Fix bugs

config.ini

@@ -12,4 +12,5 @@ reranker_model_path=/path/to/your/bce-reranker-base_v1
 [llm]
 local_llm_path=/path/to/your/internlm-chat-7b
 use_vllm=False
-stream_chat=False
\ No newline at end of file
+stream_chat=False
+tensor_parallel_size=1
\ No newline at end of file

llm_service/inferencer.py

@@ -68,7 +68,7 @@ class LLMInference:
 
     def __init__(self,
                  model,
-                 tokenzier,
+                 tokenizer,
                  device: str = 'cuda',
                  use_vllm: bool = False,
                  stream_chat: bool = False
@@ -76,7 +76,7 @@ class LLMInference:
 
         self.device = device
         self.model = model
-        self.tokenzier = tokenzier
+        self.tokenizer = tokenizer
         self.use_vllm = use_vllm
         self.stream_chat = stream_chat
 
@@ -170,16 +170,14 @@ class LLMInference:
 
 def init_model(model_path, use_vllm=False, tp_size=1):
     ## init models
-    # huggingface
-    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
-    model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True,device_map="auto").half().cuda().eval()
 
+    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
     if use_vllm:
         try:
             # vllm
             from vllm import LLM, SamplingParams
 
-            tokenizer = SamplingParams(temperature=1,
+            sampling_params = SamplingParams(temperature=1,
                                             top_p=0.95,
                                             max_tokens=1024,
                                             stop_token_ids=[tokenizer.eos_token_id])
@@ -189,10 +187,13 @@ def init_model(model_path, use_vllm=False, tp_size=1):
                         enforce_eager=True,
                         dtype="float16",
                         tensor_parallel_size=tp_size)
+            return model, sampling_params
         except Exception as e:
-            logger.error(f"fastllm initial failed, {e}")
-
-    return model, tokenizer
+            logger.error(f"vllm initial failed, {e}")
+    else:
+        # huggingface
+        model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True).half().cuda().eval()
+        return model, tokenizer
 
 
 def llm_inference(args):
@@ -204,13 +205,13 @@ def llm_inference(args):
     model_path = config['llm']['local_llm_path']
     tensor_parallel_size = config.getint('llm', 'tensor_parallel_size')
     use_vllm = config.getboolean('llm', 'use_vllm')
-    print("inference")
+    stream_chat = config.getboolean('llm', 'stream_chat')
+    logger.info(f"Get params: model_path {model_path}, use_vllm {use_vllm}, tensor_parallel_size {tensor_parallel_size}, stream_chat {stream_chat}")
 
     model, tokenzier = init_model(model_path, use_vllm, tensor_parallel_size)
     inference = LLMInference(model,
                             tokenzier,
                             use_vllm=use_vllm,
-                            tensor_parallel_size=tensor_parallel_size,
                             stream_chat=args.stream_chat)
     async def inference(request):
         start = time.time()
@@ -239,21 +240,19 @@ def infer_test(args):
     use_vllm = config.getboolean('llm', 'use_vllm')
     tensor_parallel_size = config.getint('llm', 'tensor_parallel_size')
     stream_chat = config.getboolean('llm', 'stream_chat')
+    logger.info(f"Get params: model_path {model_path}, use_vllm {use_vllm}, tensor_parallel_size {tensor_parallel_size}, stream_chat {stream_chat}")
 
     model, tokenzier = init_model(model_path, use_vllm, tensor_parallel_size)
     inference = LLMInference(model,
-                                     tokenzier,
-                                     use_vllm=use_vllm,
-                                     tensor_parallel_size=tensor_parallel_size,
-                                     stream_chat=stream_chat)
-    # prompt = "hello，please introduce yourself..."
-    prompt ='65N32-US主板清除CMOS配置的方法'
+                            tokenzier,
+                            use_vllm=use_vllm,
+                            stream_chat=stream_chat)
 
     time_first = time.time()
-    output_text = inference.chat(prompt)
+    output_text = inference.chat(args.query)
     time_second = time.time()
     logger.debug('问题:{} 回答:{} \ntimecost {} '.format(
-        prompt, output_text, time_second - time_first))
+        args.query, output_text, time_second - time_first))
 
 
 def set_envs(dcu_ids):
@@ -282,10 +281,6 @@ def parse_args():
         type=str,
         default='0,1',
         help='设置DCU卡号,卡号之间用英文逗号隔开,输入样例:"0,1,2"')
-    parser.add_argument(
-        '--stream_chat',
-        action='store_true',
-        help='启用流式对话方式')
     args = parser.parse_args()
     return args