refine config and interface based on RFC design

Signed-off-by: xin3he <xin3.he@intel.com>

Author: xin3he

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/llm/run_clm_no_trainer.py

@@ -230,8 +230,8 @@ def get_user_model():
 
     # 3.x api
     if args.approach == 'weight_only':
-        from neural_compressor.torch import RTNWeightQuantConfig, GPTQConfig, quantize
-        from neural_compressor.torch.utils.utility import get_double_quant_config
+        from neural_compressor.torch.quantization import RTNConfig, GPTQConfig, quantize
+        from neural_compressor.torch.utils import get_double_quant_config
         weight_sym = True if args.woq_scheme == "sym" else False
         double_quant_config_dict = get_double_quant_config(args.double_quant_type, weight_sym=weight_sym)
 
@@ -243,9 +243,9 @@ def get_user_model():
                         "enable_mse_search": args.woq_enable_mse_search,
                     }
                 )
-                quant_config = RTNWeightQuantConfig.from_dict(double_quant_config_dict)
+                quant_config = RTNConfig.from_dict(double_quant_config_dict)
             else:
-                quant_config = RTNWeightQuantConfig(
+                quant_config = RTNConfig(
                     weight_dtype=args.woq_dtype,
                     weight_bits=args.woq_bits,
                     weight_group_size=args.woq_group_size,
@@ -257,7 +257,7 @@ def get_user_model():
                     double_quant_sym=args.double_quant_sym,
                     double_quant_group_size=args.double_quant_group_size,
                 )
-            quant_config.set_local("lm_head", RTNWeightQuantConfig(weight_dtype="fp32"))
+            quant_config.set_local("lm_head", RTNConfig(weight_dtype="fp32"))
             user_model = quantize(
                 model=user_model, quant_config=quant_config
             )

neural_compressor/common/base_config.py

@@ -127,7 +127,7 @@ def set_local(self, operator_name: str, config: BaseConfig) -> BaseConfig:
         self.local_config[operator_name] = config
         return self
 
-    def to_dict(self, params_list=[], operator2str=None):
+    def to_dict(self):
         result = {}
         global_config = self.get_params_dict()
         if bool(self.local_config):
@@ -147,7 +147,7 @@ def get_params_dict(self):
         return result
 
     @classmethod
-    def from_dict(cls, config_dict, str2operator=None):
+    def from_dict(cls, config_dict):
         """Construct config from a dict.
 
         Args:

neural_compressor/common/utility.py

@@ -27,7 +27,7 @@
 # config name
 BASE_CONFIG = "base_config"
 COMPOSABLE_CONFIG = "composable_config"
-RTN_WEIGHT_ONLY_QUANT = "rtn_weight_only_quant"
+RTN = "rtn"
 STATIC_QUANT = "static_quant"
 GPTQ = "gptq"
 FP8_QUANT = "fp8_quant"

neural_compressor/tensorflow/utils.py

@@ -35,7 +35,7 @@ def register_algo(name):
 
     Usage example:
         @register_algo(name=example_algo)
-        def example_algo(model: torch.nn.Module, quant_config: RTNWeightQuantConfig) -> torch.nn.Module:
+        def example_algo(model: torch.nn.Module, quant_config: RTNConfig) -> torch.nn.Module:
             ...
     Args:
         name (str): The name under which the algorithm function will be registered.

neural_compressor/torch/__init__.py

@@ -11,16 +11,3 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
-from neural_compressor.torch.utils.utility import register_algo
-from neural_compressor.torch.algorithms import rtn_quantize_entry, gptq_quantize_entry
-
-from neural_compressor.torch.quantization import (
-    quantize,
-    RTNWeightQuantConfig,
-    get_default_rtn_config,
-    GPTQConfig,
-    get_default_gptq_config,
-)
-
-from neural_compressor.torch.tune import autotune, TuningConfig, get_default_tune_config

neural_compressor/torch/algorithms/__init__.py

@@ -13,5 +13,7 @@
 # limitations under the License.
 
 
-from neural_compressor.torch.algorithms.weight_only_algos import rtn_quantize_entry
-from neural_compressor.torch.algorithms.weight_only_algos import gptq_quantize_entry
+from .weight_only import (
+    rtn_quantize,
+    gptq_quantize,
+)

neural_compressor/torch/algorithms/weight_only/README.md

@@ -0,0 +1 @@
+# Demo of algorithm usage w/o INC

neural_compressor/torch/algorithms/weight_only/__init__.py

@@ -11,3 +11,6 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+from .rtn import rtn_quantize
+from .gptq import gptq_quantize