Hugging Face 模型本地推理与在线 Demo 文本输出的大小写的差异

想问一下加载从 Hugging Face 下载下来的模型进行本地推理时，会出现以上的问题。这种问题大吗？我看模型还是能够运行，就是感觉输出有些问题。

描述本地版本推理生成的文本回复，貌似无法大写，并且与 Hugging Face Demo 上的结果不一致。这是为什么？

问题:
一个 benchmark 中的例子

Observe the endoscopy image focusing on the area within the box. What is the most likely abnormality depicted in this section?
A: polyp
B: high-grade dysplasia
C: gastrointestinal tract cancer
D: barrett's esophagus

Explaining your answer with upper case

本地版回复：
a: polyp

在线 Demo 版回复：
A: POLYP

No source file, test, or inference entry point is named. Start by reproducing the benchmark prompt with the downloaded model and comparing the local inference path with the Hugging Face Demo; done means the casing and text differences are explained and, if confirmed as a defect, the affected behavior is aligned.

Written by the indexing model from the issue text.