我使用0.7.2用Qwen2vl-7B做图文理解,vllm效果完全不如hf transformer;之前的0.6.5的时候没有发现过这个问题;但是0.6.5 Qwen2.5-14b的时候进行20000token的prompt文本输入,推理效果奇差,完全是乱码输出更新到0.7.2输出就正常了。 感觉vllm多模态模型和文本模型的实现上还有一些小bug Sign up for free to join ...
register_model("Qwen2VLForConditionalGeneration", Collaborator wangxiyuan Mar 5, 2025 • edited have you tested this? I'm not sure if override Qwen2VLForConditionalGeneration works. Looks it has some error: https://github.com/vllm-project/vllm-ascend/actions/runs/13656424628/job/...
New issue Closed Description MrRace jklj077 mentioned thison Mar 25, 2024 Sign up for freeto join this conversation on GitHub.Already have an account?Sign in to comment