From 9edb0e0da341a18707b4422957cc334b12116aac Mon Sep 17 00:00:00 2001
From: Zilin Zhu <zhuzilinallen@gmail.com>
Date: Sun, 16 Nov 2025 22:02:22 +0800
Subject: [PATCH] Add SGLANG_ENABLE_REQ_POOL_LEAK_STRICT_CHECK to bypass mem
 leak check (#13339)

---
 python/sglang/srt/environ.py                  |  1 +
 .../scheduler_runtime_checker_mixin.py        | 21 ++++++++++++++++---
 python/sglang/srt/utils/common.py             | 10 +++++++++
 3 files changed, 29 insertions(+), 3 deletions(-)

diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index 86936bccb..ee5b42349 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -136,6 +136,7 @@ class Envs:
     SGLANG_LOG_GC = EnvBool(False)
     SGLANG_LOG_FORWARD_ITERS = EnvBool(False)
     SGLANG_DISABLE_REQUEST_LOGGING = EnvBool(False)
+    SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE = EnvBool(True)
 
     # Test & Debug
     SGLANG_IS_IN_CI = EnvBool(False)
diff --git a/python/sglang/srt/managers/scheduler_runtime_checker_mixin.py b/python/sglang/srt/managers/scheduler_runtime_checker_mixin.py
index e6f59e5b0..5870a13cd 100644
--- a/python/sglang/srt/managers/scheduler_runtime_checker_mixin.py
+++ b/python/sglang/srt/managers/scheduler_runtime_checker_mixin.py
@@ -7,10 +7,15 @@ import time
 from typing import TYPE_CHECKING
 
 from sglang.srt.disaggregation.utils import DisaggregationMode
+from sglang.srt.environ import envs
 from sglang.srt.managers.schedule_batch import ScheduleBatch
 from sglang.srt.mem_cache.mamba_radix_cache import MambaRadixCache
 from sglang.srt.mem_cache.swa_radix_cache import SWARadixCache
-from sglang.srt.utils.common import disable_request_logging, pyspy_dump_schedulers
+from sglang.srt.utils.common import (
+    disable_request_logging,
+    pyspy_dump_schedulers,
+    raise_error_or_warn,
+)
 
 if TYPE_CHECKING:
     from sglang.srt.managers.scheduler import Scheduler
@@ -138,7 +143,12 @@ class SchedulerRuntimeCheckerMixin:
                 f"available_size={len(self.req_to_token_pool.free_slots)}, "
                 f"total_size={self.req_to_token_pool.size}\n"
             )
-            raise ValueError(msg)
+            raise_error_or_warn(
+                self,
+                envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE,
+                "count_req_pool_leak_warnings",
+                msg,
+            )
 
     def check_memory(self: Scheduler):
         if self.is_hybrid:
@@ -150,7 +160,12 @@ class SchedulerRuntimeCheckerMixin:
 
         if memory_leak:
             msg = "token_to_kv_pool_allocator memory leak detected! " f"{token_msg}"
-            raise ValueError(msg)
+            raise_error_or_warn(
+                self,
+                envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE,
+                "count_memory_leak_warnings",
+                msg,
+            )
 
         self._check_req_pool()
 
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index babd57888..f181dafa5 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -3646,3 +3646,13 @@ def get_current_device_stream_fast():
     if cached_device_index == -1:
         cached_device_index = torch.get_device_module().current_device()
     return torch.get_device_module().current_stream(cached_device_index)
+
+
+def raise_error_or_warn(obj, strict, counter_name, message, log_interval=1000):
+    if strict:
+        raise ValueError(message)
+    else:
+        count = getattr(obj, counter_name, 0)
+        if count % log_interval == 0:
+            logger.warning(message)
+        setattr(obj, counter_name, count + 1)