compressed-tensors-nightly 0.8.1.20250105__py3-none-any.whl → 0.8.1.20250107__py3-none-any.whl

compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -17,8 +17,9 @@ import logging
 import operator
 import os
 import re
+from contextlib import contextmanager
 from copy import deepcopy
-from typing import TYPE_CHECKING, Any, Dict, Optional, TypeVar, Union
+from typing import TYPE_CHECKING, Any, Dict, Optional, Set, TypeVar, Union
 
 import compressed_tensors
 import torch
@@ -38,6 +39,7 @@ from compressed_tensors.quantization import (
     apply_quantization_config,
     load_pretrained_quantization,
 )
+from compressed_tensors.quantization.lifecycle import expand_sparse_target_names
 from compressed_tensors.quantization.quant_args import QuantizationArgs
 from compressed_tensors.quantization.utils import (
     is_module_quantized,
@@ -104,7 +106,6 @@ class ModelCompressor:
         """
         config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
         compression_config = getattr(config, QUANTIZATION_CONFIG_NAME, None)
-
         return cls.from_compression_config(compression_config)
 
     @classmethod
@@ -282,8 +283,14 @@ class ModelCompressor:
                 )
 
         if self.sparsity_compressor is not None:
+            sparse_compression_targets: Set[str] = expand_sparse_target_names(
+                model=model,
+                targets=self.sparsity_config.targets,
+                ignore=self.sparsity_config.ignore,
+            )
             compressed_state_dict = self.sparsity_compressor.compress(
-                compressed_state_dict
+                compressed_state_dict,
+                compression_targets=sparse_compression_targets,
             )
 
         # HACK: Override the dtype_byte_size function in transformers to
@@ -301,23 +308,41 @@ class ModelCompressor:
         :param model: pytorch model to load decompressed weights into
         """
         model_path = get_safetensors_folder(model_path)
+        sparse_decompressed = False
+
         if self.sparsity_compressor is not None:
+            # Sparse decompression is applied on the model_path
             dense_gen = self.sparsity_compressor.decompress(model_path)
             self._replace_weights(dense_gen, model)
             setattr(model, SPARSITY_CONFIG_NAME, self.sparsity_compressor.config)
+            sparse_decompressed = True
 
         if self.quantization_compressor is not None:
-            names_to_scheme = apply_quantization_config(model, self.quantization_config)
-            load_pretrained_quantization(model, model_path)
+            # Temporarily set quantization status to FROZEN to prevent
+            # quantization during apply_quantization_config. This ensures
+            # that the dtypes of the weights are not unintentionally updated.
+            # The status is restored after quantization params are loaded.
+            with override_quantization_status(
+                self.quantization_config, QuantizationStatus.FROZEN
+            ):
+                names_to_scheme = apply_quantization_config(
+                    model, self.quantization_config
+                )
+                load_pretrained_quantization(model, model_path)
+
+            model_path_or_state_dict = (
+                model.state_dict() if sparse_decompressed else model_path
+            )
+
             dense_gen = self.quantization_compressor.decompress(
-                model_path, names_to_scheme=names_to_scheme
+                model_path_or_state_dict, names_to_scheme=names_to_scheme
             )
             self._replace_weights(dense_gen, model)
 
-            def update_status(module):
+            def freeze_quantization_status(module):
                 module.quantization_status = QuantizationStatus.FROZEN
 
-            model.apply(update_status)
+            model.apply(freeze_quantization_status)
             setattr(model, QUANTIZATION_CONFIG_NAME, self.quantization_config)
 
     def update_config(self, save_directory: str):
@@ -367,12 +392,26 @@ class ModelCompressor:
         with open(config_file_path, "w") as config_file:
             json.dump(config_data, config_file, indent=2, sort_keys=True)
 
-    def _replace_weights(self, dense_weight_generator, model):
+    def _replace_weights(self, dense_weight_generator, model: Module):
+        """
+        Replace the weights of the model with the
+        provided dense weights.
+
+        This method iterates over the dense_weight_generator and
+        updates the corresponding weights in the model. If a parameter
+        name does not exist in the model, it will be skipped.
+
+        :param dense_weight_generator (generator): A generator that yields
+            tuples of (name, data), where 'name' is the parameter name and
+            'data' is the updated param data
+        :param model: The model whose weights are to be updated.
+        """
         for name, data in tqdm(dense_weight_generator, desc="Decompressing model"):
             split_name = name.split(".")
             prefix, param_name = ".".join(split_name[:-1]), split_name[-1]
             module = operator.attrgetter(prefix)(model)
-            update_parameter_data(module, data, param_name)
+            if hasattr(module, param_name):
+                update_parameter_data(module, data, param_name)
 
 
 def map_modules_to_quant_args(model: Module) -> Dict[str, QuantizationArgs]:
@@ -402,3 +441,23 @@ def new_dtype_byte_size(dtype):
         raise ValueError(f"`dtype` is not a valid dtype: {dtype}.")
     bit_size = int(bit_search.groups()[0])
     return bit_size // 8
+
+
+@contextmanager
+def override_quantization_status(
+    config: QuantizationConfig, status: QuantizationStatus
+):
+    """
+    Within this context, the quantization status will be set to the
+    supplied status. After the context exits, the original status
+    will be restored.
+
+    :param config: the quantization config to override
+    :param status: the status to temporarily set
+    """
+    original_status = config.quantization_status
+    config.quantization_status = status
+    try:
+        yield
+    finally:
+        config.quantization_status = original_status

compressed_tensors/compressors/quantized_compressors/base.py

@@ -13,12 +13,17 @@
 # limitations under the License.
 
 import logging
-from typing import Dict, Generator, Tuple
+from pathlib import Path
+from typing import Any, Dict, Generator, Tuple, Union
 
 import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.quantization import QuantizationArgs
-from compressed_tensors.utils import get_nested_weight_mappings, merge_names
+from compressed_tensors.utils import (
+    get_nested_mappings_from_state_dict,
+    get_nested_weight_mappings,
+    merge_names,
+)
 from safetensors import safe_open
 from torch import Tensor
 from tqdm import tqdm
@@ -113,7 +118,7 @@ class BaseQuantizationCompressor(BaseCompressor):
 
     def decompress(
         self,
-        path_to_model_or_tensors: str,
+        path_to_model_or_tensors: Union[str, Path, Dict[str, Any]],
         names_to_scheme: Dict[str, QuantizationArgs],
         device: str = "cpu",
     ) -> Generator[Tuple[str, Tensor], None, None]:
@@ -121,15 +126,25 @@ class BaseQuantizationCompressor(BaseCompressor):
         Reads a compressed state dict located at path_to_model_or_tensors
         and returns a generator for sequentially decompressing back to a
         dense state dict
-
         :param path_to_model_or_tensors: path to compressed safetensors model (directory
             with one or more safetensors files) or compressed tensors file
         :param names_to_scheme: quantization args for each quantized weight
         :param device: optional device to load intermediate weights into
         :return: compressed state dict
         """
+        if isinstance(path_to_model_or_tensors, (str, Path)):
+            yield from self._decompress_from_path(
+                path_to_model_or_tensors, names_to_scheme, device
+            )
+
+        else:
+            yield from self._decompress_from_state_dict(
+                path_to_model_or_tensors, names_to_scheme
+            )
+
+    def _decompress_from_path(self, path_to_model, names_to_scheme, device):
         weight_mappings = get_nested_weight_mappings(
-            path_to_model_or_tensors, self.COMPRESSION_PARAM_NAMES
+            path_to_model, self.COMPRESSION_PARAM_NAMES
         )
         for weight_name in weight_mappings.keys():
             weight_data = {}
@@ -137,6 +152,21 @@ class BaseQuantizationCompressor(BaseCompressor):
                 full_name = merge_names(weight_name, param_name)
                 with safe_open(safe_path, framework="pt", device=device) as f:
                     weight_data[param_name] = f.get_tensor(full_name)
+            if "weight_scale" in weight_data:
+                quant_args = names_to_scheme[weight_name]
+                decompressed = self.decompress_weight(
+                    compressed_data=weight_data, quantization_args=quant_args
+                )
+                yield merge_names(weight_name, "weight"), decompressed
+
+    def _decompress_from_state_dict(self, state_dict, names_to_scheme):
+        weight_mappings = get_nested_mappings_from_state_dict(
+            state_dict, self.COMPRESSION_PARAM_NAMES
+        )
+        for weight_name in weight_mappings.keys():
+            weight_data = {}
+            for param_name, param_value in weight_mappings[weight_name].items():
+                weight_data[param_name] = param_value
 
             if "weight_scale" in weight_data:
                 quant_args = names_to_scheme[weight_name]

compressed_tensors/compressors/quantized_compressors/naive_quantized.py

@@ -68,9 +68,9 @@ class NaiveQuantizationCompressor(BaseQuantizationCompressor):
         self,
         weight: Tensor,
         scale: Tensor,
+        quantization_args: QuantizationArgs,
         zero_point: Optional[Tensor] = None,
         g_idx: Optional[torch.Tensor] = None,
-        quantization_args: Optional[QuantizationArgs] = None,
         device: Optional[torch.device] = None,
     ) -> Dict[str, torch.Tensor]:
         """
@@ -78,9 +78,9 @@ class NaiveQuantizationCompressor(BaseQuantizationCompressor):
 
         :param weight: uncompressed weight tensor
         :param scale: quantization scale for weight
+        :param quantization_args: quantization parameters for weight
         :param zero_point: quantization zero point for weight
         :param g_idx: optional mapping from column index to group index
-        :param quantization_args: quantization parameters for weight
         :param device: optional device to move compressed output to
         :return: dictionary of compressed weight data
         """

compressed_tensors/compressors/quantized_compressors/pack_quantized.py

@@ -68,9 +68,9 @@ class PackedQuantizationCompressor(BaseQuantizationCompressor):
         self,
         weight: Tensor,
         scale: Tensor,
+        quantization_args: QuantizationArgs,
         zero_point: Optional[Tensor] = None,
         g_idx: Optional[torch.Tensor] = None,
-        quantization_args: Optional[QuantizationArgs] = None,
         device: Optional[torch.device] = None,
     ) -> Dict[str, torch.Tensor]:
         """
@@ -78,9 +78,9 @@ class PackedQuantizationCompressor(BaseQuantizationCompressor):
 
         :param weight: uncompressed weight tensor
         :param scale: quantization scale for weight
+        :param quantization_args: quantization parameters for weight
         :param zero_point: quantization zero point for weight
         :param g_idx: optional mapping from column index to group index
-        :param quantization_args: quantization parameters for weight
         :param device: optional device to move compressed output to
         :return: dictionary of compressed weight data
         """

compressed_tensors/compressors/sparse_compressors/base.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 import logging
-from typing import Dict, Generator, Tuple
+from typing import Dict, Generator, Optional, Set, Tuple
 
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.utils import get_nested_weight_mappings, merge_names
@@ -30,7 +30,8 @@ _LOGGER: logging.Logger = logging.getLogger(__name__)
 class BaseSparseCompressor(BaseCompressor):
     """
     Base class representing a sparse compression algorithm. Each child class should
-    implement compression_param_info, compress_weight and decompress_weight.
+    implement compression_param_info, compress_weight and decompress_weight; child
+    classes should also define COMPRESSION_PARAM_NAMES.
 
     Compressors support compressing/decompressing a full module state dict or a single
     quantized PyTorch leaf module.
@@ -59,11 +60,17 @@ class BaseSparseCompressor(BaseCompressor):
     :param config: config specifying compression parameters
     """
 
-    def compress(self, model_state: Dict[str, Tensor]) -> Dict[str, Tensor]:
+    def compress(
+        self,
+        model_state: Dict[str, Tensor],
+        compression_targets: Optional[Set[str]] = None,
+    ) -> Dict[str, Tensor]:
         """
         Compresses a dense state dict using bitmask compression
 
         :param model_state: state dict of uncompressed model
+        :param compression_targets: optional set of layer prefixes to compress,
+            otherwise compress all layers (for backwards compatibility)
         :return: compressed state dict
         """
         compressed_dict = {}
@@ -71,7 +78,14 @@ class BaseSparseCompressor(BaseCompressor):
             f"Compressing model with {len(model_state)} parameterized layers..."
         )
         for name, value in tqdm(model_state.items(), desc="Compressing model"):
-            compression_data = self.compress_weight(name, value)
+            if not self.should_compress(name, compression_targets):
+                compressed_dict[name] = value
+                continue
+            prefix = name
+            if prefix.endswith(".weight"):
+                prefix = prefix[: -(len(".weight"))]
+
+            compression_data = self.compress_weight(prefix, value)
             for key in compression_data.keys():
                 if key in compressed_dict:
                     _LOGGER.warn(
@@ -97,8 +111,10 @@ class BaseSparseCompressor(BaseCompressor):
         :param device: device to load decompressed weights onto
         :return: iterator for generating decompressed weights
         """
-        weight_mappings = get_nested_weight_mappings(
-            path_to_model_or_tensors, self.COMPRESSION_PARAM_NAMES
+        weight_mappings, ignored_params = get_nested_weight_mappings(
+            path_to_model_or_tensors,
+            self.COMPRESSION_PARAM_NAMES,
+            return_unmatched_params=True,
         )
         for weight_name in weight_mappings.keys():
             weight_data = {}
@@ -107,4 +123,26 @@ class BaseSparseCompressor(BaseCompressor):
                 with safe_open(safe_path, framework="pt", device=device) as f:
                     weight_data[param_name] = f.get_tensor(full_name)
             decompressed = self.decompress_weight(weight_data)
-            yield weight_name, decompressed
+            yield merge_names(weight_name, "weight"), decompressed
+
+        for ignored_param_name, safe_path in ignored_params.items():
+            with safe_open(safe_path, framework="pt", device=device) as f:
+                value = f.get_tensor(ignored_param_name)
+            yield ignored_param_name, value
+
+    @staticmethod
+    def should_compress(name: str, expanded_targets: Optional[Set[str]] = None) -> bool:
+        """
+        Check if a parameter should be compressed.
+        Currently, this only returns True for weight parameters.
+
+        :param name: name of the parameter
+        :param expanded_targets: set of layer prefixes to compress
+        :return: whether or not the parameter should be compressed
+        """
+        if expanded_targets is None:
+            return name.endswith(".weight")
+
+        return (
+            name.endswith(".weight") and name[: -(len(".weight"))] in expanded_targets
+        )

compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py

@@ -19,6 +19,7 @@ import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.compressors.sparse_compressors.base import BaseSparseCompressor
 from compressed_tensors.config import CompressionFormat
+from compressed_tensors.quantization import FP8_DTYPE
 from compressed_tensors.utils import merge_names
 from torch import Tensor
 
@@ -134,9 +135,14 @@ def bitmask_compress(tensor: Tensor) -> Tuple[Tensor, Tensor, Tensor]:
     bytemasks = tensor != 0
     row_counts = bytemasks.sum(dim=-1)
     row_offsets = torch.cumsum(row_counts, 0) - row_counts
-    values = tensor[bytemasks]
+    if tensor.dtype == FP8_DTYPE:
+        # acces raw bytes of the tensor
+        tensor_view = tensor.view(torch.int8)
+        values = tensor_view[bytemasks]
+        values = values.view(FP8_DTYPE)
+    else:
+        values = tensor[bytemasks]
     bitmasks_packed = pack_bitmasks(bytemasks)
-
     return values, bitmasks_packed, row_offsets
 
 

compressed_tensors/quantization/lifecycle/apply.py

@@ -18,7 +18,7 @@ from collections import OrderedDict, defaultdict
 from copy import deepcopy
 from typing import Dict, Iterable, List, Optional
 from typing import OrderedDict as OrderedDictType
-from typing import Union
+from typing import Set, Union
 
 import torch
 from compressed_tensors.config import CompressionFormat
@@ -52,6 +52,8 @@ __all__ = [
     "apply_quantization_config",
     "apply_quantization_status",
     "find_name_or_class_matches",
+    "expand_sparse_target_names",
+    "is_sparse_target",
 ]
 
 from compressed_tensors.quantization.utils.helpers import is_module_quantized
@@ -245,6 +247,49 @@ def apply_quantization_status(model: Module, status: QuantizationStatus):
         model.apply(compress_quantized_weights)
 
 
+def expand_sparse_target_names(
+    model: Module, targets: Iterable[str], ignore: Iterable[str]
+) -> Set[str]:
+    """
+    Finds all unique module names in the model that match the given
+    targets and ignore lists.
+
+    Note: Targets must be regexes, layer types, or full layer names.
+
+    :param model: model to search for targets in
+    :param targets: list of targets to search for
+    :param ignore: list of targets to ignore
+    :return: set of all targets that match the given targets and should
+        not be ignored
+    """
+    return {
+        name
+        for name, module in iter_named_leaf_modules(model)
+        if is_sparse_target(name, module, targets, ignore)
+    }
+
+
+def is_sparse_target(
+    name: str, module: Module, targets: Iterable[str], ignore: Iterable[str]
+) -> bool:
+    """
+    Determines if a module should be included in the targets based on the
+    targets and ignore lists.
+
+    Note: Targets must be regexes, layer types, or full layer names.
+
+    :param name: name of the module
+    :param module: the module itself
+    :param targets: list of targets to search for
+    :param ignore: list of targets to ignore
+    :return: True if the module is a target and not ignored, False otherwise
+    """
+    return bool(
+        find_name_or_class_matches(name, module, targets)
+        and not find_name_or_class_matches(name, module, ignore or [])
+    )
+
+
 def find_name_or_class_matches(
     name: str, module: Module, targets: Iterable[str], check_contains: bool = False
 ) -> List[str]:

compressed_tensors/quantization/lifecycle/forward.py

@@ -82,8 +82,8 @@ def quantize(
 def dequantize(
     x_q: torch.Tensor,
     scale: torch.Tensor,
-    zero_point: torch.Tensor = None,
-    args: QuantizationArgs = None,
+    zero_point: Optional[torch.Tensor] = None,
+    args: Optional[QuantizationArgs] = None,
     dtype: Optional[torch.dtype] = None,
     g_idx: Optional[torch.Tensor] = None,
 ) -> torch.Tensor:

compressed_tensors/utils/safetensors_load.py

@@ -16,7 +16,7 @@ import json
 import os
 import re
 import struct
-from typing import Dict, List, Optional
+from typing import Dict, List, Optional, Tuple, Union
 
 from safetensors import safe_open
 from torch import Tensor
@@ -30,10 +30,14 @@ __all__ = [
     "merge_names",
     "get_weight_mappings",
     "get_nested_weight_mappings",
+    "get_nested_mappings_from_state_dict",
     "get_quantization_state_dict",
     "is_quantization_param",
 ]
 
+WeightMappingType = Dict[str, str]
+NestedWeightMappingType = Dict[str, WeightMappingType]
+
 
 def get_safetensors_folder(
     pretrained_model_name_or_path: str, cache_dir: Optional[str] = None
@@ -92,7 +96,7 @@ def get_safetensors_header(safetensors_path: str) -> Dict[str, str]:
     return header
 
 
-def match_param_name(full_name: str, param_name: str) -> str:
+def match_param_name(full_name: str, param_name: str) -> Optional[str]:
     """
     Helper function extracting the uncompressed parameterized layer name from a
     compressed name. Assumes the compressed name was merged using merge_names.
@@ -176,38 +180,100 @@ def get_weight_mappings(path_to_model_or_tensors: str) -> Dict[str, str]:
 
 
 def get_nested_weight_mappings(
-    model_path: str, params_to_nest: List[str]
-) -> Dict[str, Dict[str, str]]:
+    model_path: str, params_to_nest: List[str], return_unmatched_params: bool = False
+) -> Union[NestedWeightMappingType, Tuple[NestedWeightMappingType, WeightMappingType]]:
     """
     Takes a path to a state dict saved in safetensors format and returns a nested
-    mapping from uncompressed parameterized layer names to the file locations of each
-    of the layers compression parameters.
+    mapping from uncompressed parameterized layer names to the file locations of
+    each layer's compression parameters.
 
-    layer.weight: {
+    Example of the nested mapping:
+    layer: {
         bitmask: file_location,
         row_offsets: file_location,
         shape: file_location,
         compressed: file_location
     }
 
-    This generalizes to cases where the model is split into multiple safetensors files
+    If other parameters are found that do not match the nested parameters, they will
+    be returned in a separate dictionary only if return_unmatched_params is True.
+    This dictionary may be needed for cases where compressors are stacked (e.g.,
+    quantization compression followed by sparse compression).
+
+    Example of the unmatched params mapping:
+    {
+        layer.weight_scale: file_location,
+        layer.input_scale: file_location
+    }
 
-    :param model_path: path to safetensors state dict, must contain either a single
-    safetensors file or multiple files with an index
-    :return: nested mapping of parameterized layer name to file location
+    This generalizes to cases where the model is split into multiple safetensors
+    files.
+
+    :param model_path: Path to the safetensors state dict, must contain either a
+        single safetensors file or multiple files with an index.
+    :param params_to_nest: List of parameter names to nest.
+    :param return_unmatched_params: If True, return a second dictionary containing
+        the remaining parameters that were not matched to the params_to_nest.
+    :return:
+        - If return_unmatched_params is False:
+            NestedWeightMappingType: A nested mapping of parameterized layer names to
+            file locations of each layer's compression parameters.
+        - If return_unmatched_params is True:
+            Tuple[NestedWeightMappingType, WeightMappingType]: A tuple containing:
+                - NestedWeightMappingType: A nested mapping of parameterized layer
+                names to file locations of each layer's compression parameters.
+                - WeightMappingType: A mapping of the remaining parameter names to
+                their file locations that were not matched to the params_to_nest.
     """
     weight_mappings = get_weight_mappings(model_path)
-
     nested_weight_mappings = {}
-    for key in weight_mappings.keys():
+    unmatched_params = {}
+
+    for key, file_location in weight_mappings.items():
+        matched = False
         for param_name in params_to_nest:
-            maybe_match = match_param_name(key, param_name)
-            if maybe_match is not None:
-                dense_param = maybe_match
+            dense_param = match_param_name(key, param_name)
+            if dense_param:
                 if dense_param not in nested_weight_mappings:
                     nested_weight_mappings[dense_param] = {}
-                nested_weight_mappings[dense_param][param_name] = weight_mappings[key]
+                nested_weight_mappings[dense_param][param_name] = file_location
+                matched = True
+        if return_unmatched_params and not matched:
+            unmatched_params[key] = file_location
+
+    if return_unmatched_params:
+        return nested_weight_mappings, unmatched_params
+    return nested_weight_mappings
 
+
+def get_nested_mappings_from_state_dict(
+    state_dict, params_to_nest
+) -> NestedWeightMappingType:
+    """
+    Takes a state dict and returns a nested mapping from uncompressed
+    parameterized layer names to the value of
+    each layer's compression parameters.
+
+    Example of the nested mapping:
+    layer: {
+        weight_scale: ...,
+        weight: ...,
+        zero_point: ...,
+    }
+
+    :param state_dict: state dict of the model
+    :param params_to_nest: List of parameter names to nest.
+    :return: Nested mapping of parameterized layer names to the value of
+        each layer's compression parameters.
+    """
+    nested_weight_mappings = {}
+    for key in state_dict.keys():
+        for param_name in params_to_nest:
+            dense_param = match_param_name(key, param_name)
+            if dense_param:
+                if dense_param not in nested_weight_mappings:
+                    nested_weight_mappings[dense_param] = {}
+                nested_weight_mappings[dense_param][param_name] = state_dict[key]
     return nested_weight_mappings
 
 

{compressed_tensors_nightly-0.8.1.20250105.dist-info → compressed_tensors_nightly-0.8.1.20250107.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.8.1.20250105
+Version: 0.8.1.20250107
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors_nightly-0.8.1.20250105.dist-info → compressed_tensors_nightly-0.8.1.20250107.dist-info}/RECORD

@@ -5,15 +5,15 @@ compressed_tensors/compressors/__init__.py,sha256=smSygTSfcfuujRrAXDc6uZm4L_ccV1
 compressed_tensors/compressors/base.py,sha256=D9TNwQcjanDiAHODPbg8JUqc66e3j50rctY7A708NEs,6743
 compressed_tensors/compressors/helpers.py,sha256=OK6qxX9j3bHwF9JfIYSGMgBJe2PWjlTA3byXKCJaTIQ,5431
 compressed_tensors/compressors/model_compressors/__init__.py,sha256=5RGGPFu4YqEt_aOdFSQYFYFDjcZFJN0CsMqRtDZz3Js,666
-compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=WmjhfGma7gswMHXLaRriyDNrefO5lCmi6rW35dCcLJM,15903
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=nsMKqjdzEttvkabpp_7Qt4mhWcmjwRYnwjQzeN2a2E4,18295
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=09UJq68Pht6Bf-4iP9xYl3tetKsncNPHD8IAGbePsr4,714
-compressed_tensors/compressors/quantized_compressors/base.py,sha256=K1KOnS6Y8nUA1-HN7VhyfsDc01nilW0WfXMUhuD-l8w,5954
-compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=MMUya3Iwarm0BkeYXqKTUnEDPiBw98GKF09QiNST45k,4960
-compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=1CLwvBlu4AtGkuo3IisD1-rQzwLiA6hE1bCc-pF_XGo,7758
+compressed_tensors/compressors/quantized_compressors/base.py,sha256=LVqSSqSjGi8LB-X13zC_0AFHc8BobGQVC0zjInDhOWE,7217
+compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=fahmPJFz49rVS7q705uQwZ0kUtdP46GuXR7nPr6uIqI,4943
+compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=OO5dceCfNVuY8A23kBg6z2wk-zGUVqR_MyLvObvT7pk,7741
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=i2TESH27l7KXeOhJ6hShIoI904XX96l-cRQiMR6MAaU,704
-compressed_tensors/compressors/sparse_compressors/base.py,sha256=Ua4rUSGyucEs-YJI5z3oIUF-zqQLrFsQ9f-qKasEdUM,4410
+compressed_tensors/compressors/sparse_compressors/base.py,sha256=9e841MQWr0j8m33ejDw_jP5_BIpQ5099x9_pvuZ-Nr0,5944
 compressed_tensors/compressors/sparse_compressors/dense.py,sha256=lSKNWRx6H7aUqaJj1j4qbXk8Gkm1UohbnvW1Rvq6Ra4,1284
-compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=4fKwCG7ZM8mUtSnjPvubzEHl-mTnxMzwjmcs7L43WLY,6622
+compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=Z9qMJ2JyUaBNQe-CXBJLuWacnHdFArrJYZEZDmW5x8o,6889
 compressed_tensors/compressors/sparse_quantized_compressors/__init__.py,sha256=4f_cwcKXB1nVVMoiKgTFAc8jAPjPLElo-Df_EDm1_xw,675
 compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py,sha256=BMIQWTLlnUvxy14iEJegtiP75WHJeOVojey9mKOK1hE,9427
 compressed_tensors/config/__init__.py,sha256=ZBqWn3r6ku1qfmlHHYp0mQueY0i7Pwhr9rbQk9dDlMc,704
@@ -27,9 +27,9 @@ compressed_tensors/quantization/quant_args.py,sha256=jwC__lSmuiJ2qSJYYZGgWgQNbZu
 compressed_tensors/quantization/quant_config.py,sha256=vx06wBo91p4LCb3Vzd-2eCTUeIf_Sz2ZXRP263eQyjQ,10385
 compressed_tensors/quantization/quant_scheme.py,sha256=eQ0JrRZ80GX69fpwW87VzPzzhajhk4mUaJScjk82OY4,6010
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
-compressed_tensors/quantization/lifecycle/apply.py,sha256=jCUSgeOBtagE5IhgIbyYMZ4kv8Rm20VGJ4IxXZ5HAnw,15066
+compressed_tensors/quantization/lifecycle/apply.py,sha256=XS4M6N1opKBybhkuQsS338QVb_CKMhUM5TUKrqoNQ0k,16517
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
-compressed_tensors/quantization/lifecycle/forward.py,sha256=QPL6-vKOFuKdKIEsVqMhsw4x552Jpm2sqO0oeChbnrM,12941
+compressed_tensors/quantization/lifecycle/forward.py,sha256=DOWouUqfaLA4Qhg-ojVVBdhhSAlgZqFC26vZARxE0ko,12961
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
 compressed_tensors/quantization/lifecycle/initialize.py,sha256=hymYtayTSumm8KCYAYPY267aWmlsJpt8oQFiRblk8qE,7452
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
@@ -41,10 +41,10 @@ compressed_tensors/utils/helpers.py,sha256=XF36-SLkXnAHh0VzbvUlAdh6a88aCQvS_WeYs
 compressed_tensors/utils/offload.py,sha256=cMmzd9IdlNbs29CReHj1PPSLUM6OWaT5YumlLT5eP3w,13845
 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=m08ANVuTBxQdoa6LufDgcNJ7wCLDJolyZljB8VEybAU,8578
+compressed_tensors/utils/safetensors_load.py,sha256=fBuoHVPoBt1mkvqFJ60zQIASX_4nhl0-6QfFS27NY8I,11430
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors_nightly-0.8.1.20250105.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.8.1.20250105.dist-info/METADATA,sha256=MqxpH595abOI0cwQmfVodd1oY1p4TvUD89SXaWkqVAg,6799
-compressed_tensors_nightly-0.8.1.20250105.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-compressed_tensors_nightly-0.8.1.20250105.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.8.1.20250105.dist-info/RECORD,,
+compressed_tensors_nightly-0.8.1.20250107.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.8.1.20250107.dist-info/METADATA,sha256=X847Gn6LonNqO0XBA0k2lHz_BGmYt7oeHJ8-k90LVsA,6799
+compressed_tensors_nightly-0.8.1.20250107.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+compressed_tensors_nightly-0.8.1.20250107.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.8.1.20250107.dist-info/RECORD,,