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

There is a number token and request level memory accounting issues reported and fixed by the community recently. It would be good to complement them with robust CI tests.

Context:

1. bugfix[schedule]: Excessive preemption occurs when preempting running requests to schedule new prefill requests. #12494
2. Fix wrong running_bs in priority scheduling #13142
3. Fix bug: Incorrect variable used in rem_total_token_offset calculatio… #13201; [Bug] Incorrect variable used in rem_total_token_offset calculation during preemption (line 700) #13111 (TestRemTotalTokenOffsetBug)

4. sglang/python/sglang/srt/managers/schedule_policy.py

   Lines 666 to 668 in 5ef7287

   	valid_running_reqs = (
   	r for r in self.running_batch.reqs if r not in self.preempt_list
   	)

   - cover this case

Related resources

No response