compressed-tensors-nightly 0.6.0.20240920__py3-none-any.whl → 0.6.0.20240921__py3-none-any.whl

compressed_tensors/config/base.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 
 from enum import Enum
-from typing import Optional
+from typing import List, Optional
 
 from compressed_tensors.registry import RegistryMixin
 from pydantic import BaseModel
@@ -37,11 +37,16 @@ class SparsityCompressionConfig(RegistryMixin, BaseModel):
     Base data class for storing sparsity compression parameters
 
     :param format: name of compression format
+    :param targets: List of layer names or layer types that aren't sparse and should
+        be ignored during compression. By default, assume all layers are targeted
+    :param ignore: List of layer names (unique) to ignore from targets. Defaults to None
     :param global_sparsity: average sparsity of the entire model
     :param sparsity_structure: structure of the sparsity, such as
     "unstructured", "2:4", "8:16" etc
     """
 
     format: str
+    targets: Optional[List[str]] = None
+    ignore: Optional[List[str]] = None
     global_sparsity: Optional[float] = 0.0
     sparsity_structure: Optional[str] = "unstructured"

compressed_tensors/quantization/quant_scheme.py

@@ -57,15 +57,9 @@ class QuantizationScheme(BaseModel):
             # default to quantizing all Linear layers
             targets = ["Linear"]
 
-        # default to 8 bit integer symmetric quantization
-        # for weights
-        weights = QuantizationArgs(num_bits=8, symmetric=True)
-
-        # default to 8 bit integer asymmetric quantization
-        input_activations = QuantizationArgs(num_bits=8, symmetric=True)
-
-        # Do not quantize the output activations
-        # by default
+        # by default, activations and weights are left unquantized
+        weights = None
+        input_activations = None
         output_activations = None
 
         return cls(
@@ -110,6 +104,7 @@ def is_preset_scheme(name: str) -> bool:
     """
     return name.upper() in PRESET_SCHEMES
 
+
 UNQUANTIZED = dict()
 
 # 8 bit integer weights and 8 bit activations quantization

{compressed_tensors_nightly-0.6.0.20240920.dist-info → compressed_tensors_nightly-0.6.0.20240921.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.6.0.20240920
+Version: 0.6.0.20240921
 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-0.6.0.20240920.dist-info → compressed_tensors_nightly-0.6.0.20240921.dist-info}/RECORD

@@ -11,7 +11,7 @@ compressed_tensors/compressors/naive_quantized.py,sha256=z3h3ca5xKCN69mahutxcbzd
 compressed_tensors/compressors/pack_quantized.py,sha256=27RVmJ2wg2dvCoawj407HSmKT3VPGJ6ujAMHlT26WlI,7571
 compressed_tensors/compressors/sparse_bitmask.py,sha256=kiDwBlFV0sJGLcIdDYxIiuF64ccgwDfqq1hWRQThYDc,8647
 compressed_tensors/config/__init__.py,sha256=ZBqWn3r6ku1qfmlHHYp0mQueY0i7Pwhr9rbQk9dDlMc,704
-compressed_tensors/config/base.py,sha256=caSZ7xZ_kgcHRMXZ5hM1i6TKbgY__CkiSjZ93imHZQ0,1562
+compressed_tensors/config/base.py,sha256=BNTFKy12isY7qblwxdi_R1f00EzgrNOXLrfxqLCPT8w,1903
 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
@@ -19,7 +19,7 @@ compressed_tensors/linear/compressed_linear.py,sha256=G0gEFfxLAUsgRcnfSV-PKz1ZBN
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=CmyVtjJeHlqCW-7R5Z7tIw6lXUrzCX6Y9bwgmMxEudY,8069
 compressed_tensors/quantization/quant_config.py,sha256=NpVu8YJ4Xw2pIQW_PGaNaml8kx1bUnxkvb0jBYWbKdE,9971
-compressed_tensors/quantization/quant_scheme.py,sha256=VRvWweqwlhjYMrKf62fXKQTeoJGhjJa3tXnE-TuFdFA,6093
+compressed_tensors/quantization/quant_scheme.py,sha256=HmR1DcFZcjuqX7KHUYI0NFXsCIzJ8sxFGH6zhYuHmEs,5870
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=MXE2E7GfIfRRfhrdGy2Og3AZOz5N59B0ZGFcsD89y6c,821
 compressed_tensors/quantization/lifecycle/apply.py,sha256=uftWFunr_CpCZM_qWfo2O1USXKB2qSYD1pBJsO8BuCU,15285
 compressed_tensors/quantization/lifecycle/calibration.py,sha256=PlS_EqCOPqJD3QKuLPXO9AOtDzXtQWvEBTynFv-FFVw,2698
@@ -45,8 +45,8 @@ 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-0.6.0.20240920.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.6.0.20240920.dist-info/METADATA,sha256=ojNYypBv_NmLJohyxTs9r0shK9LKJnism4or6-BXI1Q,6799
-compressed_tensors_nightly-0.6.0.20240920.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-compressed_tensors_nightly-0.6.0.20240920.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.6.0.20240920.dist-info/RECORD,,
+compressed_tensors_nightly-0.6.0.20240921.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.6.0.20240921.dist-info/METADATA,sha256=VvF40t86UhlxctGKFqedC0jd-Fkef3ahcnb5ZgjJD44,6799
+compressed_tensors_nightly-0.6.0.20240921.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+compressed_tensors_nightly-0.6.0.20240921.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.6.0.20240921.dist-info/RECORD,,