Update main.py

main.py

@@ -102,7 +102,7 @@ async def demo():
     return HTMLResponse(content=html_content, status_code=200)
 
 @app.get("/stream")
-async def chat(prompt = "<|user|> Write a simple express style server in golang"):
+async def chat(prompt = "<|user|> Write an express server with server sent events. <|assistant|>"):
     tokens = llm.tokenize(prompt)
     async def server_sent_events(chat_chunks, llm):
         yield prompt
@@ -114,7 +114,7 @@ async def chat(prompt = "<|user|> Write a simple express style server in golang"
 
 @app.post("/v1/chat/completions")
 async def chat(request, response_mode=None):
-    tokens = llm.tokenize(request.messages.join(''))
+    tokens = llm.tokenize(request.messages)
     async def server_sent_events(chat_chunks, llm):
         for token in llm.generate(chat_chunks):
             yield llm.detokenize(token)