fix params bug

llm_service/inferencer.py

@@ -43,7 +43,7 @@ class InferenceWrapper:
                     self.model = llm.from_hf(self.model, self.tokenizer, dtype="float16").cuda()
 
             except Exception as e:
-                logger.error(str(e))
+                logger.error(f"fastllm initial failed, {e}")
 
 
     def chat(self, prompt: str, history=[]):
@@ -58,7 +58,7 @@ class InferenceWrapper:
                                                     history,
                                                     do_sample=False)
         except Exception as e:
-            logger.error(str(e))
+            logger.error(f"chat inference failed, {e}")
         return output_text
 
 

llm_service/worker.py

@@ -13,12 +13,12 @@ class ChatAgent:
         self.reranker_model_path = config['feature_database']['reranker_model_path']
         reject_throttle = float(config['feature_database']['reject_throttle'])
         local_llm_path = config['llm']['local_llm_path']
-        accelerate = config.getboolean('llm', 'accelerate')
+        use_vllm = config.getboolean('llm', 'use_vllm')
 
         self.retriever = CacheRetriever(self.embedding_model_path,
                                         self.reranker_model_path).get(reject_throttle=reject_throttle,
                                                                       work_dir=self.work_dir)
-        self.llm_server = LLMInference(local_llm_path, tensor_parallel_size, accelerate=accelerate)
+        self.llm_server = LLMInference(local_llm_path, tensor_parallel_size, use_vllm=use_vllm)
 
     def generate_prompt(self,
                         history_pair,

server_start.py

@@ -116,20 +116,17 @@ def auto_select_dcu(config):
 def parse_args():
     parser = argparse.ArgumentParser(description='Start all services.')
     parser.add_argument('--config_path',
-                        default='/path/of/config.ini',
+                        default='./config.ini',
                         help='Config directory')
     parser.add_argument('--log_path',
-                        default='',
+                        default='./log/assistant.log',
                         help='Set log file path')
     return parser.parse_args()
 
 
 def main():
     args = parse_args()
-    log_path = './log/assistant.log'
-    if args.log_path:
-        log_path = args.log_path
-    logger.add(sink=log_path, level="DEBUG", rotation="500MB", compression="zip", encoding="utf-8", enqueue=True)
+    logger.add(sink=args.log_path, level="DEBUG", rotation="500MB", compression="zip", encoding="utf-8", enqueue=True)
     workflow(args)