Skip to content

GKD在使用vllm api-server部署教师模型时,似乎对多模态输入的处理有问题 #9011

@WhiteGrayxp

Description

@WhiteGrayxp

Checklist / 检查清单

  • I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues,确认这是一个新的问题与讨论。

Question Description / 问题描述

rt,在使用api部署教师模型,获取logprobs,使用了如下的实现,payload prompt部分似乎仅包含了文本部分token,对于多模态模型训练可能会导致教师实际没有收到图片信息(仅占位符)?

def _fetch_one(batch_idx):
        payload = {
            'model': model,
            'prompt': input_ids[batch_idx],
            'max_tokens': 1,
            'temperature': 0,
            'prompt_logprobs': topk,
        }
        try:
            resp = session.post(url, json=payload, timeout=timeout)
            resp.raise_for_status()
            prompt_logprobs_list = resp.json()['choices'][0].get('prompt_logprobs', [])
       ……

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingquestionFurther information is requested

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions