Checklist

• I searched related issues but found no solution.
• The bug persists in the latest version.
• Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
• If this is not a bug report 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.

Describe the bug

When launching the model server with extra_buffer enabled and running tests for structured output, a slow but steady increase in VRAM consumption occurs, eventually causing OOM errors and model restarts in our production environment (every few hours). No leak is observed when extra_buffer is disabled, nor when it is enabled but JSON responses are not requested.

Reproduction

mamba_scheduler_strategy=extra_buffer
and structured JSON response generation

Environment

main