fix english query bug (#840)

### What problem does this PR solve?

#834 

### Type of change

- [x] Bug Fix (non-breaking change which fixes an issue)

rag/llm/chat_model.py

@@ -20,6 +20,7 @@ from openai import OpenAI
 import openai
 from ollama import Client
 from rag.nlp import is_english
+from rag.utils import num_tokens_from_string
 
 
 class Base(ABC):
@@ -255,3 +256,46 @@ class OllamaChat(Base):
         except Exception as e:
             yield ans + "\n**ERROR**: " + str(e)
         yield 0
+
+
+class LocalLLM(Base):
+    class RPCProxy:
+        def __init__(self, host, port):
+            self.host = host
+            self.port = int(port)
+            self.__conn()
+
+        def __conn(self):
+            from multiprocessing.connection import Client
+            self._connection = Client(
+                (self.host, self.port), authkey=b'infiniflow-token4kevinhu')
+
+        def __getattr__(self, name):
+            import pickle
+
+            def do_rpc(*args, **kwargs):
+                for _ in range(3):
+                    try:
+                        self._connection.send(
+                            pickle.dumps((name, args, kwargs)))
+                        return pickle.loads(self._connection.recv())
+                    except Exception as e:
+                        self.__conn()
+                raise Exception("RPC connection lost!")
+
+            return do_rpc
+
+    def __init__(self, key, model_name="glm-3-turbo"):
+        self.client = LocalLLM.RPCProxy("127.0.0.1", 7860)
+
+    def chat(self, system, history, gen_conf):
+        if system:
+            history.insert(0, {"role": "system", "content": system})
+        try:
+            ans = self.client.chat(
+                history,
+                gen_conf
+            )
+            return ans, num_tokens_from_string(ans)
+        except Exception as e:
+            return "**ERROR**: " + str(e), 0

rag/llm/rpc_server.py

@@ -2,9 +2,10 @@ import argparse
 import pickle
 import random
 import time
+from copy import deepcopy
 from multiprocessing.connection import Listener
 from threading import Thread
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
 
 
 def torch_gc():
@@ -95,6 +96,32 @@ def chat(messages, gen_conf):
         return str(e)
 
 
+def chat_streamly(messages, gen_conf):
+    global tokenizer
+    model = Model()
+    try:
+        torch_gc()
+        conf = deepcopy(gen_conf)
+        print(messages, conf)
+        text = tokenizer.apply_chat_template(
+            messages,
+            tokenize=False,
+            add_generation_prompt=True
+        )
+        model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+        streamer = TextStreamer(tokenizer)
+        conf["inputs"] = model_inputs.input_ids
+        conf["streamer"] = streamer
+        conf["max_new_tokens"] = conf["max_tokens"]
+        del conf["max_tokens"]
+        thread = Thread(target=model.generate, kwargs=conf)
+        thread.start()
+        for _, new_text in enumerate(streamer):
+            yield new_text
+    except Exception as e:
+        yield "**ERROR**: " + str(e)
+
+
 def Model():
     global models
     random.seed(time.time())
@@ -113,6 +140,7 @@ if __name__ == "__main__":
 
     handler = RPCHandler()
     handler.register_function(chat)
+    handler.register_function(chat_streamly)
 
     models = []
     for _ in range(1):