增加稠密模型channelwise autotuning 接口

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -613,6 +613,7 @@ class CompressedTensorsLinearMethod(LinearMethodBase):
             _weight=weight_data.T.contiguous().reshape(n,-1)
             layer.weight.data=_weight
             
+        self.tritonsingleton.gen_model_json() 
         layer.scheme.process_weights_after_loading(layer)   
         
     def create_weights(self, layer: torch.nn.Module,

vllm/utils.py

@@ -1900,7 +1900,7 @@ class W8a8GetCacheJSON:
         self.quant_method=None
 
     #析构函数，最后会生成model.json的配置文件
-    def gen_model_json(self,E:int,block_size:Optional[list]=None):
+    def gen_model_json(self,E:Optional[int]=0,block_size:Optional[list]=None):
         json_dir = os.getenv('LMSLIM_TUNING_JSON', "None")
         if json_dir is not "None" and os.path.exists(json_dir):
             #生成模型配置文件