compressed-tensors 0.9.5a20250513__py3-none-any.whl → 0.9.5a20250514__py3-none-any.whl

compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -47,6 +47,9 @@ from compressed_tensors.quantization.utils import (
     iter_named_leaf_modules,
 )
 from compressed_tensors.utils import (
+    align_module_device,
+    delete_offload_parameter,
+    get_execution_device,
     get_safetensors_folder,
     has_offloaded_params,
     merge_names,
@@ -98,6 +101,9 @@ class ModelCompressor:
     :param quantization_config: config specifying quantization compression parameters
     """
 
+    sparsity_config: Optional[SparsityCompressionConfig] = None
+    quantization_config: Optional[QuantizationConfig] = None
+
     @classmethod
     def from_pretrained(
         cls,
@@ -261,6 +267,8 @@ class ModelCompressor:
                 quantization_config.format, config=quantization_config
             )
 
+    # ----- used by hf quantizer ----- #
+
     def get_missing_module_keys(self, model: Module) -> List[str]:
         """
         Identifies the expected missing weight keys in the compressed state_dict.
@@ -270,7 +278,6 @@ class ModelCompressor:
         This function determines which weight keys are missing based on the
         applied compression techniques.
 
-
         :param model: The PyTorch model to check for missing keys.
         :return: A list of missing keys expected in the compressed state_dict.
         """
@@ -362,8 +369,124 @@ class ModelCompressor:
 
         return list(unexpected_keys)
 
+    # ----- model memory compression/decompression pathways ----- #
+
+    def compress_model(self, model: Module):
+        """
+        Compress a model in memory. Because the model structure is modified in place,
+        this method is more memory-efficient than `self.compress`
+
+        :param model: model containing parameters to compress
+        """
+        module_to_scheme = map_module_to_scheme(model)
+        sparse_compression_targets: Set[str] = expand_target_names(
+            model=model,
+            targets=self.sparsity_config.targets if self.sparsity_config else [],
+            ignore=self.sparsity_config.ignore if self.sparsity_config else [],
+        )
+
+        for prefix, module in tqdm(model.named_modules(), desc="Compressing model"):
+            if prefix in module_to_scheme or prefix in sparse_compression_targets:
+                # in the future, support compression on same device
+                with align_module_device(module, execution_device="cpu"):
+                    state_dict = module.state_dict(prefix=f"{prefix}.")
+
+                # quantization first
+                if prefix in module_to_scheme:
+                    state_dict = self.quantization_compressor.compress(
+                        state_dict,
+                        names_to_scheme=module_to_scheme,
+                        show_progress=False,
+                    )
+
+                # sparsity second
+                if prefix in sparse_compression_targets:
+                    state_dict = self.sparsity_compressor.compress(
+                        state_dict,
+                        compression_targets=sparse_compression_targets,
+                        show_progress=False,
+                    )
+
+                # remove any existing parameters
+                device = get_execution_device(module)
+                for name, _ in list(module.named_parameters()):
+                    delattr(module, name)
+
+                # replace with compressed parameters
+                for name, value in state_dict.items():
+                    name = name.removeprefix(f"{prefix}.")
+                    value = value.to(device)
+                    param = torch.nn.Parameter(value, requires_grad=False)
+                    register_offload_parameter(module, name, param)
+
+                module.quantization_status = QuantizationStatus.COMPRESSED
+
+    def decompress_model(self, model: Module):
+        """
+        Decompress a model in memory. Because the model structure is modified in place,
+        this method does not require loading some compression parameters from disk
+
+        :param model: model containing parameters to compress
+        """
+        module_to_scheme = map_module_to_scheme(model)
+        sparse_compression_targets: Set[str] = expand_target_names(
+            model=model,
+            targets=self.sparsity_config.targets if self.sparsity_config else [],
+            ignore=self.sparsity_config.ignore if self.sparsity_config else [],
+        )
+
+        for prefix, module in tqdm(model.named_modules(), desc="Decompressing model"):
+            if prefix in module_to_scheme or prefix in sparse_compression_targets:
+                # in the future, support decompression on same device
+                with align_module_device(module, execution_device="cpu"):
+                    state_dict = module.state_dict(prefix=f"{prefix}.")
+
+                # sparsity first
+                if prefix in sparse_compression_targets:
+                    # sparse_compression_targets are automatically inferred by this fn
+                    generator = self.sparsity_compressor.decompress_from_state_dict(
+                        state_dict,
+                    )
+                    # generates (param_path, param_val)
+                    # of compressed and unused params
+                    state_dict = {key: value for key, value in generator}
+
+                # quantization second
+                if prefix in module_to_scheme:
+                    generator = self.quantization_compressor.decompress_from_state_dict(
+                        state_dict,
+                        names_to_scheme=module_to_scheme,
+                    )
+                    # generates (mod_path, {param_name, param_val})
+                    # of compressed params and used params, but not unused params
+                    # some used params are removed by get_unexpected_file_keys
+                    state_dict = {
+                        merge_names(module_path, param_name): param_value
+                        for module_path, compressed_data in generator
+                        for param_name, param_value in compressed_data.items()
+                    }
+
+                # remove any existing parameters
+                device = get_execution_device(module)
+                for name, _ in list(module.named_parameters()):
+                    delete_offload_parameter(module, name)
+
+                # replace with decompressed parameters
+                for name, value in state_dict.items():
+                    name = name.removeprefix(f"{prefix}.")
+                    value = value.to(device)
+                    param = torch.nn.Parameter(value, requires_grad=False)
+                    register_offload_parameter(module, name, param)
+
+                module.quantization_status = QuantizationStatus.FROZEN
+
+    # ----- state dict compression pathways ----- #
+
     def compress(
-        self, model: Module, state_dict: Optional[Dict[str, Tensor]] = None
+        self,
+        model: Module,
+        state_dict: Optional[Dict[str, Tensor]] = None,
+        show_progress: bool = False,
     ) -> Dict[str, Tensor]:
         """
         Compresses a dense state dict or model with sparsity and/or quantization
@@ -379,7 +502,9 @@ class ModelCompressor:
         if self.quantization_compressor is not None:
             module_to_scheme = map_module_to_scheme(model)
             state_dict = self.quantization_compressor.compress(
-                state_dict, names_to_scheme=module_to_scheme
+                state_dict,
+                names_to_scheme=module_to_scheme,
+                show_progress=show_progress,
             )
 
             # TODO: consider sparse compression to also be compression
@@ -397,6 +522,7 @@ class ModelCompressor:
             state_dict = self.sparsity_compressor.compress(
                 state_dict,
                 compression_targets=sparse_compression_targets,
+                show_progress=show_progress,
             )
 
         # HACK: Override the dtype_byte_size function in transformers to
@@ -406,6 +532,8 @@ class ModelCompressor:
 
         return state_dict
 
+    # ----- disk decompression pathways ----- #
+
     def decompress(self, model_path: str, model: Module):
         """
         Overwrites the weights in model with weights decompressed from model_path

compressed_tensors/compressors/quantized_compressors/base.py

@@ -23,7 +23,6 @@ from compressed_tensors.utils import (
     get_nested_mappings_from_state_dict,
     get_nested_weight_mappings,
     merge_names,
-    remove_suffix,
 )
 from safetensors import safe_open
 from torch import Tensor
@@ -71,6 +70,7 @@ class BaseQuantizationCompressor(BaseCompressor):
         self,
         model_state: Dict[str, Tensor],
         names_to_scheme: Dict[str, QuantizationScheme],
+        show_progress: bool = False,
         **kwargs,
     ) -> Dict[str, Tensor]:
         """
@@ -79,18 +79,21 @@ class BaseQuantizationCompressor(BaseCompressor):
         :param model_state: state dict of uncompressed model
         :param names_to_scheme: quantization args for each quantized weight, needed for
             quantize function to calculate bit depth
+        :param show_progress: whether to show tqdm progress
         :return: compressed state dict
         """
+        uncompressed_names = list(model_state.keys())
         compressed_dict = {}
         save_device = "cpu"
 
-        uncompressed_names = list(model_state.keys())
-        for name in tqdm(uncompressed_names, desc="Compressing with quantization"):
+        # compress values
+        desc = "Compressing with quantization"
+        for name in tqdm(uncompressed_names, desc=desc, disable=(not show_progress)):
             value = model_state[name]
 
             # compress weights
             if name.endswith("weight"):
-                prefix = remove_suffix(name, "weight")
+                prefix = name.removesuffix("weight")
 
                 # gather qparams
                 scale = model_state.get(prefix + "weight_scale", None)
@@ -182,7 +185,7 @@ class BaseQuantizationCompressor(BaseCompressor):
             )
 
         else:
-            yield from self._decompress_from_state_dict(
+            yield from self.decompress_from_state_dict(
                 path_to_model_or_tensors, names_to_scheme
             )
 
@@ -209,7 +212,11 @@ class BaseQuantizationCompressor(BaseCompressor):
                 weight_data["weight"] = decompressed
                 yield module_path, weight_data
 
-    def _decompress_from_state_dict(self, state_dict, names_to_scheme):
+    def decompress_from_state_dict(
+        self,
+        state_dict: Dict[str, torch.Tensor],
+        names_to_scheme: Dict[str, QuantizationScheme],
+    ) -> Generator[Tuple[str, Dict[str, torch.Tensor]], None, None]:
         weight_mappings = get_nested_mappings_from_state_dict(
             state_dict, self.compression_param_names
         )
@@ -219,7 +226,7 @@ class BaseQuantizationCompressor(BaseCompressor):
                 weight_data[param_name] = param_value
 
             if "weight_scale" in weight_data:
-                quant_args = names_to_scheme[module_path]
+                quant_args = names_to_scheme[module_path].weights
                 decompressed = self.decompress_weight(
                     compressed_data=weight_data, quantization_args=quant_args
                 )

compressed_tensors/compressors/sparse_compressors/base.py

@@ -16,7 +16,11 @@ import logging
 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
+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
@@ -63,6 +67,7 @@ class BaseSparseCompressor(BaseCompressor):
         self,
         model_state: Dict[str, Tensor],
         compression_targets: Optional[Set[str]] = None,
+        show_progress: bool = False,
     ) -> Dict[str, Tensor]:
         """
         Compresses a dense state dict using bitmask compression
@@ -76,7 +81,11 @@ class BaseSparseCompressor(BaseCompressor):
         _LOGGER.debug(
             f"Compressing model with {len(model_state)} parameterized layers..."
         )
-        for name, value in tqdm(model_state.items(), desc="Compressing model"):
+        for name, value in tqdm(
+            model_state.items(),
+            desc="Compressing with sparsity",
+            disable=(not show_progress),
+        ):
             if not self.should_compress(name, compression_targets):
                 compressed_dict[name] = value
                 continue
@@ -124,15 +133,15 @@ class BaseSparseCompressor(BaseCompressor):
             self.compression_param_names,
             return_unmatched_params=True,
         )
-        for weight_name in weight_mappings.keys():
+        for module_path in weight_mappings.keys():
             weight_data = {}
-            for param_name, safe_path in weight_mappings[weight_name].items():
-                full_name = merge_names(weight_name, param_name)
+            for param_name, safe_path in weight_mappings[module_path].items():
+                full_name = merge_names(module_path, param_name)
                 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 merge_names(weight_name, "weight"), decompressed
+            yield merge_names(module_path, "weight"), decompressed
 
         for ignored_param_name, safe_path in ignored_params.items():
             should_skip = False
@@ -146,6 +155,35 @@ class BaseSparseCompressor(BaseCompressor):
                     value = f.get_tensor(ignored_param_name)
                 yield ignored_param_name, value
 
+    def decompress_from_state_dict(
+        self,
+        state_dict: Dict[str, Tensor],
+    ) -> Generator[Tuple[str, Dict[str, Tensor]], None, None]:
+        """
+        Decompress the state dict of a module (or model)
+
+        Unlike `self.decompress`, this function does not need to explicitly skip params
+        via params_to_skip_load because it is more convenient for its only caller
+        (ModelCompressor.decompress_model) to retrieve all unused param keys
+
+        :param state_dict: state dict containing parameters to decompress
+        :return: Generator of (param_path, param_val)
+        """
+        weight_mappings, ignored_params = get_nested_mappings_from_state_dict(
+            state_dict, self.compression_param_names, return_unmatched_params=True
+        )
+
+        for module_path in weight_mappings.keys():
+            weight_data = {}
+            for param_name, param_value in weight_mappings[module_path].items():
+                weight_data[param_name] = param_value
+
+            decompressed = self.decompress_weight(weight_data)
+            yield merge_names(module_path, "weight"), decompressed
+
+        for ignored_param_path, ignored_param_value in ignored_params.items():
+            yield ignored_param_path, ignored_param_value
+
     @staticmethod
     def should_compress(name: str, expanded_targets: Optional[Set[str]] = None) -> bool:
         """

compressed_tensors/compressors/sparse_compressors/dense.py

@@ -40,3 +40,10 @@ class DenseCompressor(BaseCompressor):
         self, path_to_model_or_tensors: str, device: str = "cpu", **kwargs
     ) -> Generator[Tuple[str, Tensor], None, None]:
         return iter([])
+
+    def decompress_from_state_dict(
+        self,
+        state_dict: Dict[str, Tensor],
+    ) -> Generator[Tuple[str, Dict[str, Tensor]], None, None]:
+        for key, value in state_dict.items():
+            yield key, value

compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from dataclasses import dataclass
-from typing import Dict, List, Tuple, Union
+from typing import Dict, Generator, List, Tuple, Union
 
 import torch
 from compressed_tensors.compressors.base import BaseCompressor
@@ -202,11 +202,7 @@ def sparse24_bitmask_decompress(
     decompressed_tensor = torch.zeros(original_shape, dtype=values.dtype)
     decompressed_tensor = decompressed_tensor.to(values.device)
     values = values.flatten()
-    if decompressed_tensor.dtype == FP8_DTYPE:
-        decompressed_tensor[bytemasks_unpacked] = values
-        decompressed_tensor = decompressed_tensor.cuda()
-    else:
-        decompressed_tensor[bytemasks_unpacked] = values
+    decompressed_tensor[bytemasks_unpacked] = values
     return decompressed_tensor
 
 

compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py

@@ -125,6 +125,7 @@ class Marlin24Compressor(BaseCompressor):
         self,
         model_state: Dict[str, Tensor],
         names_to_scheme: Dict[str, QuantizationScheme],
+        show_progress: bool = False,
         **kwargs,
     ) -> Dict[str, Tensor]:
         """
@@ -134,6 +135,7 @@ class Marlin24Compressor(BaseCompressor):
         :param model_state: state dict of uncompressed model
         :param names_to_scheme: quantization scheme for each quantized weight, needed
             for quantize function to calculate bit depth
+        :param show_progress: whether to show tqdm progress
         :return: compressed state dict
         """
         self.validate_quant_compatability(names_to_scheme)
@@ -144,7 +146,9 @@ class Marlin24Compressor(BaseCompressor):
             f"Compressing model with {len(model_state)} parameterized layers..."
         )
 
-        for name, value in tqdm(model_state.items(), desc="Compressing model"):
+        for name, value in tqdm(
+            model_state.items(), desc="Compressing model", disable=(not show_progress)
+        ):
             if name.endswith(weight_suffix):
                 prefix = name[: -(len(weight_suffix))]
                 scale = model_state.get(merge_names(prefix, "weight_scale"), None)

compressed_tensors/linear/compressed_linear.py

@@ -23,6 +23,7 @@ from compressed_tensors.quantization import (
     initialize_module_for_quantization,
 )
 from compressed_tensors.utils import register_offload_parameter
+from compressed_tensors.utils.offload import get_execution_device
 from torch import Tensor
 from torch.nn import Parameter
 from torch.nn.functional import linear
@@ -60,7 +61,7 @@ class CompressedLinear(Linear):
         """
         module.__class__ = CompressedLinear
         module.compressor = BaseCompressor.load_from_registry(quantization_format)
-        device = next(module.parameters()).device
+        init_device = get_execution_device(module)
 
         # this will initialize all the scales and zero points
         initialize_module_for_quantization(
@@ -79,7 +80,7 @@ class CompressedLinear(Linear):
         # populate compressed weights and quantization parameters
         for name, (shape, dtype) in compression_params.items():
             param = Parameter(
-                torch.empty(shape, device=device, dtype=dtype), requires_grad=False
+                torch.empty(shape, device=init_device, dtype=dtype), requires_grad=False
             )
             register_offload_parameter(module, name, param)
 

compressed_tensors/utils/helpers.py

@@ -38,7 +38,6 @@ __all__ = [
     "shard_tensor",
     "pack_bitmasks",
     "unpack_bitmasks",
-    "remove_suffix",
 ]
 
 FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
@@ -329,9 +328,3 @@ def unpack_bitmasks(
     )
 
     return unpacked_bitmasks_torch
-
-
-def remove_suffix(value: str, suffix: str) -> str:
-    # can replace with str.removesuffix in python3.9+
-    assert value.endswith(suffix)
-    return value[: -len(suffix)]

compressed_tensors/utils/safetensors_load.py

@@ -35,6 +35,7 @@ __all__ = [
     "is_quantization_param",
 ]
 
+NestedStateDictType = Dict[str, Dict[str, Tensor]]
 WeightMappingType = Dict[str, str]
 NestedWeightMappingType = Dict[str, WeightMappingType]
 
@@ -249,8 +250,10 @@ def get_nested_weight_mappings(
 
 
 def get_nested_mappings_from_state_dict(
-    state_dict, params_to_nest: Iterable[str]
-) -> NestedWeightMappingType:
+    state_dict: Dict[str, Tensor],
+    params_to_nest: Iterable[str],
+    return_unmatched_params: bool = False,
+) -> Union[NestedStateDictType, Tuple[NestedStateDictType, Dict[str, Tensor]]]:
     """
     Takes a state dict and returns a nested mapping from uncompressed
     parameterized layer names to the value of
@@ -266,16 +269,26 @@ def get_nested_mappings_from_state_dict(
     :param state_dict: state dict of the model
     :param params_to_nest: Iterable of parameter names to nest.
     :return: Nested mapping of parameterized layer names to the value of
-        each layer's compression parameters.
+        each layer's compression parameters. If `return_unmatched_params`, then
+        also return a dictionary mapping unused parameter names to their values
     """
     nested_weight_mappings = {}
+    unmatched_params = {}
+
     for key in state_dict.keys():
+        matched = False
         for param_name in params_to_nest:
             module_path = match_param_name(key, param_name)
             if module_path:
                 if module_path not in nested_weight_mappings:
                     nested_weight_mappings[module_path] = {}
                 nested_weight_mappings[module_path][param_name] = state_dict[key]
+                matched = True
+        if return_unmatched_params and not matched:
+            unmatched_params[key] = state_dict[key]
+
+    if return_unmatched_params:
+        return nested_weight_mappings, unmatched_params
     return nested_weight_mappings
 
 

compressed_tensors/version.py

@@ -17,5 +17,5 @@ __version__: str
 __version_tuple__: VERSION_TUPLE
 version_tuple: VERSION_TUPLE
 
-__version__ = version = '0.9.5.a20250513'
+__version__ = version = '0.9.5.a20250514'
 __version_tuple__ = version_tuple = (0, 9, 5)

{compressed_tensors-0.9.5a20250513.dist-info → compressed_tensors-0.9.5a20250514.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compressed-tensors
-Version: 0.9.5a20250513
+Version: 0.9.5a20250514
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed_tensors-0.9.5a20250513.dist-info → compressed_tensors-0.9.5a20250514.dist-info}/RECORD

@@ -1,30 +1,30 @@
 compressed_tensors/__init__.py,sha256=UtKmifNeBCSE2TZSAfduVNNzHY-3V7bLjZ7n7RuXLOE,812
 compressed_tensors/base.py,sha256=73HYH7HY7O2roC89yG_piPFnZwrBfn_i7HmKl90SKc0,875
-compressed_tensors/version.py,sha256=anvjogcrhtUkFpir0cHcNUxTI9qeqiV_ELg-b1mW8Xc,521
+compressed_tensors/version.py,sha256=DipxWeTHSaI4zbOq4qATqt0oCJgWFClVEMZaHMo3xZE,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
 compressed_tensors/compressors/model_compressors/__init__.py,sha256=5RGGPFu4YqEt_aOdFSQYFYFDjcZFJN0CsMqRtDZz3Js,666
-compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=uh3Rbyqhjvt8o8On6ioOn6utBKv2siRRmAvgM1lDrxU,26555
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=L-xfTooii1ubjMJeNQp2QL9_Q_T2Z6L1oqYdBYotDAc,32393
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=KvaFBL_Q84LxRGJOV035M8OBoCkAx8kOkfphswgkKWk,745
-compressed_tensors/compressors/quantized_compressors/base.py,sha256=n0L2QH2_Y1vWtLeQ0uV78y2lV4bviFEAtUKODl8L_nw,8828
+compressed_tensors/compressors/quantized_compressors/base.py,sha256=4YWT95GIhHETI7glsk_ITrnUzzN1MhEypt-0z9eKqOI,9134
 compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=fd0KlkSx6bvZ3xwIkK3jEUdPSUPs56Eua4dEDOtzKW0,5150
 compressed_tensors/compressors/quantized_compressors/nvfp4_quantized.py,sha256=Gw-lVzk5jrKUlM5UTCiJBmhM5gHzB9mn8r298MVUbDI,6395
 compressed_tensors/compressors/quantized_compressors/pack_quantized.py,sha256=SPIHlk8ewip2LcjgkCw02K21EkfUSFSd9qQqL0Pt5eM,11162
 compressed_tensors/compressors/sparse_compressors/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
-compressed_tensors/compressors/sparse_compressors/base.py,sha256=PMiWIaW2XSF_esYJlQ12RVW7opeAzavdbkRFtelMFX0,6655
-compressed_tensors/compressors/sparse_compressors/dense.py,sha256=_uW_HISeDNz4yboSZWoh6GwrkUE6HFibzPQSKrHOCkg,1505
-compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py,sha256=mEKSSgpXookqYSJw3mlyP6cYYKD-eaIvpQMvi4JO6TY,8807
+compressed_tensors/compressors/sparse_compressors/base.py,sha256=YNZWcHjDleAlqbgRZQ6oJf44MQb_UDNvJGOqhl26uFA,8098
+compressed_tensors/compressors/sparse_compressors/dense.py,sha256=rPaxbP7P52prWNs4lGaiBbpNvsQLElFMwOrq1oBP2Yg,1733
+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
-compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py,sha256=7VRLmtUTg1iJl3mXiOzLPi1RgIOhMISPAwzVi8v2QF0,9951
+compressed_tensors/compressors/sparse_quantized_compressors/marlin_24.py,sha256=7F9J6wgkecitK5hHuqjetZ18HExHIF4QIw1wgm2Y6U8,10099
 compressed_tensors/config/__init__.py,sha256=8sOoZ6xvYSC79mBvEtO8l6xk4PC80d29AnnJiGMrY2M,737
 compressed_tensors/config/base.py,sha256=p3glQHvC2fjodf_SvlelVrTWSIjGXgGC86t8oVOlMng,3529
 compressed_tensors/config/dense.py,sha256=NgSxnFCnckU9-iunxEaqiFwqgdO7YYxlWKR74jNbjks,1317
 compressed_tensors/config/sparse_24_bitmask.py,sha256=Lhj39zT2V1hxftprvxvneyhv45ShlXOKd75DBbDTyTE,1401
 compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5ynVAUeiiYpS1Gt8,1308
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
-compressed_tensors/linear/compressed_linear.py,sha256=_m6XpNcI53eeSHO8VdiuAM6UBTdpDhn5Ivd8iRMwEKc,3980
+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=CepGBAURFGxBzTyFXxHwsUs6wYEJ46_jPbEvJYMG0Tw,10491
 compressed_tensors/quantization/quant_config.py,sha256=MxSUcb5dOqMN6LFyD5K2h8X0TvEtcWIAoiUJqD2dHGE,10159
@@ -40,14 +40,14 @@ compressed_tensors/quantization/utils/helpers.py,sha256=-wX0H7zVysJ67jRRCGbx6Bfx
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
 compressed_tensors/registry/registry.py,sha256=vRcjVB1ITfSbfYUaGndBBmqhip_5vsS62weorVg0iXo,11896
 compressed_tensors/utils/__init__.py,sha256=gS4gSU2pwcAbsKj-6YMaqhm25udFy6ISYaWBf-myRSM,808
-compressed_tensors/utils/helpers.py,sha256=Le3LWskSQRr7pw8fWy5qmfDKYlKiQFy0id83uVi4fUg,10610
+compressed_tensors/utils/helpers.py,sha256=RrNvzD08naEjEiXdU-FdZjQVda1nQywu1hA_GCDj0vg,10415
 compressed_tensors/utils/offload.py,sha256=JNQ66_6vhSsizhlUaMgyEdBuFolYxbgUuT1mAZrCfKY,15436
 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=kkkUDmS1H40MFy6FDP-DFGiAYbtqke6bKE7YrAtORtA,11499
+compressed_tensors/utils/safetensors_load.py,sha256=DMfZBuUbA6qp_BG_zIWT3ckiEE33K9ob34s-OgzReO4,12057
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors-0.9.5a20250513.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors-0.9.5a20250513.dist-info/METADATA,sha256=gMyJF3nQLJu7DCEbM0LDyD2GeuKwu6zxewO6qST7_s8,7004
-compressed_tensors-0.9.5a20250513.dist-info/WHEEL,sha256=DnLRTWE75wApRYVsjgc6wsVswC54sMSJhAEd4xhDpBk,91
-compressed_tensors-0.9.5a20250513.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors-0.9.5a20250513.dist-info/RECORD,,
+compressed_tensors-0.9.5a20250514.dist-info/licenses/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors-0.9.5a20250514.dist-info/METADATA,sha256=wEGdTDwjdDlrcGnYaYXPdatEZj_dykpwIcdAwDxqEtk,7004
+compressed_tensors-0.9.5a20250514.dist-info/WHEEL,sha256=QZxptf4Y1BKFRCEDxD4h2V0mBFQOVFLFEpvxHmIs52A,91
+compressed_tensors-0.9.5a20250514.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors-0.9.5a20250514.dist-info/RECORD,,

{compressed_tensors-0.9.5a20250513.dist-info → compressed_tensors-0.9.5a20250514.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (80.4.0)
+Generator: setuptools (80.6.0)
 Root-Is-Purelib: true
 Tag: py3-none-any