增加w8a8的triton调度支持

vllm/model_executor/models/llama.py

@@ -681,6 +681,7 @@ class LlamaForCausalLM(nn.Module, SupportsLoRA):
                 
                         json_file=self.tritonsingleton.get_w8a8json_name(n,k)
                         configs_dict=self.tritonsingleton.get_triton_cache(json_file,n,k)
+                        if configs_dict:
                             all_json.update(configs_dict)
                                               
             if self.w8a8_strategy==1:

vllm/model_executor/models/qwen.py

@@ -1122,6 +1122,7 @@ class QWenLMHeadModel(nn.Module, SupportsMultiModal):
                 
                         json_file=self.tritonsingleton.get_w8a8json_name(n,k)
                         configs_dict=self.tritonsingleton.get_triton_cache(json_file,n,k)
+                        if configs_dict:
                             all_json.update(configs_dict)
                                               
             if self.w8a8_strategy==1:

vllm/utils.py

@@ -1414,4 +1414,6 @@ class W8a8GetCacheJSON:
         return configs_dict
   
     def get_w8a8json_name(self,n,k):
-        return self.triton_json_dir+f"/W8A8_{n}_{k}_DCUK100AI.json"
+        device_name = current_platform.get_device_name().replace(" ", "_")
+        return self.triton_json_dir+f"/W8A8_{n}_{k}_DCU{device_name}.json"
+        
\ No newline at end of file