compressed-tensors 0.8.0__tar.gz → 0.8.1__tar.gz

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors
-Version: 0.8.0
+Version: 0.8.1
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/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-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/compressors/quantized_compressors/naive_quantized.py

@@ -93,9 +93,11 @@ class NaiveQuantizationCompressor(BaseQuantizationCompressor):
                 args=quantization_args,
                 dtype=quantization_args.pytorch_dtype(),
             )
+        else:
+            quantized_weight = weight
 
-            if device is not None:
-                quantized_weight = quantized_weight.to(device)
+        if device is not None:
+            quantized_weight = quantized_weight.to(device)
 
         return {"weight": quantized_weight}
 

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/compressors/quantized_compressors/pack_quantized.py

@@ -94,6 +94,8 @@ class PackedQuantizationCompressor(BaseQuantizationCompressor):
                 args=quantization_args,
                 dtype=torch.int8,
             )
+        else:
+            quantized_weight = weight
 
         packed_weight = pack_to_int32(quantized_weight, quantization_args.num_bits)
         weight_shape = torch.tensor(weight.shape)

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/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-0.8.0 → compressed-tensors-0.8.1}/src/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-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/quantization/quant_args.py

@@ -17,6 +17,7 @@ from enum import Enum
 from typing import Any, Dict, Optional, Union
 
 import torch
+from compressed_tensors.utils import Aliasable
 from pydantic import BaseModel, Field, field_validator, model_validator
 
 
@@ -53,17 +54,29 @@ class QuantizationStrategy(str, Enum):
     TOKEN = "token"
 
 
-class ActivationOrdering(str, Enum):
+class ActivationOrdering(Aliasable, str, Enum):
     """
     Enum storing strategies for activation ordering
 
     Group: reorder groups and weight\n
-    Weight: only reorder weight, not groups. Slightly lower latency and
-    accuracy compared to group actorder\n
+    Weight: only reorder weight, not groups. Slightly lower accuracy but also lower
+    latency when compared to group actorder\n
+    Dynamic: alias for Group\n
+    Static: alias for Weight\n
     """
 
     GROUP = "group"
     WEIGHT = "weight"
+    # aliases
+    DYNAMIC = "dynamic"
+    STATIC = "static"
+
+    @staticmethod
+    def get_aliases() -> Dict[str, str]:
+        return {
+            "dynamic": "group",
+            "static": "weight",
+        }
 
 
 class QuantizationArgs(BaseModel, use_enum_values=True):

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/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-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/quantization/quant_scheme.py

@@ -13,14 +13,14 @@
 # limitations under the License.
 
 from copy import deepcopy
-from typing import List, Optional
+from typing import Any, Dict, List, Optional
 
 from compressed_tensors.quantization.quant_args import (
     QuantizationArgs,
     QuantizationStrategy,
     QuantizationType,
 )
-from pydantic import BaseModel
+from pydantic import BaseModel, model_validator
 
 
 __all__ = [
@@ -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
@@ -47,27 +47,20 @@ 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,
-        )
+    @model_validator(mode="after")
+    def validate_model_after(model: "QuantizationArgs") -> Dict[str, Any]:
+        inputs = model.input_activations
+        outputs = model.output_activations
+
+        if inputs is not None:
+            if inputs.actorder is not None:
+                raise ValueError("Cannot apply actorder to input activations")
+
+        if outputs is not None:
+            if outputs.actorder is not None:
+                raise ValueError("Cannot apply actorder to output activations")
+
+        return model
 
 
 """

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/utils/helpers.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Any, Optional
+from typing import Any, Dict, Optional
 
 import torch
 from transformers import AutoConfig
@@ -24,6 +24,7 @@ __all__ = [
     "tensor_follows_mask_structure",
     "replace_module",
     "is_compressed_tensors_config",
+    "Aliasable",
 ]
 
 FSDP_WRAPPER_NAME = "_fsdp_wrapped_module"
@@ -119,3 +120,34 @@ def is_compressed_tensors_config(compression_config: Any) -> bool:
         return isinstance(compression_config, CompressedTensorsConfig)
     except ImportError:
         return False
+
+
+class Aliasable:
+    """
+    A mixin for enums to allow aliasing of enum members
+
+    Example:
+    >>> class MyClass(Aliasable, int, Enum):
+    >>>     ...
+    """
+
+    @staticmethod
+    def get_aliases() -> Dict[str, str]:
+        raise NotImplementedError()
+
+    def __eq__(self, other):
+        if isinstance(other, self.__class__):
+            aliases = self.get_aliases()
+            return self.value == other.value or (
+                aliases.get(self.value, self.value)
+                == aliases.get(other.value, other.value)
+            )
+        else:
+            aliases = self.get_aliases()
+            self_value = aliases.get(self.value, self.value)
+            other_value = aliases.get(other, other)
+            return self_value == other_value
+
+    def __hash__(self):
+        canonical_value = self.aliases.get(self.value, self.value)
+        return hash(canonical_value)

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors/version.py

@@ -17,7 +17,7 @@ Functionality for storing and setting the version info for SparseML
 """
 
 
-version_base = "0.8.0"
+version_base = "0.8.1"
 is_release = True  # change to True to set the generated version as a release version
 
 

{compressed-tensors-0.8.0 → compressed-tensors-0.8.1}/src/compressed_tensors.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: compressed-tensors
-Version: 0.8.0
+Version: 0.8.1
 Summary: Library for utilization of compressed safetensors of neural network models
 Home-page: https://github.com/neuralmagic/compressed-tensors
 Author: Neuralmagic, Inc.