[Data] No token logprobs for ray.data.llm

[zhangzheyu13]

Description

I'm trying to use ray.data.llm to do DP inference with multi-gpu following this wiki: https://docs.ray.io/en/latest/data/working-with-llms.html

I would like to get the logprobs of the output tokens, but the response does not contain that, even if I enabled log_probs in sampling_params (I think this works in native vllm):

guided_decoding_params = dict(choice=["true", "false"])
self.sampling_params = dict(
    max_tokens=16,
    temperature=0,
    stop_token_ids=stop_token_ids,
    guided_decoding=guided_decoding_params,
    logprobs=1,
)
processor = build_llm_processor(
    self.vllm_engine_processor_config,
    preprocess=lambda row: dict(
        messages=[
            {"role": "user", "content": self._record_to_prompt(row)},
        ],
        sampling_params=self.sampling_params,
    ),
    postprocess=lambda row: self._get_llm_prediction(row),
)

I think it'd make sense to pass the logprobs to the ray data response if vllm already supports it. If there's a way to access it that I missed please point out as well, thanks

Use case

Being able to access logprobs is important in ray data offline inference