Update main.py

main.py

@@ -21,7 +21,7 @@ MAX_TOKENS = 2000
 TEMPERATURE = 0.7
 TOP_P = 0.95
 
-async def respond(message, history: list[tuple[str, str]]):
+def respond(message, history: list[tuple[str, str]]):
     messages = [{"role": "system", "content": SYSTEM_MESSAGE}]
     
     for val in history:
@@ -32,7 +32,7 @@ async def respond(message, history: list[tuple[str, str]]):
 
     messages.append({"role": "user", "content": message})
 
-    response = await client.chat_completion(
+    response = client.chat_completion(
         messages,
         max_tokens=MAX_TOKENS,
         stream=True,
@@ -40,12 +40,8 @@ async def respond(message, history: list[tuple[str, str]]):
         top_p=TOP_P,
     )
 
-    if hasattr(response, "__aiter__"):  # Check if the response supports async iteration
-        async for message in response:
-            yield message.choices[0].delta.content
-    else:
-        for message in response:  # Handle regular iteration
-            yield message.choices[0].delta.content
+    for message in response:  # Handle regular iteration
+        yield message.choices[0].delta.content
 
 @app.post("/generate/")
 async def generate(request: Request):
@@ -58,7 +54,7 @@ async def generate(request: Request):
 
     response_generator = respond(prompt, history)
     final_response = ""
-    async for part in response_generator:
+    for part in response_generator:
         final_response += part
 
     return JSONResponse(content={"response": final_response})