]
-        return HadamardTransform(weight, args)
+        factory_kwargs = {"construct_device": exec_device}
+        weight = self.weights.get(size, dtype, device, factory_kwargs=factory_kwargs)
+        perm = self.perms[weight] if self.scheme.randomize else None
+        return HadamardTransform(weight, perm, args)
 
-    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = deterministic_hadamard_matrix(size, dtype, device)
-        data = data.to(dtype=dtype, device=device)
+    def _create_weight(
+        self,
+        size: int,
+        dtype: dtype,
+        device: device,
+        construct_device: device,
+    ) -> Parameter:
+        # construct on execution device, cache on offload device
+        data = deterministic_hadamard_matrix(size, dtype, construct_device)
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)
 
+    def _create_permutation(self, weight: Parameter) -> Parameter:
+        data = torch.randperm(weight.size(0), generator=self.generator)
+        return Parameter(data, requires_grad=False)
+
 
 class HadamardTransform(TransformBase):
-    def __init__(self, weight: Parameter, args: TransformArgs):
+    def __init__(
+        self, weight: Parameter, perm: Union[Parameter, None], args: TransformArgs
+    ):
         super().__init__()
         self.weight = weight
+        self.perm = perm
         self.args = args
 
     def forward(self, value: Tensor) -> Tensor:
-        if not self.args.inverse:
-            weight = self.weight
-        else:
-            weight = self.weight.T
+        weight = self.weight
+
+        if self.perm is not None:
+            weight = weight[self.perm][:, self.perm]
+
+        if self.args.inverse:
+            weight = weight.T
 
         return apply_transform_weight(weight, value, self.args.location)

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/transform/factory/matrix_multiply.py

@@ -62,6 +62,7 @@ class RandomMatrixFactory(TransformFactory):
         return RandomMatrixTransform(weight, args)
 
     def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
+        # TODO: verify that weight is invertible (has non-zero determinant)
         data = torch.rand(
             (size, size), generator=self.generator, dtype=dtype, device=device
         )

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/transform/factory/random_hadamard.py

@@ -28,7 +28,14 @@ class RandomHadamardFactory(HadamardFactory):
     :param seed: random seed used to transform weight randomization
     """
 
-    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
-        data = random_hadamard_matrix(size, dtype, device, self.generator)
-        data = data.to(dtype=dtype, device=device)
+    def _create_weight(
+        self,
+        size: int,
+        dtype: dtype,
+        device: device,
+        construct_device: device,
+    ) -> Parameter:
+        # construct on execution device, cache on offload device
+        data = random_hadamard_matrix(size, dtype, construct_device, self.generator)
+        data = data.to(device=device)
         return Parameter(data, requires_grad=self.scheme.requires_grad)

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/transform/transform_config.py

@@ -49,7 +49,7 @@ QUIP = TransformConfig(
                     inverse=True,
                 ),
             ],
-            randomize_modules=True,
+            randomize=True,
         ),
         "u": TransformScheme(
             type="hadamard",
@@ -62,7 +62,7 @@ QUIP = TransformConfig(
                     targets=["Linear"], location="output", inverse=True  # non-mergable
                 ),
             ],
-            randomize_modules=True,
+            randomize=True,
         ),
     }
 )

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/transform/transform_scheme.py

@@ -31,13 +31,12 @@ class TransformScheme(BaseModel):
         (see `Transforms.registered_names()`)
     :param apply: list of TransformationArgs containing the information about the
         modules that should be targeted by the specified transform
-    :param randomize_modules: True if unique transforms should be applied to each
-        unique module targeted by `apply`, otherwise reuse transform weights where
-        applicable
+    :param randomize: True if uniquely randomized transform weights should be used,
+        otherwise use identical transform weights where applicable
     :param requires_grad: True if weights include gradients for training
     """
 
     type: str
     apply: List[TransformArgs] = Field(default_factory=list)
-    randomize_modules: bool = Field(default=False)
+    randomize: bool = Field(default=False)
     requires_grad: bool = Field(default=False)

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/utils/helpers.py

@@ -15,10 +15,11 @@
 import contextlib
 import warnings
 from functools import wraps
-from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Mapping, Optional
 
 import numpy
 import torch
+from frozendict import frozendict
 from transformers import AutoConfig
 
 
@@ -373,11 +374,23 @@ class ParameterizedDefaultDict(dict):
 
     def __init__(self, default_factory: Callable[[Any], Any]):
         self.default_factory = default_factory
+        self._factory_kwargs = frozendict()
 
-    def __missing__(self, key):
+    def __missing__(self, key: Any) -> Any:
         if isinstance(key, tuple):
-            value = self.default_factory(*key)
+            value = self.default_factory(*key, **self._factory_kwargs)
         else:
-            value = self.default_factory(key)
+            value = self.default_factory(key, **self._factory_kwargs)
         self[key] = value
         return value
+
+    def get(self, *args, factory_kwargs: Mapping = frozendict()) -> Any:
+        """
+        Similar to `__getitem__`, but allows passing kwargs to factory function
+
+        :param \\*args: args whose tuple will value will be treated as key
+        :param factory_kwargs: keyword arguments to pass to `default_factory`
+        :return: dictionary entry for given key
+        """
+        with patch_attr(self, "_factory_kwargs", factory_kwargs):
+            return self[args]

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.3.a20250703'
+__version__ = version = '0.10.3.a20250707'
 __version_tuple__ = version_tuple = (0, 10, 3)

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707/src/compressed_tensors.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.3a20250703
+Version: 0.10.3a20250707
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.
@@ -11,6 +11,7 @@ License-File: LICENSE
 Requires-Dist: torch>=1.7.0
 Requires-Dist: transformers
 Requires-Dist: pydantic>=2.0
+Requires-Dist: frozendict
 Provides-Extra: dev
 Requires-Dist: black==22.12.0; extra == "dev"
 Requires-Dist: isort==5.8.0; extra == "dev"

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/src/compressed_tensors.egg-info/requires.txt

@@ -1,6 +1,7 @@
 torch>=1.7.0
 transformers
 pydantic>=2.0
+frozendict
 
 [accelerate]
 accelerate

{compressed_tensors-0.10.3a20250703 → compressed_tensors-0.10.3a20250707}/tests/test_transform/test_transform_scheme.py

@@ -24,7 +24,7 @@ def test_basic_scheme():
         type="hadamard",
         apply=[basic_args],
     )
-    assert not scheme.randomize_modules
+    assert not scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 1
     assert isinstance(scheme.apply[0], TransformArgs)
@@ -43,10 +43,10 @@ def test_multiple_groups_global():
     scheme = TransformScheme(
         type="hadamard",
         apply=[embedding_args, linear_args],
-        randomize_modules=True,
+        randomize=True,
     )
 
-    assert scheme.randomize_modules
+    assert scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 2
     assert isinstance(scheme.apply[0], TransformArgs)
@@ -69,6 +69,6 @@ def test_multiple_groups():
         apply=apply,
     )
 
-    assert not scheme.randomize_modules
+    assert not scheme.randomize
     assert scheme.type == "hadamard"
     assert len(scheme.apply) == 20