compressed-tensors-nightly 0.8.1.20250110__py3-none-any.whl → 0.8.1.20250112__py3-none-any.whl

compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -310,7 +310,10 @@ class ModelCompressor:
         model_path = get_safetensors_folder(model_path)
         sparse_decompressed = False
 
-        if self.sparsity_compressor is not None:
+        if (
+            self.sparsity_compressor is not None
+            and self.sparsity_config.format != CompressionFormat.dense.value
+        ):
             # Sparse decompression is applied on the model_path
             dense_gen = self.sparsity_compressor.decompress(model_path)
             self._replace_weights(dense_gen, model)

compressed_tensors/compressors/sparse_compressors/__init__.py

@@ -15,4 +15,5 @@
 
 from .base import *
 from .dense import *
+from .sparse_24_bitmask import *
 from .sparse_bitmask import *

compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py

@@ -0,0 +1,238 @@
+# 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 dataclasses import dataclass
+from typing import Dict, List, Tuple, Union
+
+import torch
+from compressed_tensors.compressors.base import BaseCompressor
+from compressed_tensors.compressors.sparse_compressors.base import BaseSparseCompressor
+from compressed_tensors.config import CompressionFormat, SparsityStructure
+from compressed_tensors.quantization import FP8_DTYPE
+from compressed_tensors.utils import merge_names, pack_bitmasks, unpack_bitmasks
+from torch import Tensor
+
+
+__all__ = [
+    "Sparse24BitMaskCompressor",
+    "Sparse24BitMaskTensor",
+    "sparse24_bitmask_compress",
+    "sparse24_bitmask_decompress",
+    "get_24_bytemasks",
+]
+
+
+@BaseCompressor.register(name=CompressionFormat.sparse_24_bitmask.value)
+class Sparse24BitMaskCompressor(BaseSparseCompressor):
+    """
+    Compression for sparse models using bitmasks. Non-zero weights are stored in a 2d
+    values tensor, with their locations stored in a 2d bitmask
+    """
+
+    COMPRESSION_PARAM_NAMES = [
+        "shape",
+        "compressed",
+        "bitmask",
+    ]
+
+    def compress_weight(self, name, value):
+        bitmask_tensor = Sparse24BitMaskTensor.from_dense(
+            value, self.config.sparsity_structure
+        )
+        bitmask_dict = bitmask_tensor.dict(name_prefix=name, device="cpu")
+        return bitmask_dict
+
+    def decompress_weight(self, weight_data):
+        data = Sparse24BitMaskTensor.from_compressed_data(**weight_data)
+        decompressed = data.decompress()
+        return decompressed
+
+
+@dataclass
+class Sparse24BitMaskTensor:
+    """
+    Owns compressions and decompression for a single 2:4 sparse
+    bitmask compressed tensor.
+
+    :param shape: shape of dense tensor
+    :param compressed: 2d tensor of non-zero values
+    :param bitmask: 2d bitmask of non-zero values
+    """
+
+    shape: List[int]
+    compressed: Tensor
+    bitmask: Tensor
+
+    @staticmethod
+    def from_dense(
+        tensor: Tensor,
+        sparsity_structure: Union[SparsityStructure, str] = SparsityStructure.TWO_FOUR,
+    ) -> "Sparse24BitMaskTensor":
+        """
+        :param tensor: dense tensor to compress
+        :return: instantiated compressed tensor
+        """
+        shape = list(tensor.shape)
+        compressed, bitmask = sparse24_bitmask_compress(
+            tensor.cpu(), sparsity_structure=sparsity_structure
+        )
+        return Sparse24BitMaskTensor(
+            shape=shape,
+            compressed=compressed,
+            bitmask=bitmask,
+        )
+
+    @staticmethod
+    def from_compressed_data(
+        shape: Union[List[int], Tensor], compressed: Tensor, bitmask: Tensor
+    ) -> "Sparse24BitMaskTensor":
+        """
+        :param shape: shape of the dense tensor (can be a list or a tensor)
+        :param compressed: 2d tensor of non-zero values
+        :param bitmask: 2d bitmask of non-zero values
+        :return: instantiated Sparse24BitMaskTensor
+        """
+        if isinstance(shape, Tensor):
+            shape = shape.tolist()
+        return Sparse24BitMaskTensor(
+            shape=shape, compressed=compressed, bitmask=bitmask
+        )
+
+    def decompress(self) -> Tensor:
+        """
+        :return: reconstructed dense tensor
+        """
+        return sparse24_bitmask_decompress(self.compressed, self.bitmask, self.shape)
+
+    def curr_memory_size_bytes(self) -> int:
+        """
+        :return: size in bytes required to store compressed tensor on disk
+        """
+
+        def sizeof_tensor(a: Tensor) -> int:
+            return a.element_size() * a.nelement()
+
+        return sizeof_tensor(self.compressed) + sizeof_tensor(self.bitmask)
+
+    def dict(self, name_prefix: str, device: str = "cpu") -> Dict[str, Tensor]:
+        """
+        :param name_prefix: name of original tensor to store compressed weight as
+        :return: dict of compressed data for the stored weight
+        """
+        if name_prefix.endswith(".weight"):
+            name_prefix = name_prefix[: -len(".weight")]
+        return {
+            merge_names(name_prefix, "shape"): torch.tensor(
+                self.shape, device=device
+            ).reshape(-1, 1),
+            merge_names(name_prefix, "compressed"): self.compressed.to(device),
+            merge_names(name_prefix, "bitmask"): self.bitmask.to(device),
+        }
+
+    def __repr__(self) -> str:
+        return f"BitMaskTensor(shape={self.shape}, compressed=True)"
+
+
+def sparse24_bitmask_compress(
+    tensor: Tensor,
+    sparsity_structure: Union[SparsityStructure, str] = SparsityStructure.TWO_FOUR,
+) -> Tuple[Tensor, Tensor, Tensor]:
+    """
+    Compresses a dense tensor using bitmask compression
+
+    :param tensor: dense 2D tensor to compress
+    :param sparsity_structure: structure of sparsity in the tensor, defaults
+        to unstructured, can also be set to `2:4`
+    :return: tuple of compressed data representing tensor
+    """
+    assert len(tensor.shape) == 2, "Only 2D tensors are supported"
+    assert (
+        SparsityStructure(sparsity_structure) == SparsityStructure.TWO_FOUR
+    ), "Only 2:4 sparsity is supported"
+
+    bytemasks = get_24_bytemasks(tensor=tensor)
+
+    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]
+
+    num_rows, num_cols = tensor.shape
+    compressed_values = values.reshape(num_rows, num_cols // 2)
+    bitmasks_packed = pack_bitmasks(bytemasks)
+    return compressed_values, bitmasks_packed
+
+
+def sparse24_bitmask_decompress(
+    values: Tensor, bitmasks: Tensor, original_shape: torch.Size
+) -> Tensor:
+    """
+    Reconstructs a dense tensor from a compressed one
+
+    :param values: 1d tensor of non-zero values
+    :param bitmasks: 2d int8 tensor flagging locations of non-zero values in the
+    tensors original shape
+    :param original_shape: shape of the dense tensor
+    :return: decompressed dense tensor
+    """
+    bytemasks_unpacked = unpack_bitmasks(bitmasks, original_shape)
+
+    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
+    return decompressed_tensor
+
+
+def get_24_bytemasks(tensor):
+    """
+    Generate a 2:4 sparsity mask for the given tensor.
+
+    This function creates a mask where exactly 2 out of every 4 elements are
+    preserved based on their magnitudes. The preserved elements are the ones
+    with the highest absolute values in each group of 4 elements.
+
+    :param tensor: The input tensor for which the 2:4 sparsity mask is to be created.
+                   The tensor can be of any shape but its total number of elements
+                   must be a multiple of 4.
+    :return: A boolean tensor of the same shape as the input tensor, where `True`
+             indicates the preserved elements and `False` indicates the pruned elements.
+    :raises ValueError: If the total number of elements in the tensor is not a
+                        multiple of 4.
+    """
+    original_dtype = tensor.dtype
+    if tensor.dtype == FP8_DTYPE:
+        tensor = tensor.view(torch.int8)
+    original_shape = tensor.shape
+    num_elements = tensor.numel()
+
+    if num_elements % 4 != 0:
+        raise ValueError("Tensor size must be a multiple of 4 for TWO_FOUR sparsity")
+
+    reshaped_tensor = tensor.view(-1, 4)
+    abs_tensor = reshaped_tensor.abs()
+    topk_indices = abs_tensor.topk(2, dim=1).indices
+    mask = torch.zeros_like(reshaped_tensor, dtype=torch.bool)
+    mask.scatter_(1, topk_indices, True)
+    mask = mask.view(original_shape)
+    tensor = tensor.view(original_dtype)
+
+    return mask

compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py

@@ -14,13 +14,12 @@
 
 from typing import Dict, List, Tuple, Union
 
-import numpy
 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 compressed_tensors.utils import merge_names, pack_bitmasks, unpack_bitmasks
 from torch import Tensor
 
 
@@ -29,8 +28,6 @@ __all__ = [
     "BitmaskTensor",
     "bitmask_compress",
     "bitmask_decompress",
-    "pack_bitmasks",
-    "unpack_bitmasks",
 ]
 
 
@@ -164,37 +161,3 @@ def bitmask_decompress(
     decompressed_tensor[bytemasks_unpacked] = values
 
     return decompressed_tensor
-
-
-def pack_bitmasks(bytemasks: Tensor) -> Tensor:
-    """
-    Converts a bytemask tensor to a bitmask tensor to reduce memory. Shape RxC will be
-    compressed to R x ceil(C/8)
-    :param bytemasks: mask tensor where each byte corresponds to a weight
-    :return: mask tensor where each bit corresounds to a weight
-    """
-    packed_bits_numpy = numpy.packbits(bytemasks.numpy(), axis=-1, bitorder="little")
-    packed_bits_torch = torch.from_numpy(packed_bits_numpy)
-
-    return packed_bits_torch
-
-
-def unpack_bitmasks(packed_bitmasks: Tensor, original_shape: torch.Size) -> Tensor:
-    """
-    Converts a bitmask tensor back to a bytemask tensor for use during decompression
-
-    :param packed_bitmasks: mask tensor where each bit corresponds to a weight
-    :param original_shape: dense shape to decompress to
-    :return: boolean mask of weights in the original dense shape
-    """
-    # Unpack the bits
-    unpacked_bits = numpy.unpackbits(
-        packed_bitmasks.numpy(), axis=-1, count=original_shape[-1], bitorder="little"
-    )
-
-    # Reshape to match the original shape
-    unpacked_bitmasks_torch = torch.from_numpy(
-        unpacked_bits.reshape(original_shape).astype(bool)
-    )
-
-    return unpacked_bitmasks_torch

compressed_tensors/config/__init__.py

@@ -15,4 +15,5 @@
 # flake8: noqa
 from .base import *
 from .dense import *
+from .sparse_24_bitmask import *
 from .sparse_bitmask import *

compressed_tensors/config/base.py

@@ -26,6 +26,7 @@ __all__ = ["SparsityCompressionConfig", "CompressionFormat", "SparsityStructure"
 class CompressionFormat(Enum):
     dense = "dense"
     sparse_bitmask = "sparse-bitmask"
+    sparse_24_bitmask = "sparse-24-bitmask"
     int_quantized = "int-quantized"
     float_quantized = "float-quantized"
     naive_quantized = "naive-quantized"

compressed_tensors/config/sparse_24_bitmask.py

@@ -0,0 +1,40 @@
+# 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
+
+from compressed_tensors.config import (
+    CompressionFormat,
+    SparsityCompressionConfig,
+    SparsityStructure,
+)
+
+
+__all__ = ["Sparse24BitMaskConfig"]
+
+
+@SparsityCompressionConfig.register(name=CompressionFormat.sparse_24_bitmask.value)
+class Sparse24BitMaskConfig(SparsityCompressionConfig):
+    """
+    Configuration for storing a 24 sparse model using
+    bytemask compression
+
+    :param global_sparsity: average sparsity of the entire model
+    :param sparsity_structure: structure of the sparsity, should always be
+        "2:4" for this compression format
+    """
+
+    format: str = CompressionFormat.sparse_24_bitmask.value
+    global_sparsity: Optional[float] = 0.0
+    sparsity_structure: Optional[str] = SparsityStructure.TWO_FOUR.value

compressed_tensors/utils/helpers.py

@@ -14,8 +14,9 @@
 
 import warnings
 from functools import wraps
-from typing import Any, Callable, Dict, Optional
+from typing import Any, Callable, Dict, List, Optional
 
+import numpy
 import torch
 from transformers import AutoConfig
 
@@ -29,6 +30,10 @@ __all__ = [
     "getattr_chain",
     "deprecated",
     "Aliasable",
+    "combine_shards",
+    "shard_tensor",
+    "pack_bitmasks",
+    "unpack_bitmasks",
 ]
 
 FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
@@ -214,3 +219,108 @@ class Aliasable:
     def __hash__(self):
         canonical_value = self.aliases.get(self.value, self.value)
         return hash(canonical_value)
+
+
+def shard_tensor(
+    tensor: torch.Tensor, shard_sizes: List[int], dim: int = 0
+) -> List[torch.Tensor]:
+    """
+    Shards a tensor into a list of tensors along a given dimension.
+
+    raises: ValueError: If the sum of shard_sizes does not match the
+        size of the tensor along the given dimension.
+
+    :param tensor: The input tensor to shard.
+    :param shard_sizes : List of sizes for each shard along the specified dimension.
+    :param dim : The dimension along which to shard the tensor.
+    :returns: A list of tensors sharded along the specified dimension.
+    """
+    if sum(shard_sizes) != tensor.size(dim):
+        raise ValueError(
+            "Sum of shard_sizes must equal the size of the tensor "
+            "along the specified dimension."
+        )
+
+    shards = []
+    start_idx = 0
+
+    for size in shard_sizes:
+        end_idx = start_idx + size
+        shard = tensor.narrow(dim, start_idx, size)
+        shards.append(shard)
+        start_idx = end_idx
+
+    return shards
+
+
+def combine_shards(shards, dim=0):
+    """
+    Combine decompressed shards along a given dimension using `narrow`.
+
+    :param shards: List of decompressed shard tensors.
+    :param dim: Dimension to combine along (default: 0).
+    :return: Combined decompressed tensor.
+    """
+    if not shards:
+        raise ValueError("The list of shards is empty.")
+
+    # Assert that all shards have the same dtype
+    shard_dtypes = {shard.dtype for shard in shards}
+    if len(shard_dtypes) > 1:
+        raise ValueError("All shards must have the same dtype.")
+
+    # Determine the total shape of the combined tensor
+    total_shape = list(shards[0].shape)
+    total_shape[dim] = sum(shard.shape[dim] for shard in shards)
+
+    # Create the combined tensor
+    combined = torch.zeros(total_shape, dtype=shards[0].dtype, device=shards[0].device)
+
+    # Fill the combined tensor using narrow
+    shard_offset = 0
+    for shard in shards:
+        shard_size = shard.shape[dim]
+        combined.narrow(dim, shard_offset, shard_size).copy_(shard)
+        shard_offset += shard_size
+
+    return combined
+
+
+def pack_bitmasks(bytemasks: torch.Tensor) -> torch.Tensor:
+    """
+    Converts a bytemask tensor to a bitmask tensor to reduce memory. Shape RxC will be
+    compressed to R x ceil(C/8)
+
+    :param bytemasks: mask tensor where each byte corresponds to a weight
+    :return: mask tensor where each bit corresounds to a weight
+    """
+    packed_bits_numpy = numpy.packbits(bytemasks.numpy(), axis=-1, bitorder="little")
+    packed_bits_torch = torch.from_numpy(packed_bits_numpy)
+
+    return packed_bits_torch
+
+
+def unpack_bitmasks(
+    packed_bitmasks: torch.Tensor, original_shape: torch.Size
+) -> torch.Tensor:
+    """
+    Converts a bitmask tensor back to a bytemask tensor for use during decompression
+
+    :param packed_bitmasks: mask tensor where each bit corresponds to a weight
+    :param original_shape: dense shape to decompress to
+    :return: boolean mask of weights in the original dense shape
+    """
+    # Unpack the bits
+    unpacked_bits = numpy.unpackbits(
+        packed_bitmasks.cpu().numpy(),
+        axis=-1,
+        count=original_shape[-1],
+        bitorder="little",
+    )
+
+    # Reshape to match the original shape
+    unpacked_bitmasks_torch = torch.from_numpy(
+        unpacked_bits.reshape(original_shape).astype(bool)
+    )
+
+    return unpacked_bitmasks_torch

{compressed_tensors_nightly-0.8.1.20250110.dist-info → compressed_tensors_nightly-0.8.1.20250112.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.8.1.20250110
+Version: 0.8.1.20250112
 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-0.8.1.20250110.dist-info → compressed_tensors_nightly-0.8.1.20250112.dist-info}/RECORD

@@ -5,20 +5,22 @@ 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=nsMKqjdzEttvkabpp_7Qt4mhWcmjwRYnwjQzeN2a2E4,18295
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=3WyzAW2Rm_uLprxwO2QH6FR76W6Mk4r2yedayaSZHhw,18396
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=09UJq68Pht6Bf-4iP9xYl3tetKsncNPHD8IAGbePsr4,714
 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/__init__.py,sha256=Atuz-OdEgn8OCUhx7Ovd6gXdyImAI186uCR-uR0t_Nk,737
 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=Z9qMJ2JyUaBNQe-CXBJLuWacnHdFArrJYZEZDmW5x8o,6889
+compressed_tensors/compressors/sparse_compressors/sparse_24_bitmask.py,sha256=3M0FI8gY_T8iNmp9oSEHoVjr_AwdercdRd3R9hzltVM,8512
+compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=7zSr9bqkpuH1ivQpxtYBNxXIoElal7Jo1nSKpZN_IFk,5633
 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
-compressed_tensors/config/base.py,sha256=3bFAdwDZjOt-U3fneOeL8dRci-PS8DqstnXuQVtkfiQ,3435
+compressed_tensors/config/__init__.py,sha256=8sOoZ6xvYSC79mBvEtO8l6xk4PC80d29AnnJiGMrY2M,737
+compressed_tensors/config/base.py,sha256=R3iUmFf1MslEjin5LgwQbmfJHIsS7Uw0UIxfn780uqY,3479
 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=MJa-UfoKhIkdUWRD1shrXXri2cOwR5GK0a4t4bNYosM,3268
@@ -37,14 +39,14 @@ compressed_tensors/quantization/utils/helpers.py,sha256=DBP-sGRpGAY01K0LFE7qqonN
 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=XF36-SLkXnAHh0VzbvUlAdh6a88aCQvS_WeYs9Lfio8,6827
+compressed_tensors/utils/helpers.py,sha256=OODitCQuSKH6Ux_8Ff05pSrKzaai1t8IERNPKTtiD1A,10321
 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=fBuoHVPoBt1mkvqFJ60zQIASX_4nhl0-6QfFS27NY8I,11430
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors_nightly-0.8.1.20250110.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.8.1.20250110.dist-info/METADATA,sha256=B1dwfh4KtSBBAhnhxvyKqw6hzmzWfnrW4M8RnPtyRyE,6799
-compressed_tensors_nightly-0.8.1.20250110.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
-compressed_tensors_nightly-0.8.1.20250110.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.8.1.20250110.dist-info/RECORD,,
+compressed_tensors_nightly-0.8.1.20250112.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.8.1.20250112.dist-info/METADATA,sha256=pq2PSfcDi6Nd2HW_UQdaHAYGGDo8X4Ko948pS3B1fj0,6799
+compressed_tensors_nightly-0.8.1.20250112.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+compressed_tensors_nightly-0.8.1.20250112.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.8.1.20250112.dist-info/RECORD,,