Allow ModelCompressor.from_pretrained to load from quantization_confi…

…g, not compression config (#207)

src/compressed_tensors/compressors/model_compressors/model_compressor.py

@@ -24,7 +24,6 @@
 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 @@
     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 @@ def from_pretrained(
         :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 @@ def compress(
             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 @@ def _replace_weights(self, dense_weight_generator, model):
             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):

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 @@ def from_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
         )
 

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

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]]]