Update main.py

main.py

@@ -53,7 +53,7 @@ class Message(BaseModel):
 
 class ChatCompletionRequest(BaseModel):
     messages: List[Message]
-    max_tokens: int = 100
+    max_tokens: int = 250
 
 @app.post("/v1/completions")
 async def completion(request: ChatCompletionRequest, response_mode=None):
@@ -66,7 +66,7 @@ async def chat(request: ChatCompletionRequest):
     tokens = llm.tokenize(combined_messages)
     
     try:
-        chat_chunks = llm.generate(tokens, max_tokens=request.max_tokens)
+        chat_chunks = llm.generate(tokens, max_new_tokens=request.max_tokens)
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))