compressed-tensors-nightly 0.7.1.20241023__py3-none-any.whl → 0.7.1.20241025__py3-none-any.whl

compressed_tensors/config/base.py

@@ -12,16 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from enum import Enum
+from enum import Enum, unique
 from typing import List, Optional
 
 from compressed_tensors.registry import RegistryMixin
 from pydantic import BaseModel
 
 
-__all__ = ["SparsityCompressionConfig", "CompressionFormat"]
+__all__ = ["SparsityCompressionConfig", "CompressionFormat", "SparsityStructure"]
 
 
+@unique
 class CompressionFormat(Enum):
     dense = "dense"
     sparse_bitmask = "sparse-bitmask"
@@ -32,6 +33,63 @@ class CompressionFormat(Enum):
     marlin_24 = "marlin-24"
 
 
+@unique
+class SparsityStructure(Enum):
+    """
+    An enumeration to represent different sparsity structures.
+
+    Attributes
+    ----------
+    TWO_FOUR : str
+        Represents a 2:4 sparsity structure.
+    ZERO_ZERO : str
+        Represents a 0:0 sparsity structure.
+    UNSTRUCTURED : str
+        Represents an unstructured sparsity structure.
+
+    Examples
+    --------
+    >>> SparsityStructure('2:4')
+    <SparsityStructure.TWO_FOUR: '2:4'>
+
+    >>> SparsityStructure('unstructured')
+    <SparsityStructure.UNSTRUCTURED: 'unstructured'>
+
+    >>> SparsityStructure('2:4') == SparsityStructure.TWO_FOUR
+    True
+
+    >>> SparsityStructure('UNSTRUCTURED') == SparsityStructure.UNSTRUCTURED
+    True
+
+    >>> SparsityStructure(None) == SparsityStructure.UNSTRUCTURED
+    True
+
+    >>> SparsityStructure('invalid')
+    Traceback (most recent call last):
+        ...
+    ValueError: invalid is not a valid SparsityStructure
+    """
+
+    TWO_FOUR = "2:4"
+    UNSTRUCTURED = "unstructured"
+    ZERO_ZERO = "0:0"
+
+    def __new__(cls, value):
+        obj = object.__new__(cls)
+        obj._value_ = value.lower() if value is not None else value
+        return obj
+
+    @classmethod
+    def _missing_(cls, value):
+        # Handle None and case-insensitive values
+        if value is None:
+            return cls.UNSTRUCTURED
+        for member in cls:
+            if member.value == value.lower():
+                return member
+        raise ValueError(f"{value} is not a valid {cls.__name__}")
+
+
 class SparsityCompressionConfig(RegistryMixin, BaseModel):
     """
     Base data class for storing sparsity compression parameters

compressed_tensors/quantization/lifecycle/calibration.py

@@ -64,14 +64,12 @@ def set_module_for_calibration(module: Module, quantize_weights_upfront: bool =
                 quantization_args=module.quantization_scheme.weights,
             )
 
-        observer = module.weight_observer
-
-        g_idx = getattr(module, "weight_g_idx", None)
-
         offloaded = is_module_offloaded(module)
         if offloaded:
             module._hf_hook.pre_forward(module)
 
+        observer = module.weight_observer
+        g_idx = getattr(module, "weight_g_idx", None)
         scale, zero_point = observer(module.weight, g_idx=g_idx)
         update_parameter_data(module, scale, "weight_scale")
         update_parameter_data(module, zero_point, "weight_zero_point")

compressed_tensors/quantization/lifecycle/frozen.py

@@ -14,7 +14,6 @@
 
 
 from compressed_tensors.quantization.quant_config import QuantizationStatus
-from compressed_tensors.quantization.utils import is_kv_cache_quant_scheme
 from torch.nn import Module
 
 

compressed_tensors/quantization/observers/helpers.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from collections import Counter
-from typing import Optional, Tuple
+from typing import Tuple
 
 import torch
 from compressed_tensors.quantization.quant_args import (

compressed_tensors/quantization/observers/mse.py

@@ -70,7 +70,9 @@ class MovingAverageMSEObserver(Observer):
             absolute_min_val = torch.amin(observed, dim=reduce_dims, keepdims=True)
             absolute_max_val = torch.amax(observed, dim=reduce_dims, keepdims=True)
 
-        best = torch.full_like(absolute_min_val, torch.finfo(absolute_min_val.dtype).max)
+        best = torch.full_like(
+            absolute_min_val, torch.finfo(absolute_min_val.dtype).max
+        )
         min_val = torch.ones_like(absolute_min_val)
         max_val = torch.zeros_like(absolute_max_val)
         for i in range(int(self.maxshrink * self.grid)):

{compressed_tensors_nightly-0.7.1.20241023.dist-info → compressed_tensors_nightly-0.7.1.20241025.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.7.1.20241023
+Version: 0.7.1.20241025
 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-0.7.1.20241023.dist-info → compressed_tensors_nightly-0.7.1.20241025.dist-info}/RECORD

@@ -17,7 +17,7 @@ compressed_tensors/compressors/sparse_compressors/sparse_bitmask.py,sha256=4fKwC
 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=akqE7eW8CLTslpWRxERaZ8R0TSm1lS7D1bgZXKL0xi8,9427
 compressed_tensors/config/__init__.py,sha256=ZBqWn3r6ku1qfmlHHYp0mQueY0i7Pwhr9rbQk9dDlMc,704
-compressed_tensors/config/base.py,sha256=BNTFKy12isY7qblwxdi_R1f00EzgrNOXLrfxqLCPT8w,1903
+compressed_tensors/config/base.py,sha256=3bFAdwDZjOt-U3fneOeL8dRci-PS8DqstnXuQVtkfiQ,3435
 compressed_tensors/config/dense.py,sha256=NgSxnFCnckU9-iunxEaqiFwqgdO7YYxlWKR74jNbjks,1317
 compressed_tensors/config/sparse_bitmask.py,sha256=pZUboRNZTu6NajGOQEFExoPknak5ynVAUeiiYpS1Gt8,1308
 compressed_tensors/linear/__init__.py,sha256=fH6rjBYAxuwrTzBTlTjTgCYNyh6TCvCqajCz4Im4YrA,617
@@ -29,17 +29,17 @@ compressed_tensors/quantization/quant_config.py,sha256=NCiMvUMnnz5kTyAkDylxjtEGQ
 compressed_tensors/quantization/quant_scheme.py,sha256=5ggPz5sqEfTUgvJJeiPIINA74QtO-08hb3szsm7UHGE,6000
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=MXE2E7GfIfRRfhrdGy2Og3AZOz5N59B0ZGFcsD89y6c,821
 compressed_tensors/quantization/lifecycle/apply.py,sha256=czaayvpeUYyWRJhO_klffw6esptOgA9sBKL5TWQcRdw,15805
-compressed_tensors/quantization/lifecycle/calibration.py,sha256=gPSD3kiH4VuU6nq-OLbOmhBGaMXsebEwLm4PkEnUhf0,3043
+compressed_tensors/quantization/lifecycle/calibration.py,sha256=fJ2RDL3E4hmWR8v8nYhq_tv31K8WV00o_4Y3xr7c37Y,3041
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=8GjOnx4rwOZZqSDTdnejNOY2DVTjNDzH0DfY_rQam6k,16575
-compressed_tensors/quantization/lifecycle/frozen.py,sha256=8myzxsz5h5Odh5cIB2lDHb7xLRYBYnAhA1PO8YGuCtM,1839
+compressed_tensors/quantization/lifecycle/frozen.py,sha256=71TsgS0Uxku0NomdWOBJsVfXCGTne-Gx9zUEMsCmw5Q,1764
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
 compressed_tensors/quantization/lifecycle/initialize.py,sha256=lKoFy18PjbSklyum7f4hoLuWtHShBKax7JDTBzPlCqM,8839
 compressed_tensors/quantization/observers/__init__.py,sha256=DYrttzq-8MHLZUzpX-xzzm4hrw6HcXkMkux82KBKb1M,738
 compressed_tensors/quantization/observers/base.py,sha256=5ovQicWPYHjIxr6-EkQ4lgOX0PpI9g23iSzKpxjM1Zg,8420
-compressed_tensors/quantization/observers/helpers.py,sha256=o9hg4E9b5cCb5PaEAj6jHiUWkNrKtYtv0b1pGg-T9B4,5516
+compressed_tensors/quantization/observers/helpers.py,sha256=nUFdNEIACiPBfFwNYDGCXOvw6tf7j6jfTvDwImHKMPg,5506
 compressed_tensors/quantization/observers/min_max.py,sha256=sQXqU3z-voxIDfR_9mQzwQUflZj2sASm_G8CYaXntFw,3865
-compressed_tensors/quantization/observers/mse.py,sha256=9JRbvXo0VKLrgsTNuVlQ7AV87wwjRUuQludG0v7IJbI,6058
+compressed_tensors/quantization/observers/mse.py,sha256=G5Y9v4MqXUVcKxBSmCFFW3p_7rlu-6scqLIN88ng-sE,6080
 compressed_tensors/quantization/utils/__init__.py,sha256=VdtEmP0bvuND_IGQnyqUPc5lnFp-1_yD7StKSX4x80w,656
 compressed_tensors/quantization/utils/helpers.py,sha256=y4LEyC2oUd876ZMdALWKGH3Ct5EgBJZV4id_NUjTGH8,9531
 compressed_tensors/registry/__init__.py,sha256=FwLSNYqfIrb5JD_6OK_MT4_svvKTN_nEhpgQlQvGbjI,658
@@ -51,8 +51,8 @@ compressed_tensors/utils/permutations_24.py,sha256=kx6fsfDHebx94zsSzhXGyCyuC9sVy
 compressed_tensors/utils/permute.py,sha256=V6tJLKo3Syccj-viv4F7ZKZgJeCB-hl-dK8RKI_kBwI,2355
 compressed_tensors/utils/safetensors_load.py,sha256=m08ANVuTBxQdoa6LufDgcNJ7wCLDJolyZljB8VEybAU,8578
 compressed_tensors/utils/semi_structured_conversions.py,sha256=XKNffPum54kPASgqKzgKvyeqWPAkair2XEQXjkp7ho8,13489
-compressed_tensors_nightly-0.7.1.20241023.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.7.1.20241023.dist-info/METADATA,sha256=VTd017y76OKUEeK4DGjn_qbgympazs5s85NMVClrTUM,6799
-compressed_tensors_nightly-0.7.1.20241023.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-compressed_tensors_nightly-0.7.1.20241023.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.7.1.20241023.dist-info/RECORD,,
+compressed_tensors_nightly-0.7.1.20241025.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.7.1.20241025.dist-info/METADATA,sha256=c5Ij0aprxvjscvNlBWuoUIUkNE0Fd2sakpFBRaoZI2A,6799
+compressed_tensors_nightly-0.7.1.20241025.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+compressed_tensors_nightly-0.7.1.20241025.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.7.1.20241025.dist-info/RECORD,,