[CI] Use smaller models in CI

[b8zhong]

Checklist

• If this is not a feature request but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
• Please use English. Otherwise, it will be closed.

Motivation

If the model is the exact same architecture, we can use a smaller model to speed up CI? Currently some tests are taking a long time.

Also we can increase the BS for many accuracy related tests.

E.g, Gemma 27B -> 1B

We may also have to adjust accuracy related tests if this happens.

Related resources

No response

[ishandhanani]

Can also use qwen 0.6b. Only 1.2GB

[b8zhong]

@ishandhanani Yeah I agree, though actually I tried 0.6B for a recent PR, the problem it seems to be non instruct, and the GSM8K accy varies too much (i ended up just using https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)

[github-actions]

This issue has been automatically closed due to inactivity. Please feel free to reopen it if needed.