[Bug] corrupted size vs. prev_size in fastbins

[tuilakhanh]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.

Describe the bug

Run Qwen3.5 with turbomind engine.

After several time, lmdeploy crash with logs:

corrupted size vs. prev_size in fastbins
Aborted (core dumped)

Reproduction

It is currently unknown.

Environment

-

Error traceback

[lzhangzz]

Having trouble reproducing the issue.

What setting you are testing? e.g. model / GPU used, prompt length and any parameter setting that is different from the default.

[lzhangzz]

@tuilakhanh Also, you may pass --async 0 to see if disabling the async model execution helps.

[tuilakhanh]

@lzhangzz I haven't encountered this error lately, however, I am having an issue with dense models like Qwen-27B. In some cases, the model's response gets stuck in a loop like this."