compressed-tensors-nightly 0.6.0.20240920__tar.gz → 0.6.0.20240921__tar.gz

{compressed-tensors-nightly-0.6.0.20240920/src/compressed_tensors_nightly.egg-info → compressed-tensors-nightly-0.6.0.20240921}/PKG-INFO

@@ -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 → compressed-tensors-nightly-0.6.0.20240921}/src/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-nightly-0.6.0.20240920 → compressed-tensors-nightly-0.6.0.20240921}/src/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 → compressed-tensors-nightly-0.6.0.20240921/src/compressed_tensors_nightly.egg-info}/PKG-INFO

@@ -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.