fix master conflict

Signed-off-by: sdp <sdp@9049fa09fd7b.jf.intel.com>

Author: sdp

neural_compressor/torch/algorithms/weight_only/gptq.py

@@ -1045,7 +1045,8 @@ def convert(self, model, *args, **kwargs):
         self.gptq_quantizer.remove_prepare_for_calibration()
 
         q_model, gptq_config = self.gptq_quantizer.execute_quantization()
-        q_model = q_model.to(self.model_device)
+        if not self.gptq_quantizer.use_layer_wise:
+            q_model = q_model.to(self.model_device)
         q_model.gptq_config = gptq_config
         logger.info("GPTQ quantizing done.")
         return q_model

neural_compressor/torch/algorithms/weight_only/rtn.py

@@ -146,7 +146,8 @@ def convert(
                 if dtype == "fp32":
                     continue
                 # Move modules to the accelerator device layer-by-layer
-                m.to(device)
+                if not use_layer_wise:
+                    m.to(device)
                 ### FP8 cast part
                 if dtype in ["fp8_e5m2", "fp8_e5m2fnuz", "fp8_e4m3fn", "fp8_e4m3fnuz"]:
                     logger.debug("Cast module {} to FP8 using qdq mode, no scaling".format(name))
@@ -200,7 +201,6 @@ def convert(
                 weight = m.weight.detach()
             if use_mse_search:
                 quantile = search_clip(m, bits, group_size, scheme, dtype, use_full_range)
-            start_quant = time.time()
             int_weight, scale, zp = quant_tensor(
                 weight,
                 dtype=dtype,
@@ -212,8 +212,6 @@ def convert(
                 full_range=use_full_range,
                 **double_quant_config,
             )
-            quant_int_time = time.time() - start_quant
-            total_quant_int_time += quant_int_time
             int_weight = int_weight.t_().contiguous() if transpose else int_weight
             scale = scale.t_().contiguous() if transpose else scale
             zp = zp.t_().contiguous() if transpose and zp is not None else zp
@@ -248,7 +246,9 @@ def convert(
             else:
                 set_module(model, name, new_module)
             # Move modules back to the model device layer-by-layer
-            m.to(model_device)
-            new_module.to(model_device)
-        model.to(model_device)
+            if not use_layer_wise:
+                m.to(model_device)
+                new_module.to(model_device)
+        if not use_layer_wise:
+            model.to(model_device)
         return model

neural_compressor/torch/quantization/config.py

@@ -200,7 +200,7 @@ def to_config_mapping(
         self, config_list: List[BaseConfig] = None, model_info: List[Tuple[str, str]] = None
     ) -> OrderedDictType[Union[str, str], OrderedDictType[str, BaseConfig]]:
         if not self.quant_lm_head:
-            self.set_local(LM_HEAD_NAMES, RTNConfig(dtype="fp32"))
+            self.set_local(LM_HEAD_NAMES, RTNConfig(dtype="fp32", use_layer_wise=self.use_layer_wise, model_path=self.model_path))
         config_mapping = super().to_config_mapping(config_list, model_info)
         return config_mapping
 
@@ -363,7 +363,7 @@ def to_config_mapping(
         self, config_list: List[BaseConfig] = None, model_info: List[Tuple[str, str]] = None
     ) -> OrderedDictType[Union[str, str], OrderedDictType[str, BaseConfig]]:
         if not self.quant_lm_head:
-            self.set_local(LM_HEAD_NAMES, GPTQConfig(dtype="fp32"))
+            self.set_local(LM_HEAD_NAMES, GPTQConfig(dtype="fp32", use_layer_wise=self.use_layer_wise, model_path=self.model_path))
         config_mapping = super().to_config_mapping(config_list, model_info)
         return config_mapping
 
@@ -385,7 +385,7 @@ def get_config_set_for_tuning(cls) -> Union[None, "GPTQConfig", List["GPTQConfig
     @classmethod
     def get_predefined_configs(cls) -> Dict[torch_utils.ProcessorType, "GPTQConfig"]:
         pre_defined_configs: Dict[torch_utils.ProcessorType, GPTQConfig] = {}
-        pre_defined_configs[torch_utils.ProcessorType.Client] = cls(use_layer_wise=True)
+        pre_defined_configs[torch_utils.ProcessorType.Client] = cls(use_layer_wise=True)#, model_path=self.model_path)
         pre_defined_configs[torch_utils.ProcessorType.Server] = cls()
         return pre_defined_configs
 
@@ -508,7 +508,7 @@ def to_config_mapping(
         self, config_list: List[BaseConfig] = None, model_info: List[Tuple[str, str]] = None
     ) -> OrderedDictType[Union[str, str], OrderedDictType[str, BaseConfig]]:
         if not self.quant_lm_head:
-            self.set_local(LM_HEAD_NAMES, AWQConfig(dtype="fp32"))
+            self.set_local(LM_HEAD_NAMES, AWQConfig(dtype="fp32", use_layer_wise=self.use_layer_wise, model_path=self.model_path))
         config_mapping = super().to_config_mapping(config_list, model_info)
         return config_mapping
 
@@ -815,7 +815,7 @@ def get_config_set_for_tuning(cls) -> Union[None, "AutoRoundConfig", List["AutoR
     @classmethod
     def get_predefined_configs(cls) -> Dict[torch_utils.ProcessorType, "AutoRoundConfig"]:
         pre_defined_configs: Dict[torch_utils.ProcessorType, AutoRoundConfig] = {}
-        pre_defined_configs[torch_utils.ProcessorType.Client] = cls(use_layer_wise=True)
+        pre_defined_configs[torch_utils.ProcessorType.Client] = cls(use_layer_wise=True, model_path=self.model_path)
         pre_defined_configs[torch_utils.ProcessorType.Server] = cls()
         return pre_defined_configs
 

test/3x/torch/quantization/weight_only/test_rtn.py

@@ -44,7 +44,7 @@ def setup_class(self):
         self.label = self.tiny_gptj(self.example_inputs)[0]
         # test_default_config
         model = copy.deepcopy(self.tiny_gptj)
-        quant_config = get_default_rtn_config()
+        quant_config = get_default_rtn_config("Server")
         model = prepare(model, quant_config)
         model = convert(model)
         # record q_label for comparison
@@ -172,7 +172,6 @@ def test_layer_wise(self):
         model = load_empty_model("hf-internal-testing/tiny-random-GPTJForCausalLM")
         quant_config = RTNConfig(
             use_layer_wise=True,
-            model_path="hf-internal-testing/tiny-random-GPTJForCausalLM",
         )
         model = prepare(model, quant_config)
         model = convert(model)