compressed-tensors 0.10.2a20250609__py3-none-any.whl → 0.10.2a20250612__py3-none-any.whl

compressed_tensors/compressors/sparse_compressors/base.py

@@ -13,9 +13,8 @@
 # limitations under the License.
 
 import logging
-from typing import TYPE_CHECKING, Dict, Generator, Optional, Set, Tuple
+from typing import Dict, Generator, Optional, Set, Tuple
 
-import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.utils import (
     get_nested_mappings_from_state_dict,
@@ -27,10 +26,6 @@ from torch import Tensor
 from tqdm import tqdm
 
 
-if TYPE_CHECKING:
-    from compressed_tensors.quantization import QuantizationScheme
-
-
 __all__ = ["BaseSparseCompressor"]
 
 _LOGGER: logging.Logger = logging.getLogger(__name__)
@@ -205,16 +200,3 @@ class BaseSparseCompressor(BaseCompressor):
         return (
             name.endswith(".weight") and name[: -(len(".weight"))] in expanded_targets
         )
-
-    def decompress_module_from_state_dict(
-        self,
-        prefix: str,
-        state_dict: Dict[str, torch.Tensor],
-        scheme: "QuantizationScheme",
-    ) -> Dict[str, torch.Tensor]:
-        """
-        This function is implemented as a workaround because of how
-        `ModelCompressor.quantization_compressor` can be set to either
-        an instance of `BaseQuantizationCompressor` or `BaseSparseCompressor`.
-        """
-        return state_dict.copy()

compressed_tensors/compressors/sparse_compressors/dense.py

@@ -12,13 +12,18 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Dict, Generator, Tuple
+from typing import TYPE_CHECKING, Dict, Generator, Tuple
 
+import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.config import CompressionFormat
 from torch import Tensor
 
 
+if TYPE_CHECKING:
+    from compressed_tensors.quantization import QuantizationScheme
+
+
 @BaseCompressor.register(name=CompressionFormat.dense.value)
 class DenseCompressor(BaseCompressor):
     """
@@ -47,3 +52,16 @@ class DenseCompressor(BaseCompressor):
     ) -> Generator[Tuple[str, Dict[str, Tensor]], None, None]:
         for key, value in state_dict.items():
             yield key, value
+
+    def decompress_module_from_state_dict(
+        self,
+        prefix: str,
+        state_dict: Dict[str, torch.Tensor],
+        scheme: "QuantizationScheme",
+    ) -> Dict[str, torch.Tensor]:
+        """
+        This function is implemented as a workaround because of how
+        `ModelCompressor.quantization_compressor` can be set to either
+        an instance of `BaseQuantizationCompressor` or `DenseCompressor`.
+        """
+        return state_dict.copy()

compressed_tensors/quantization/lifecycle/apply.py

@@ -183,9 +183,7 @@ def apply_quantization_config(
                         replace_module(model, name, compressed_linear)
 
             # target matched - add layer and scheme to target list
-            submodule.quantization_scheme = _scheme_from_targets(
-                target_to_scheme, targets, name
-            )
+            submodule.quantization_scheme = scheme
 
             names_to_scheme[name] = submodule.quantization_scheme
 

compressed_tensors/transform/__init__.py

@@ -18,3 +18,8 @@
 from .transform_args import *
 from .transform_scheme import *
 from .transform_config import *
+
+from .factory.base import *
+from .factory.hadamard import *
+from .factory.matrix_multiply import *
+from .factory.random_hadamard import *

compressed_tensors/transform/factory/__init__.py

@@ -0,0 +1,13 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

compressed_tensors/transform/factory/base.py

@@ -0,0 +1,164 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from typing import Optional
+
+import torch
+import torch.nn.utils.parametrize as P
+from compressed_tensors.quantization.lifecycle import is_target  # TODO: move to utils
+from compressed_tensors.registry.registry import RegistryMixin, T
+from compressed_tensors.transform import (
+    TransformArgs,
+    TransformLocation,
+    TransformScheme,
+)
+from compressed_tensors.utils import (
+    align_module_device,
+    has_offloaded_params,
+    patch_attr,
+    register_offload_module,
+    update_offload_parameter,
+)
+from torch import Tensor
+from torch.nn import Module, Parameter
+
+
+__all__ = ["TransformFactory", "TransformBase"]
+
+
+class TransformFactory(RegistryMixin, ABC):
+    """
+    Abstract factory base used to create and apply transforms to a model
+
+    :param name: name associated with transform scheme
+    :param scheme: transform scheme which defines how transforms should be created
+    :param seed: random seed used to transform weight randomization
+    """
+
+    def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
+        self.name = name
+        self.scheme = scheme
+        self.generator = torch.Generator()
+        if seed is not None:
+            self.generator.manual_seed(seed)
+
+    @classmethod
+    def from_scheme(cls: type[T], scheme: TransformScheme, **kwargs) -> T:
+        """
+        Create a transform factory from a scheme
+
+        :param scheme: defines how transforms should be created
+        :param kwargs: TransformFactory constructor arguments
+        :return: subclass of `TransformFactory` corresponding to the scheme type
+        """
+        constructor = cls.get_value_from_registry(name=scheme.type)
+        return constructor(scheme=scheme, **kwargs)
+
+    @abstractmethod
+    def create_transform(self, module: Module, args: TransformArgs) -> "TransformBase":
+        """
+        Abstract method which defines how a transform should be created. May utilize
+        caching to maximize shared memory
+
+        :param module: parent module that transform will be applied to
+        :param args: defines how the transform will be applied to the module
+        :return: instance of TransformBase
+        """
+        raise NotImplementedError()
+
+    def apply_to_model(self, model: Module):
+        """
+        Create transforms and apply them to the model
+
+        :param model: module to apply transforms to
+        """
+        for arg in self.scheme.apply:
+            for name, module in list(model.named_modules()):
+                if is_target(name, module, arg.targets, arg.ignore):
+                    self._apply_to_module(module, arg)
+
+    def _apply_to_module(self, module: Module, args: TransformArgs):
+        """
+        Create transforms and apply them to the module
+
+        :param module: target module to apply transforms to
+        :param args: defines how the transform will be applied to the target module
+        """
+        # create transform as submodule
+        transform_name = f"{self.name}_{args.location.value}"
+        transform = self.create_transform(module, args)
+        register_offload_module(module, transform_name, transform)  # (1)
+
+        # register input transformation hook
+        if args.location == TransformLocation.INPUT:
+
+            def input_hook(_, args):
+                input = args[0]
+                return transform(input)
+
+            module.register_forward_pre_hook(input_hook, prepend=True)
+
+        # eagerly apply transformation to weight
+        elif args.location in (
+            TransformLocation.WEIGHT_INPUT,
+            TransformLocation.WEIGHT_OUTPUT,
+        ):
+            assert isinstance(module, torch.nn.Linear)
+            assert module.bias is None
+
+            with torch.no_grad(), align_module_device(module):
+                update_offload_parameter(module, "weight", transform(module.weight))
+
+            if self.scheme.requires_grad:
+                # for training, the weight changes with every forward pass
+                # so we can leverage parametrization to propagate the gradient
+                if has_offloaded_params(module):
+                    raise ValueError("Offloaded training is not supported")
+                P.register_parametrization(module, "weight", transform)
+
+        # register output transformation hook
+        elif args.location == TransformLocation.OUTPUT:
+
+            def output_hook(_, _input, output):
+                return transform(output)
+
+            module.register_forward_hook(output_hook)
+
+        # other locations such as q_attn and k_attn have not been implemented
+        else:
+            raise NotImplementedError()
+
+        # (1) even in the `weight` cases, this submodule attachment is needed in order
+        # to support saving in the frozen state
+
+
+class TransformBase(Module, ABC):
+    """
+    Represents the application of a transform accord to TransformArgs
+    """
+
+    args: TransformArgs
+    weight: Parameter
+
+    @abstractmethod
+    def forward(self, value: Tensor) -> Tensor:
+        raise NotImplementedError()
+
+    def right_inverse(self, value: Tensor) -> Tensor:
+        with patch_attr(self.args, "inverse", not self.args.inverse):
+            return self.forward(value)
+
+    def __repr__(self):
+        return f"{self.__class__.__name__}(inverse={self.args.inverse})"

compressed_tensors/transform/factory/hadamard.py

@@ -0,0 +1,79 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+from compressed_tensors.transform import TransformArgs, TransformScheme
+from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
+from compressed_tensors.transform.utils.hadamard import deterministic_hadamard_matrix
+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.helpers import ParameterizedDefaultDict
+from torch import Tensor, device, dtype
+from torch.nn import Linear, Module, Parameter
+
+
+@TransformFactory.register("hadamard")
+class HadamardFactory(TransformFactory):
+    """
+    Factory used to apply hadamard transforms to a model
+
+    :param name: name associated with transform scheme
+    :param scheme: transform scheme which defines how transforms should be created
+    :param seed: random seed used to transform weight randomization
+    """
+
+    def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
+        super().__init__(name, scheme, seed)
+        self.weights = ParameterizedDefaultDict(self._create_weight)
+
+    def create_transform(self, module: Module, args: TransformArgs):
+        """
+        Create a HadamardTransform for applying to a module. Transforms with the same
+        size, dtype, and device are cached
+
+        :param module: parent module that transform will be applied to
+        :param args: defines how the transform will be applied to the module
+        """
+        assert isinstance(module, Linear)
+        size = get_matrix_size(module, args.location)
+        dtype = module.weight.dtype
+        device = get_offloaded_device(module)
+
+        weight = self.weights[size, dtype, device]
+        return HadamardTransform(weight, args)
+
+    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
+        data = deterministic_hadamard_matrix(size)
+        data = data.to(dtype=dtype, device=device)
+        return Parameter(data, requires_grad=self.scheme.requires_grad)
+
+
+class HadamardTransform(TransformBase):
+    def __init__(self, weight: Parameter, args: TransformArgs):
+        super().__init__()
+        self.weight = weight
+        self.args = args
+
+    def forward(self, value: Tensor) -> Tensor:
+        if not self.args.inverse:
+            weight = self.weight
+        else:
+            weight = self.weight.T
+
+        return apply_transform_weight(weight, value, self.args.location)

compressed_tensors/transform/factory/matrix_multiply.py

@@ -0,0 +1,90 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import torch
+from compressed_tensors.transform import TransformArgs, TransformScheme
+from compressed_tensors.transform.factory.base import TransformBase, TransformFactory
+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.helpers import ParameterizedDefaultDict
+from torch import Tensor, device, dtype
+from torch.nn import Linear, Module, Parameter
+
+
+@TransformFactory.register("random-matrix")
+class RandomMatrixFactory(TransformFactory):
+    """
+    Factory used to apply random matrix transforms to a model
+
+    :param name: name associated with transform scheme
+    :param scheme: transform scheme which defines how transforms should be created
+    :param seed: random seed used to transform weight randomization
+    """
+
+    def __init__(self, name: str, scheme: TransformScheme, seed: Optional[int] = None):
+        super().__init__(name, scheme, seed)
+        self.weights = ParameterizedDefaultDict(self._create_weight)
+        self.inverses = ParameterizedDefaultDict(self._create_inverse)
+
+    def create_transform(self, module: Module, args: TransformArgs):
+        """
+        Create a RandomMatrixTransform for applying to a module. Transforms with the
+        same size, dtype, and device are cached
+
+        :param module: parent module that transform will be applied to
+        :param args: defines how the transform will be applied to the module
+        """
+        assert isinstance(module, Linear)
+        size = get_matrix_size(module, args.location)
+        dtype = module.weight.dtype
+        device = get_offloaded_device(module)
+
+        weight = self.weights[size, dtype, device]
+        if args.inverse:
+            weight = self.inverses[weight]
+
+        return RandomMatrixTransform(weight, args)
+
+    def _create_weight(self, size: int, dtype: dtype, device: device) -> Parameter:
+        data = torch.rand(
+            (size, size), generator=self.generator, dtype=dtype, device=device
+        )
+        return Parameter(data, requires_grad=self.scheme.requires_grad)
+
+    def _create_inverse(self, weight: Parameter) -> Parameter:
+        data = high_precision_invert(weight.data)
+        return Parameter(data, requires_grad=False)
+
+
+class RandomMatrixTransform(TransformBase):
+    def __init__(self, weight: Tensor, args: TransformArgs):
+        super().__init__()
+        self.weight = weight  # is an inverse if args.inverse
+        self.args = args
+
+    def forward(self, value: Tensor) -> Parameter:
+        return apply_transform_weight(self.weight, value, self.args.location)
+
+    def right_inverse(self, value: Tensor) -> Tensor:
+        inverse = high_precision_invert(self.weight)
+        return apply_transform_weight(inverse, value, self.args.location)
+
+
+def high_precision_invert(weight: Tensor) -> Tensor:
+    return torch.linalg.inv(weight.to(torch.float32)).to(weight.dtype)

compressed_tensors/transform/factory/random_hadamard.py

@@ -0,0 +1,34 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from compressed_tensors.transform import HadamardFactory, TransformFactory
+from compressed_tensors.transform.utils.hadamard import random_hadamard_matrix
+from torch import device, dtype
+from torch.nn import Parameter
+
+
+@TransformFactory.register("random-hadamard")
+class RandomHadamardFactory(HadamardFactory):
+    """
+    Factory used to apply random hadamard transforms to a model
+
+    :param name: name associated with transform scheme
+    :param scheme: transform scheme which defines how transforms should be created
+    :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, self.generator)
+        data = data.to(dtype=dtype, device=device)
+        return Parameter(data, requires_grad=self.scheme.requires_grad)

compressed_tensors/utils/offload.py

@@ -14,27 +14,29 @@
 """
 Utilities associated with offloading functionality provided by `accelerate`.
 
-| ----------------------------------------------------------------------------------------------------- | # noqa: E501
-| Operation | Without offloading support             | With offloading support                          | # noqa: E501
-| --------- | -------------------------------------- | ------------------------------------------------ | # noqa: E501
-| Add       | module.register_parameter(name, param) | register_offload_parameter(module, name, param)  | # noqa: E501
-| Check     | N/A                                    | has_offloaded_params(module)                     | # noqa: E501
-| Onload    | N/A                                    | with align_module_device(module)                 | # noqa: E501
-| Update    | module.name.data.copy_(new_data)       | update_offload_parameter(module, name, new_data) | # noqa: E501
-| Delete    | del module.name                        | delete_offload_parameter(module, name)           | # noqa: E501
-| ----------------------------------------------------------------------------------------------------- | # noqa: E501
+| ------------------------------------------------------------------------------------------------------ | # noqa: E501
+| Operation  | Without offloading support             | With offloading support                          | # noqa: E501
+| ---------- | -------------------------------------- | ------------------------------------------------ | # noqa: E501
+| Add        | module.register_parameter(name, param) | register_offload_parameter(module, name, param)  | # noqa: E501
+| Check      | N/A                                    | has_offloaded_params(module)                     | # noqa: E501
+| Onload     | N/A                                    | with align_module_device(module)                 | # noqa: E501
+| Update     | module.name.data.copy_(new_data)       | update_offload_parameter(module, name, new_data) | # noqa: E501
+| Delete     | del module.name                        | delete_offload_parameter(module, name)           | # noqa: E501
+| Add Module | module.register_module(name, child)    | register_offload_module(name, child)             | # noqa: E501
+| Del Module | del module.name                        | delete_offload_module(module, name)              | # noqa: E501
+| ------------------------------------------------------------------------------------------------------ | # noqa: E501
 """
 
 import contextlib
 import warnings
 from functools import wraps
-from typing import Any, Callable, Dict, Iterable, List, Literal, Optional, Union
+from operator import attrgetter
+from typing import Any, Callable, Dict, Iterable, Literal, Optional, Union
 
 import torch
 
 
 try:
-    from accelerate import dispatch_model
     from accelerate.hooks import (
         AlignDevicesHook,
         add_hook_to_module,
@@ -45,10 +47,12 @@ try:
     from accelerate.utils import (
         OffloadedWeightsLoader,
         PrefixedDataset,
+        find_tied_parameters,
         set_module_tensor_to_device,
     )
 
     _has_accelerate = True
+
 except ImportError:
     _has_accelerate = False
     AlignDevicesHook = None
@@ -58,8 +62,8 @@ except ImportError:
     PrefixedDataset = None
     set_module_tensor_to_device = None
     named_module_tensors = None
-    dispatch_model = None
     attach_align_device_hook = None
+    find_tied_parameters = None
 
 
 __all__ = [
@@ -78,14 +82,13 @@ __all__ = [
     "align_module_device",
     "register_offload_module",
     "delete_offload_module",
-    "force_cpu_offload",
+    "offloaded_dispatch",
 ]
 
 
 def check_accelerate(fallback: Any):
     def decorator(func: Callable[[Any], Any]):
         if not _has_accelerate:
-
             if fallback == "error":
 
                 @wraps(func)
@@ -479,46 +482,44 @@ def delete_offload_module(base: torch.nn.Module, name: str):
 
 
 @check_accelerate(fallback="error")
-def force_cpu_offload(
-    module: torch.nn.Module, execution_device: torch.device
+def offloaded_dispatch(
+    module: torch.nn.Module,
+    execution_device: torch.device,
+    offload_device: Union[torch.device, Literal["disk"]] = torch.device("cpu"),
 ) -> torch.nn.Module:
     """
-    Force cpu offloading a module, primarily used for testing
+    Unlike `dispatch_model`, this function forces a module (and its submodules) to
+    offload all parameters and replace them with meta tensors, utiliizing the
+    `AlignDevicesHook` to control onloading and offloading.
 
     :param module: module containing parameters to offload
-    :param execution_device: execution device submodules
-    :return: module with hooks to perform cpu offloading
+    :param execution_device: device that modules will onload and execute on
+    :param offload_device: device that module parameters will offload to
+    :return: module with offloading device hooks
     """
-    # edge case: there is a bug in `dispatch_model` which causes
-    # the function to only work if the model contains submodules
-    if next(module.children(), None) is None:
-        attach_align_device_hook(
-            module,
-            execution_device=execution_device,
-            offload=True,
-            weights_map=module.state_dict(),
-            tied_params_map={},
-        )
-        return module
-
-    device_map = {}
-
-    def collect_device_map(name: List[str], module: torch.nn.Module):
-        if next(module.parameters(recurse=False), None) is not None:
-            device_map[".".join(name)] = "cpu"
-            return
-
-        else:
-            for submodule_name, submodule in module.named_children():
-                name.append(submodule_name)
-                collect_device_map(name, submodule)
-                name.pop()
-
-    collect_device_map([], module)
-
-    return dispatch_model(
-        module, device_map, main_device=execution_device, force_hooks=True
+    if offload_device == "disk":
+        raise NotImplementedError("Disk offloading is not currently supported")
+
+    # create weights map
+    weights_map = OffloadedWeightsLoader(state_dict=module.state_dict(), device="cpu")
+
+    # create tied params map
+    tied_params = find_tied_parameters(module)
+    tied_params_map = {}
+    for group in tied_params:
+        for param_name in group:
+            data_ptr = attrgetter(param_name)(module).data_ptr()
+            tied_params_map[data_ptr] = {}
+
+    # recursively attaches hooks to all submodules
+    attach_align_device_hook(
+        module,
+        execution_device=execution_device,
+        offload=True,
+        weights_map=weights_map,
+        tied_params_map=tied_params_map,
     )
+    return module
 
 
 """ Upstreamed Functions """

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.10.2.a20250609'
+__version__ = version = '0.10.2.a20250612'
 __version_tuple__ = version_tuple = (0, 10, 2)

{compressed_tensors-0.10.2a20250609.dist-info → compressed_tensors-0.10.2a20250612.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.10.2a20250609
+Version: 0.10.2a20250612
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.10.2a20250609.dist-info → compressed_tensors-0.10.2a20250612.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=LTla_Pj0LNUtkAT7ETF33M0rqtov9c8Q6T5tjzwvHOs,523
+compressed_tensors/version.py,sha256=F2izwCTRKbiv1mAW6qD3TbJD5cXQrz4zRmew4qZ4Ud0,523
 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
@@ -12,8 +12,8 @@ compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=0
 compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=Gw-lVzk5jrKUlM5UTCiJBmhM5gHzB9mn8r298MVUbDI,6395
 compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=_66tQ8bxslDUdas-ULORXblPw9kdNNn1UJJU9-ZOGPY,11380
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
-compressed_tensors/compressors/sparse_compressors/base.py,sha256=4-7Hm2uJ4qYtl4hi6GqwegGg2gYNhbq7noledK3h86Q,8684
-compressed_tensors/compressors/sparse_compressors/dense.py,sha256=rPaxbP7P52prWNs4lGaiBbpNvsQLElFMwOrq1oBP2Yg,1733
+compressed_tensors/compressors/sparse_compressors/base.py,sha256=YNZWcHjDleAlqbgRZQ6oJf44MQb_UDNvJGOqhl26uFA,8098
+compressed_tensors/compressors/sparse_compressors/dense.py,sha256=-OujJ1e0iXBvxYVULrIGvAZ9l-IC0mXczZRnimQdgo4,2314
 compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py,sha256=4cwkj40SFrXEyE_jyt2xjz3R-gTdU9uMpMFUKo1pRBA,8643
 compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=S8vW0FI9ep_XtUQOxj0P5utJt3vKEYOHjWEPp-Xd9aY,5820
 compressed_tensors/compressors/sparse_quantized_compressors/__init__.py,sha256=4f_cwcKXB1nVVMoiKgTFAc8jAPjPLElo-Df_EDm1_xw,675
@@ -30,7 +30,7 @@ compressed_tensors/quantization/quant_args.py,sha256=2OpiiSdl4KidzNmjx7J8UlQoAYm
 compressed_tensors/quantization/quant_config.py,sha256=aFi6PKqmEX9iP9O8GVn3mEUjRDEwk_hOCbmmiq-j9oU,10198
 compressed_tensors/quantization/quant_scheme.py,sha256=IDWa1GWUbUdWCo8j78Jz6svYF5hLz89J2PVYWBBnXRc,7102
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
-compressed_tensors/quantization/lifecycle/apply.py,sha256=DOoxH4jM8r0270GGGUFOpRrgwaisiJi7TV-Q6E8qM8E,18067
+compressed_tensors/quantization/lifecycle/apply.py,sha256=v7D0TJU_eLT20Odn_J1VCPo2twll2ra-wxlEGBKB2OA,17990
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=JWOQ-03bsgh9_nnOLAjmLZ0S8bFQA-GjwDK6YUBwcrU,14883
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
@@ -39,22 +39,27 @@ compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5
 compressed_tensors/quantization/utils/helpers.py,sha256=bqxNL2NU1XVsSxNzmDVZE3zd65PlLFq1Ir-RHwff8G0,17840
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=0s15BxdGgzBv8RL4kUJCYcuDOFUh_KZYvNvLEeRqWTc,11956
-compressed_tensors/transform/__init__.py,sha256=oa5VdrE-GtDYYceXNSwj5X_ropoXLLukm6Aufcc9WhY,747
+compressed_tensors/transform/__init__.py,sha256=mtUOzwq-H7fXGi7sMmfe7zj83fjMg_LAu4DjTZ5vaHk,886
 compressed_tensors/transform/transform_args.py,sha256=8-Ab5_dFfdObfwVCgrWrEWcoVRzXmMBSDSUxjftI-Ss,3177
 compressed_tensors/transform/transform_config.py,sha256=6JA8VFcoz4EGHOev6thj51OuB7K2gKUUazWjrVPYDLc,2144
 compressed_tensors/transform/transform_scheme.py,sha256=c7NAuLDL0itFgUfBMNShegMI9bzKL7s4LR3QJTHsXLs,1733
+compressed_tensors/transform/factory/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
+compressed_tensors/transform/factory/base.py,sha256=yVrYWEnrr2RFWE5AjSNeXzO9aXc443dTNMVSxuLztz8,5940
+compressed_tensors/transform/factory/hadamard.py,sha256=tuFVpKsv__SeDj-QwKxtipLvjb993DOSIFvWcUh42Ww,3124
+compressed_tensors/transform/factory/matrix_multiply.py,sha256=0g4sYC_tOmCjOomae2gl54UTXiFdl0mCCkmbqIRX8yw,3613
+compressed_tensors/transform/factory/random_hadamard.py,sha256=6kqr9z6kFc-2qRNskhWRsLGTDT_NfNAkFcTLMqQJcWA,1484
 compressed_tensors/transform/utils/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
 compressed_tensors/transform/utils/hadamard.py,sha256=SmPZmnHtc5N36gJA5EbM1T65uf4w1_flgl7SWBeg_W8,5642
 compressed_tensors/transform/utils/utils.py,sha256=PRPTYwPs2nnNaQMq2GEbC4QYKHFKlZwaRyPgdDhl66g,2992
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
 compressed_tensors/utils/helpers.py,sha256=cPg-ikdeA92aIGwBONg8GmPNvcGlFhozyJVwsRiXBTA,11981
-compressed_tensors/utils/offload.py,sha256=aGcBv0UWLuauY_-HBSei5Gm0xsZOfWn7NeX76tcLf30,20319
+compressed_tensors/utils/offload.py,sha256=myV7iC75gA8A3BGgwR3uoeaJkIC9oigKp9CcqsHsVJc,20686
 compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVyah6BUUir_StT28,2530
 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-0.10.2a20250609.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.10.2a20250609.dist-info/METADATA,sha256=_OH8bysJ6ymrDyfkHXrKuZNb47qh38hfdCCOt7RRYeM,7005
-compressed_tensors-0.10.2a20250609.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-compressed_tensors-0.10.2a20250609.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.10.2a20250609.dist-info/RECORD,,
+compressed_tensors-0.10.2a20250612.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.10.2a20250612.dist-info/METADATA,sha256=541wdYU5905X69fwti-7pubCIzjsENQnbOxpJt4X2qQ,7005
+compressed_tensors-0.10.2a20250612.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+compressed_tensors-0.10.2a20250612.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.10.2a20250612.dist-info/RECORD,,