Chat title is "Untitled" when using reasoning model

[Leon-Sander]

What happened?

Hello, I am using vllm with Qwen/QwQ-32B-AWQ as model.
For the vllm docker image I added the parameters
"--reasoning-parser", "deepseek_r1",
"--enable-reasoning"
Now I see the "thinking" field in the UI, but the titles do not work anymore. In the vllm logs further down you can see that vllm gets the request and seems to return an answer.

Version Information

Running from docker
ghcr.io/danny-avila/librechat-dev latest 07698c5407cb 2 days ago 950MB
ghcr.io/danny-avila/librechat-rag-api-dev latest 96061ddf7d66 4 weeks ago 7.81GB
ghcr.io/danny-avila/librechat-rag-api-dev-lite latest 817faebddae9 4 weeks ago 1.31GB

Steps to Reproduce

Setup Librechat with vllm in the librecaht.yaml:

  custom:
    - name: "vllm" # Internal identifier for this endpoint
      apiKey: "vllm" # Dummy API key (vLLM OpenAI endpoint often doesn't require one)
      baseURL: "http://host.docker.internal:8080/v1" # Corrected host, protocol, and standard base path
      models:
        # Replace "your-model-id-served-by-vllm" with the actual model ID
        # Example: "mistralai/Mistral-7B-Instruct-v0.1"
        default:
          - "Qwen/QwQ-32B-AWQ"
        # fetch: true will attempt to query the /v1/models endpoint
        fetch: true
      titleConvo: true
      # titleModel: "current_model" # Use the currently selected model (usually best)
      # To use a specific model for titles (must be in the default list or fetched):
      titleModel: "current_model"
      summarize: true
      summaryModel: "current_model" # Use the currently selected model if summarize were true
      forcePrompt: false
      modelDisplayLabel: "vLLM" # How the endpoint group appears in the UI

Start Librechat and vllm, here the vllm docker compose:

services:
  vllm_qwq32bawq:
    image: vllm/vllm-openai:v0.8.5.post1
    command: [
      "--device", "cuda",
      "--model", "Qwen/QwQ-32B-AWQ",
      "--host", "0.0.0.0",
      "--port", "8080",
      "--quantization", "awq_marlin",
      "--gpu-memory-utilization", "0.95",
      "--max-model-len", "32768",
      "--enable-prefix-caching",
      "--enable-auto-tool-choice",
      "--tool-call-parser", "hermes",
      "--reasoning-parser", "deepseek_r1",
      "--enable-reasoning"
    ]
    ports:
      - "8080:8080"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

    ipc: host

What browsers are you seeing the problem on?

Firefox, Brave
Pretty sure its not browser related.

Relevant log output

I didnt find relevant log outputs in the libreChat docker container, is it in another conatainer?

vllm logs:

INFO 05-07 07:08:36 [logger.py:39] Received request chatcmpl-c01ad60ffa0f464d988e5680ccd4eaee: prompt: '<|im_start|>user\ntest<|im_end|>\n<|im_start|>assistant\n<think>\n', params: SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0.6, top_p=0.95, top_k=40, min_p=0.0, seed=None, stop=[], stop_token_ids=[], bad_words=[], include_stop_str_in_output=False, ignore_eos=False, max_tokens=32757, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None, guided_decoding=None, extra_args=None), prompt_token_ids: None, lora_request: None, prompt_adapter_request: None.
INFO:     172.18.0.1:54588 - "POST /v1/chat/completions HTTP/1.1" 200 OK
INFO 05-07 07:08:36 [async_llm.py:239] Added request chatcmpl-c01ad60ffa0f464d988e5680ccd4eaee.
INFO 05-07 07:08:38 [logger.py:39] Received request chatcmpl-db4a2663ae8d46f3ab06f487a366bed7: prompt: '<|im_start|>system\nPlease generate a concise, 5-word-or-less title for the conversation, using its same language, with no punctuation. Apply title case conventions appropriate for the language. Never directly mention the language name or the word "title"\n\n||>User:\n"test"\n||>Response:\n"":::thinking\\n好的，用户发来了“test”，看起来可能是测试消息或者误发。我需要确认他们是否需要帮助。先回复一个友好的询问，看看他们具体需要什么帮助。保持简洁，避免冗长，确保用户知道可以继续说明他们的需求。同时，保持专业和友好的语气，让用户感到被重视和支持。接下来应该等待用户的进一步回复，再根据具体问题提供相应的帮助。\\n:::\\n\\n\\n您好！看起来您发送了一个测试消息。如果您有任何问题或需要帮助，请随时告诉我，我会尽力提供支持。😊""\n\n||>Title:<|im_end|>\n<|im_start|>assistant\n<think>\n', params: SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0.2, top_p=0.95, top_k=40, min_p=0.0, seed=None, stop=[], stop_token_ids=[], bad_words=[], include_stop_str_in_output=False, ignore_eos=False, max_tokens=16, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None, guided_decoding=None, extra_args=None), prompt_token_ids: None, lora_request: None, prompt_adapter_request: None.
INFO 05-07 07:08:38 [async_llm.py:239] Added request chatcmpl-db4a2663ae8d46f3ab06f487a366bed7.
INFO:     172.18.0.1:54588 - "POST /v1/chat/completions HTTP/1.1" 200 OK

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

Reasoning models are really not recommended for titles. The title method uses limited tokens, and there is a limited time to generate the title, so more than likely you will "Untitled" (a title was not generated).

Look into customizing this here: https://www.librechat.ai/docs/configuration/librechat_yaml/object_structure/custom_endpoint#titlemodel

[Leon-Sander]

Well we are working with local models and the one reasoning model we are using is fully utilizing my gpu, so I cant run a seperate smaller non reasoning model just for the titles. Would be nice to be able to also generate titles with the reasoning model.

[Nindaleth]

It's hard for me as well to use a different local model for titles as I'm already resource constrained locally and I'm obviously reluctant to use a remote model for this. I usually get something like "<think> So the user wants to know about" when using a local reasoning model.

Some of the modern reasoning models allow adding "/no_think" to the user prompt text, setting a enable_thinking parameter of the request or starting the assistant response with "<think> </think>", would some of that be viable?

[Originalimoc]

@danny-avila
This really needs a solution since Qwen3 is currently best local(32B super smart and 30B-A3B is 100tk/s+ super fast) you can get and like QwQ/R1 also needs this. So my 2 cents here:
1\Add /no_think to that title prompt if detected as a Qwen3 model even for all models I see no reason this will break other models.
2\And for other model you can use prefill(last message is assistant, content is eg: "The suitable title is " + pass add_generation_prompt = false for supported endpoint).