Port torch GPTQ to 3.x (#1408)

Signed-off-by: yiliu30 <[email protected]>

.azure-pipelines/scripts/ut/run_itrex.sh

@@ -6,6 +6,7 @@ echo "run itrex ut..."
 
 # prepare itrex
 git clone https://github.com/intel/intel-extension-for-transformers.git /intel-extension-for-transformers
+cd /intel-extension-for-transformers && git rev-parse --short HEAD
 bash /intel-extension-for-transformers/.github/workflows/script/prepare_env.sh
 bash /intel-extension-for-transformers/.github/workflows/script/install_binary.sh
 

neural_compressor/common/base_config.py

@@ -118,12 +118,16 @@ def from_dict(cls, config_dict, str2operator=None):
         Returns:
             The constructed config.
         """
-        config = cls(**config_dict.get(GLOBAL, {}))
-        operator_config = config_dict.get(LOCAL, {})
-        if operator_config:
-            for op_name, op_config in operator_config.items():
-                config.set_local(op_name, cls(**op_config))
-        return config
+        if GLOBAL not in config_dict and LOCAL not in config_dict:
+            config = cls(**config_dict)
+            return config
+        else:
+            config = cls(**config_dict.get(GLOBAL, {}))
+            operator_config = config_dict.get(LOCAL, {})
+            if operator_config:
+                for op_name, op_config in operator_config.items():
+                    config.set_local(op_name, cls(**op_config))
+            return config
 
     @classmethod
     def to_diff_dict(cls, instance) -> Dict[str, Any]:
@@ -201,11 +205,11 @@ def to_config_mapping(
             global_config = config.global_config
             op_type_config_dict, op_name_config_dict = config._get_op_name_op_type_config()
             for op_name, op_type in model_info:
-                config_mapping.setdefault(op_type, OrderedDict())[op_name] = global_config
+                config_mapping[(op_type, op_name)] = global_config
                 if op_type in op_type_config_dict:
-                    config_mapping[op_type][op_name] = op_name_config_dict[op_type]
+                    config_mapping[(op_type, op_name)] = op_name_config_dict[op_type]
                 if op_name in op_name_config_dict:
-                    config_mapping[op_type][op_name] = op_name_config_dict[op_name]
+                    config_mapping[(op_type, op_name)] = op_name_config_dict[op_name]
         return config_mapping
 
     @staticmethod
@@ -234,9 +238,15 @@ def to_dict(self, params_list=[], operator2str=None):
         return result
 
     @classmethod
-    def from_dict(cls, config_dict, str2operator=None):
-        # TODO(Yi)
-        pass
+    def from_dict(cls, config_dict: OrderedDict[str, Dict], config_registry: Dict[str, BaseConfig]):
+        assert len(config_dict) >= 1, "The config dict must include at least one configuration."
+        num_configs = len(config_dict)
+        name, value = next(iter(config_dict.items()))
+        config = config_registry[name].from_dict(value)
+        for _ in range(num_configs - 1):
+            name, value = next(iter(config_dict.items()))
+            config += config_registry[name].from_dict(value)
+        return config
 
     def to_json_string(self, use_diff: bool = False) -> str:
         return json.dumps(self.to_dict(), indent=2) + "\n"

neural_compressor/common/utility.py

@@ -26,4 +26,5 @@
 BASE_CONFIG = "base_config"
 COMPOSABLE_CONFIG = "composable_config"
 RTN_WEIGHT_ONLY_QUANT = "rtn_weight_only_quant"
+GPTQ = "gptq"
 DUMMY_CONFIG = "dummy_config"

neural_compressor/torch/__init__.py

@@ -13,12 +13,14 @@
 # limitations under the License.
 
 from neural_compressor.torch.utils import register_algo
-from neural_compressor.torch.algorithms import rtn_quantize_entry
+from neural_compressor.torch.algorithms import rtn_quantize_entry, gptq_quantize_entry
 
 from neural_compressor.torch.quantization import (
     quantize,
     RTNWeightQuantConfig,
     get_default_rtn_config,
     DummyConfig,
     get_default_dummy_config,
+    GPTQConfig,
+    get_default_gptq_config,
 )

neural_compressor/torch/algorithms/__init__.py

@@ -13,4 +13,5 @@
 # limitations under the License.
 
 
-from neural_compressor.torch.algorithms.rtn_quantize import rtn_quantize_entry
+from neural_compressor.torch.algorithms.weight_only_algos import rtn_quantize_entry
+from neural_compressor.torch.algorithms.weight_only_algos import gptq_quantize_entry