Skip to content

发现问题,采用原始的 transformer推理,vllm offline推理以及vllm serve推理,结果相差很大 #76

@AbbottKilig

Description

@AbbottKilig

transformers
vllm local method: https://github.com/QwenLM/Qwen3-VL-Embedding/blob/main/examples/reranker_vllm.ipynb
vllm openai method: https://github.com/vllm-project/vllm/blob/main/examples/pooling/score/vision_rerank_api_online.py
以方法 1 为标准答案,我发现方法 2 与之差别不大,但方法 3 的结果却非常奇怪:得分都在 0.5 倍左右。请问有人能提供一些建议吗?查询和文档均为图像。

vllm issue区也遇到相似的问题
vllm-project/vllm#34502

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions