compressed-tensors-nightly 0.8.0.20241122__py3-none-any.whl → 0.8.0.20241124__py3-none-any.whl

compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -24,7 +24,6 @@ import compressed_tensors
 import torch
 import transformers
 from compressed_tensors.base import (
-    COMPRESSION_CONFIG_NAME,
     COMPRESSION_VERSION_NAME,
     QUANTIZATION_CONFIG_NAME,
     QUANTIZATION_METHOD_NAME,
@@ -39,6 +38,7 @@ from compressed_tensors.quantization import (
     apply_quantization_config,
     load_pretrained_quantization,
 )
+from compressed_tensors.quantization.quant_args import QuantizationArgs
 from compressed_tensors.quantization.utils import (
     is_module_quantized,
     iter_named_leaf_modules,
@@ -103,12 +103,14 @@ class ModelCompressor:
         :return: compressor for the configs, or None if model is not compressed
         """
         config = AutoConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
-        compression_config = getattr(config, COMPRESSION_CONFIG_NAME, None)
+        compression_config = getattr(config, QUANTIZATION_CONFIG_NAME, None)
+
         return cls.from_compression_config(compression_config)
 
     @classmethod
     def from_compression_config(
-        cls, compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"]
+        cls,
+        compression_config: Union[Dict[str, Any], "CompressedTensorsConfig"],
     ):
         """
         :param compression_config:
@@ -265,7 +267,11 @@ class ModelCompressor:
             state_dict = model.state_dict()
 
         compressed_state_dict = state_dict
-        quantized_modules_to_args = map_modules_to_quant_args(model)
+
+        quantized_modules_to_args: Dict[
+            str, QuantizationArgs
+        ] = map_modules_to_quant_args(model)
+
         if self.quantization_compressor is not None:
             compressed_state_dict = self.quantization_compressor.compress(
                 state_dict, names_to_scheme=quantized_modules_to_args
@@ -369,7 +375,13 @@ class ModelCompressor:
             update_parameter_data(module, data, param_name)
 
 
-def map_modules_to_quant_args(model: Module) -> Dict:
+def map_modules_to_quant_args(model: Module) -> Dict[str, QuantizationArgs]:
+    """
+    Given a pytorch model, map out the submodule name (usually linear layers)
+     to the QuantizationArgs
+
+    :param model: pytorch model
+    """
     quantized_modules_to_args = {}
     for name, submodule in iter_named_leaf_modules(model):
         if is_module_quantized(submodule):

compressed_tensors/linear/compressed_linear.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from typing import Dict, Tuple
+
 import torch
 from compressed_tensors.compressors.base import BaseCompressor
 from compressed_tensors.quantization import (
@@ -53,7 +55,7 @@ class CompressedLinear(Linear):
         )
 
         # get the shape and dtype of compressed parameters
-        compression_params = module.compressor.compression_param_info(
+        compression_params: Dict[str, Tuple] = module.compressor.compression_param_info(
             module.weight.shape, quantization_scheme.weights
         )
 

compressed_tensors/quantization/lifecycle/apply.py

@@ -106,7 +106,8 @@ def apply_quantization_config(
     model: Module, config: Union[QuantizationConfig, None], run_compressed: bool = False
 ) -> OrderedDict:
     """
-    Initializes the model for quantization in-place based on the given config
+    Initializes the model for quantization in-place based on the given config.
+    Optionally coverts quantizable modules to compressed_linear modules
 
     :param model: model to apply quantization config to
     :param config: quantization config

compressed_tensors/quantization/quant_config.py

@@ -132,9 +132,9 @@ class QuantizationConfig(BaseModel):
         `k_proj` and `v_proj` in their names. If this is not the case
         and kv_cache_scheme != None, the quantization of kv cache will fail
     :global_compression_ratio: optional informational config to report the model
-    compression ratio acheived by the quantization config
+        compression ratio acheived by the quantization config
     :ignore: optional list of layers to ignore from config_groups. Layers in this list
-    are not quantized even if they match up with a target in config_groups
+        are not quantized even if they match up with a target in config_groups
     """
 
     config_groups: Dict[str, Union[QuantizationScheme, List[str]]]

compressed_tensors/quantization/quant_scheme.py

@@ -36,7 +36,7 @@ class QuantizationScheme(BaseModel):
     of modules should be quantized
 
     :param targets: list of modules to apply the QuantizationArgs to, can be layer
-    names, layer types or a regular expression
+    names, layer types or a regular expression, typically ["Linear"]
     :param weights: quantization config for layer weights
     :param input_activations: quantization config for layer inputs
     :param output_activations: quantization config for layer outputs
@@ -47,28 +47,6 @@ class QuantizationScheme(BaseModel):
     input_activations: Optional[QuantizationArgs] = None
     output_activations: Optional[QuantizationArgs] = None
 
-    @classmethod
-    def default_scheme(
-        cls,
-        targets: Optional[List[str]] = None,
-    ):
-
-        if targets is None:
-            # default to quantizing all Linear layers
-            targets = ["Linear"]
-
-        # by default, activations and weights are left unquantized
-        weights = None
-        input_activations = None
-        output_activations = None
-
-        return cls(
-            targets=targets,
-            weights=weights,
-            input_activations=input_activations,
-            output_activations=output_activations,
-        )
-
 
 """
 Pre-Set Quantization Scheme Args

{compressed_tensors_nightly-0.8.0.20241122.dist-info → compressed_tensors_nightly-0.8.0.20241124.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors-nightly
-Version: 0.8.0.20241122
+Version: 0.8.0.20241124
 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-0.8.0.20241122.dist-info → compressed_tensors_nightly-0.8.0.20241124.dist-info}/RECORD

@@ -5,7 +5,7 @@ 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=XJgPsq8KiDfiR4e8bSI38lmoOd2ApqRk1aPcXS2obqY,15600
+compressed_tensors/compressors/model_compressors/model_compressor.py,sha256=sxh1TvW1Bp9YJE41hW0XZfd0kYYB85nhJvBLVRTDcV0,15886
 compressed_tensors/compressors/quantized_compressors/__init__.py,sha256=09UJq68Pht6Bf-4iP9xYl3tetKsncNPHD8IAGbePsr4,714
 compressed_tensors/compressors/quantized_compressors/base.py,sha256=K1KOnS6Y8nUA1-HN7VhyfsDc01nilW0WfXMUhuD-l8w,5954
 compressed_tensors/compressors/quantized_compressors/naive_quantized.py,sha256=Mmfr-hap-4zw7CzE1mXi0UirknqGidNxw38GGWVgTqM,4916
@@ -21,13 +21,13 @@ compressed_tensors/config/base.py,sha256=3bFAdwDZjOt-U3fneOeL8dRci-PS8DqstnXuQVt
 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
-compressed_tensors/linear/compressed_linear.py,sha256=0jTTf6XxOAjAYs3tvFtgiNMAO4W10sSeR-pdH2M413g,3218
+compressed_tensors/linear/compressed_linear.py,sha256=MJa-UfoKhIkdUWRD1shrXXri2cOwR5GK0a4t4bNYosM,3268
 compressed_tensors/quantization/__init__.py,sha256=83J5bPB7PavN2TfCoW7_vEDhfYpm4TDrqYO9vdSQ5bk,760
 compressed_tensors/quantization/quant_args.py,sha256=osjNwCSB6tcyH9Qeg5sHEiB-bHyi3XJ8TzkGVJuGTc4,8711
-compressed_tensors/quantization/quant_config.py,sha256=NCiMvUMnnz5kTyAkDylxjtEGQnjgsIYIeNR2zyHEdTQ,10371
-compressed_tensors/quantization/quant_scheme.py,sha256=5ggPz5sqEfTUgvJJeiPIINA74QtO-08hb3szsm7UHGE,6000
+compressed_tensors/quantization/quant_config.py,sha256=K6kOZ6LDXpFlqsVzR4NEATV6y6Ea83rJWnNyVlvw-pI,10379
+compressed_tensors/quantization/quant_scheme.py,sha256=o3SaeNKWFVqbOtk0cLlwHh_YenU4hnteXFtH5ey9zLk,5452
 compressed_tensors/quantization/lifecycle/__init__.py,sha256=_uItzFWusyV74Zco_pHLOTdE9a83cL-R-ZdyQrBkIyw,772
-compressed_tensors/quantization/lifecycle/apply.py,sha256=pdCqxXnVw7HoDDanaOtek13g8x_nb54CBUlfuMdhFG4,14993
+compressed_tensors/quantization/lifecycle/apply.py,sha256=jCUSgeOBtagE5IhgIbyYMZ4kv8Rm20VGJ4IxXZ5HAnw,15066
 compressed_tensors/quantization/lifecycle/compressed.py,sha256=Fj9n66IN0EWsOAkBHg3O0GlOQpxstqjCcs0ttzMXrJ0,2296
 compressed_tensors/quantization/lifecycle/forward.py,sha256=QPL6-vKOFuKdKIEsVqMhsw4x552Jpm2sqO0oeChbnrM,12941
 compressed_tensors/quantization/lifecycle/helpers.py,sha256=C0mhy2vJ0fCjVeN4kFNhw8Eq1wkteBGHiZ36RVLThRY,944
@@ -43,8 +43,8 @@ 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-0.8.0.20241122.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-compressed_tensors_nightly-0.8.0.20241122.dist-info/METADATA,sha256=4kWPSxMdTxocUah2jo1PG8Y_hO-rCf60bj_sVukBsDw,6799
-compressed_tensors_nightly-0.8.0.20241122.dist-info/WHEEL,sha256=bFJAMchF8aTQGUgMZzHJyDDMPTO3ToJ7x23SLJa1SVo,92
-compressed_tensors_nightly-0.8.0.20241122.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
-compressed_tensors_nightly-0.8.0.20241122.dist-info/RECORD,,
+compressed_tensors_nightly-0.8.0.20241124.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+compressed_tensors_nightly-0.8.0.20241124.dist-info/METADATA,sha256=WfFSOxQOHVkPwaftGCghUA9za3NTcO7HeWRyBPbXGxk,6799
+compressed_tensors_nightly-0.8.0.20241124.dist-info/WHEEL,sha256=tZoeGjtWxWRfdplE7E3d45VPlLNQnvbKiYnx7gwAy8A,92
+compressed_tensors_nightly-0.8.0.20241124.dist-info/top_level.txt,sha256=w2i-GyPs2s1UwVxvutSvN_lM22SXC2hQFBmoMcPnV7Y,19
+compressed_tensors_nightly-0.8.0.20241124.dist-info/RECORD,,

{compressed_tensors_nightly-0.8.0.20241122.dist-info → compressed_tensors_nightly-0.8.0.20241124.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.45.0)
+Generator: bdist_wheel (0.45.1)
 Root-Is-Purelib: true
 Tag: py3-none-any