Possible invalid request formatting for `max_completion_tokens`

[stmcginnis]

Just looking in to this, but wanted to report it in case it was a known issue or someone has more information.

While running guidellm against a locally running vllm serve, I am seeing a very large amount of these log messages in the vLLM output:

WARNING 06-27 14:57:36 [protocol.py:58] The following fields were present in the request but ignored: {'max_completion_tokens'}

Running a request manually against the endpoint is happy, with no errors in the vllm logs:

curl -k -s -H "Content-Type: application/json" http://localhost:8000/v1/chat/completions -d '{"model":"llama3.1-8b-instruct","messages":[{"role":"user","content":"What is an AI tensorized weight?"}],"max_completion_tokens":35}' | jq .

{
  "id": "chatcmpl-e00ce83f-121f-482a-b43c-5c4494ad29ae",
  "object": "chat.completion",
  "created": 1751037186,
  "model": "llama3.1-8b-instruct",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "reasoning_content": null,
        "content": "A tensorized weight, also known as a tensor weight or a weight tensor, is a type of weight used in artificial neural networks (ANNs) and deep learning models.",
        "tool_calls": []
      },
      "logprobs": null,
      "finish_reason": "length",
      "stop_reason": null
    }
  ],
  "usage": {
    "prompt_tokens": 43,
    "total_tokens": 78,
    "completion_tokens": 35,
    "prompt_tokens_details": null
  },
  "prompt_logprobs": null
}

Which leads me to believe the prompt being formed by guidellm must be placing max_completion_tokens somewhere other than as a top level property of the request struct.

[sjmonson]

Tldr: you can ignore that error. We set both max_completion_tokens and max_tokens to the same value.

The standard for setting a max tokens is kinda of messy. Some model servers support max_tokens as the flag and some support max_completion_tokens. In the official OpenAI documentation the legacy endpoint only supports max_tokens while the chat endpoint supports max_tokens for older models and max_completion_tokens for everything. Ollama supports only max_completion_tokens

vLLM started with max_completion_tokens but at some point switched to max_tokens and throws a harmless warning for the former. We kept both in GuideLLM for compatibility, but some model servers cough cough Ollama don't like that. As of now if you have an issue you can set --backend_args="{'remove_from_body':['max_completion_tokens']}" to remove it from the request. In the future we may support custom request templates and default to only having max_tokens (#194).