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,