From 2e6670739973533ac4bc2810613d4d4d3d564b1c Mon Sep 17 00:00:00 2001
From: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
Date: Fri, 10 Jul 2026 08:53:18 +0800
Subject: [PATCH] Fix pipeline-parallel abort missing in-flight requests in
 non-current microbatch slots (#29405)

Co-authored-by: burling <3637497+burling@users.noreply.github.com>
Co-authored-by: zhaotyer <89376832+zhaotyer@users.noreply.github.com>
---
 python/sglang/srt/managers/scheduler.py       |  9 +-
 .../scheduler/test_scripted_pp_abort.py       | 94 +++++++++++++++++++
 2 files changed, 99 insertions(+), 4 deletions(-)
 create mode 100644 test/manual/scheduler/test_scripted_pp_abort.py

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index a872dbbcb..c43c8c5eb 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -3966,12 +3966,13 @@ class Scheduler(
                 self.disagg_decode_prealloc_queue.retracted_queue = remaining_retracted
 
         # Delete requests in the running batch
-        if self.cur_batch is self.running_batch or self.cur_batch is None:
-            reqs = self.running_batch.reqs
+        if self.ps.pp_size == 1:
+            inflight_batches = [self.running_batch, self.cur_batch]
         else:
-            reqs = self.running_batch.reqs + self.cur_batch.reqs
+            inflight_batches = [*self.running_mbs, *self.mbs]
 
-        for req in reqs:
+        inflight_reqs = {r for b in inflight_batches if b is not None for r in b.reqs}
+        for req in inflight_reqs:
             if not req.finished() and (
                 recv_req.abort_all or req.rid.startswith(recv_req.rid)
             ):
diff --git a/test/manual/scheduler/test_scripted_pp_abort.py b/test/manual/scheduler/test_scripted_pp_abort.py
new file mode 100644
index 000000000..06ab4505a
--- /dev/null
+++ b/test/manual/scheduler/test_scripted_pp_abort.py
@@ -0,0 +1,94 @@
+import unittest
+
+from sglang.test.scripted_runtime.context import ScriptedContext
+from sglang.test.scripted_runtime.req_handle import ScriptedReqHandle
+from sglang.test.scripted_runtime.test_case import ScriptedTestCase
+from sglang.test.scripted_runtime_chunked_helpers import (
+    DEFAULT_CHUNK_SIZE,
+    base_engine_kwargs,
+)
+
+
+def _drain_until_released(t: ScriptedContext, *handles: ScriptedReqHandle):
+    for _ in range(12):
+        if all(
+            h.kv_pages == 0
+            and h.lock_refs == 0
+            and (h.req is None or h.req.req_pool_idx is None)
+            for h in handles
+        ):
+            return
+        yield
+
+
+def _advance_until_in_running_mbs(
+    t: ScriptedContext, *handles: ScriptedReqHandle, max_steps: int = 800
+):
+    rids = {h.rid for h in handles}
+    present: set[str] = set()
+    for _ in range(max_steps):
+        present = {
+            req.rid
+            for mb in t.scheduler.running_mbs
+            for req in mb.reqs
+            if req.rid in rids
+        }
+        if present == rids:
+            return
+        yield
+    raise AssertionError(
+        f"reqs did not all reach a running_mbs decode slot within {max_steps} "
+        f"steps; present={present!r} wanted={rids!r}"
+    )
+
+
+class TestAbortPPCrossSlot(ScriptedTestCase):
+    ENGINE_KWARGS = base_engine_kwargs(
+        chunked_prefill_size=DEFAULT_CHUNK_SIZE,
+        pp_size=4,
+        pp_max_micro_batch_size=1,
+    )
+
+    def test_abort_all_reaches_running_reqs_in_all_microbatch_slots(self):
+        """abort_all must abort running reqs in every PP microbatch slot, not just the current one (needs >2 slots)."""
+        self.server.execute_script(
+            self._script_abort_all_reaches_running_reqs_in_all_microbatch_slots
+        )
+
+    @staticmethod
+    def _script_abort_all_reaches_running_reqs_in_all_microbatch_slots(
+        t: ScriptedContext,
+    ):
+        reqs = [
+            t.start_req(
+                prompt_len=16,
+                max_new_tokens=512,
+                ignore_eos=True,
+                prompt_token=310 + i,
+            )
+            for i in range(4)
+        ]
+        yield from _advance_until_in_running_mbs(t, *reqs)
+
+        slot_of = {}
+        for slot_id, mb in enumerate(t.scheduler.running_mbs):
+            for req in mb.reqs:
+                slot_of[req.rid] = slot_id
+        slots = {slot_of[r.rid] for r in reqs}
+        assert len(slots) == len(reqs), (
+            f"setup invalid: reqs must each occupy a distinct mb slot to exercise "
+            f"the cross-slot abort scan; slot_of={slot_of!r}"
+        )
+
+        t.abort_all()
+        yield from _drain_until_released(t, *reqs)
+
+        alive = {r.rid: r.kv_pages for r in reqs if r.kv_pages != 0}
+        assert not alive, (
+            f"abort_all left running reqs alive in non-current mb slots (only the "
+            f"current slot + stale cur_batch were scanned): still_holding_kv={alive!r}"
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()