From 9416ee60ddad97c921376dbbb72108cca579014f Mon Sep 17 00:00:00 2001
From: Liangsheng Yin <hnyls2002@gmail.com>
Date: Fri, 31 Oct 2025 10:56:05 +0800
Subject: [PATCH] Reserved abortion API when retracting (#12425)

---
 python/sglang/srt/managers/io_struct.py         | 2 +-
 python/sglang/srt/managers/schedule_batch.py    | 6 ++++--
 python/sglang/srt/managers/scheduler.py         | 9 ++++++++-
 python/sglang/srt/managers/tokenizer_manager.py | 2 +-
 4 files changed, 14 insertions(+), 5 deletions(-)

diff --git a/python/sglang/srt/managers/io_struct.py b/python/sglang/srt/managers/io_struct.py
index fbee33d9b..3e740621f 100644
--- a/python/sglang/srt/managers/io_struct.py
+++ b/python/sglang/srt/managers/io_struct.py
@@ -1230,7 +1230,7 @@ class AbortReq(BaseReq):
     abort_all: bool = False
     # The finished reason data
     finished_reason: Optional[Dict[str, Any]] = None
-    abort_reason: Optional[str] = None
+    abort_message: Optional[str] = None
 
     def __post_init__(self):
         # FIXME: This is a hack to keep the same with the old code
diff --git a/python/sglang/srt/managers/schedule_batch.py b/python/sglang/srt/managers/schedule_batch.py
index deb6c3930..6e3756e5c 100644
--- a/python/sglang/srt/managers/schedule_batch.py
+++ b/python/sglang/srt/managers/schedule_batch.py
@@ -1445,7 +1445,9 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         evict_from_tree_cache(self.tree_cache, num_tokens)
         return self._is_available_size_sufficient(num_tokens)
 
-    def retract_decode(self, server_args: ServerArgs):
+    def retract_decode(
+        self, server_args: ServerArgs
+    ) -> Tuple[List[Req], float, List[Req]]:
         """Retract the decoding requests when there is not enough memory."""
         sorted_indices = list(range(len(self.reqs)))
 
@@ -1513,7 +1515,7 @@ class ScheduleBatch(ScheduleBatchDisaggregationDecodeMixin):
         )  # avoid zero division
         new_estimate_ratio = min(1.0, new_estimate_ratio)
 
-        return retracted_reqs, new_estimate_ratio
+        return retracted_reqs, new_estimate_ratio, []
 
     def release_req(self, idx: int, remaing_req_count: int, server_args: ServerArgs):
         req = self.reqs[idx]
diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 39777a2ae..58fa79859 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -1902,9 +1902,16 @@ class Scheduler(
             TEST_RETRACT and self.forward_ct % TEST_RETRACT_INTERVAL == 0
         ):
             old_ratio = self.new_token_ratio
-            retracted_reqs, new_token_ratio = batch.retract_decode(self.server_args)
+            retracted_reqs, new_token_ratio, reqs_to_abort = batch.retract_decode(
+                self.server_args
+            )
             self.num_retracted_reqs = len(retracted_reqs)
             self.new_token_ratio = new_token_ratio
+            for req in reqs_to_abort:
+                abort_reason: FINISH_ABORT = req.to_finish
+                self.send_to_tokenizer.send_output(
+                    AbortReq(abort_message=abort_reason.message, rid=req.rid), req
+                )
 
             logger.info(
                 "KV cache pool is full. Retract requests. "
diff --git a/python/sglang/srt/managers/tokenizer_manager.py b/python/sglang/srt/managers/tokenizer_manager.py
index d5d6b2ef1..0f2e4c78e 100644
--- a/python/sglang/srt/managers/tokenizer_manager.py
+++ b/python/sglang/srt/managers/tokenizer_manager.py
@@ -1781,7 +1781,7 @@ class TokenizerManager(TokenizerCommunicatorMixin):
         state = self.rid_to_state[recv_obj.rid]
         state.finished = True
 
-        abort_message = recv_obj.abort_reason or "Abort in waiting queue"
+        abort_message = recv_obj.abort_message or "Abort in waiting queue"
         finish_reason = {
             "type": "abort",
             "message": abort_message,