| 123456789101112131415161718 |
- diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
- index b57c93299a..c0b0571c8a 100644
- --- a/python/sglang/srt/managers/scheduler.py
- +++ b/python/sglang/srt/managers/scheduler.py
- @@ -3944,7 +3944,12 @@ class Scheduler(
- candidate_beam_width = (
- req.beam_group.beam_width if req.beam_group is not None else None
- )
- - if len(adder.can_run_list) >= self.get_num_allocatable_reqs(
- + # mimo26: a continuing chunked request already holds its req slot; only
- + # requests that still need one count against the free slots. Counting it
- + # set batch_is_full with a slot free, and that flag stays set until a
- + # request finishes, so the next waiting request sat out the whole decode.
- + new_slot_reqs = sum(1 for r in adder.can_run_list if not r.kv.holds_kv)
- + if new_slot_reqs >= self.get_num_allocatable_reqs(
- running_bs,
- candidate_beam_width,
- running_batch=running_batch,
|
