Enhance runtime memory check in CI (#15192)

test/registered/spec/eagle/test_eagle_infer_b.py

@@ -316,6 +316,12 @@ class TestEAGLEServerPageSize(TestEAGLEServerBasic):
         "--attention-backend=flashinfer",
     ]
 
+    @classmethod
+    def setUpClass(cls):
+        # Runtime check only supported for topk=1, and can help to find a leak.
+        with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1):
+            super().setUpClass()
+
 
 class TestEAGLEServerPageSizeTopk(TestEAGLEServerBasic):
     # default topk=8 and tokens=64

test/registered/spec/eagle/test_eagle_infer_beta.py

@@ -57,7 +57,9 @@ class TestEagleServerBase(CustomTestCase, MatchedStopMixin):
             *[str(i) for i in range(1, cls.max_running_requests + 1)],
         ]
         launch_args.extend(cls.other_launch_args)
-        with envs.SGLANG_ENABLE_SPEC_V2.override(True):
+        with envs.SGLANG_ENABLE_SPEC_V2.override(
+            True
+        ), envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1):
             cls.process = popen_launch_server(
                 cls.model,
                 cls.base_url,

test/srt/test_radix_attention.py

@@ -84,4 +84,5 @@ class TestRadixCacheNonOverlapLPM(TestRadixCacheFCFS):
 
 if __name__ == "__main__":
     envs.SGLANG_TEST_RETRACT.set(True)
+    envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.set(1)
     unittest.main()

test/srt/test_retract_decode.py

@@ -27,7 +27,9 @@ class TestRetractDecode(CustomTestCase):
         cls.model = DEFAULT_MODEL_NAME_FOR_TEST
         cls.base_url = DEFAULT_URL_FOR_TEST
         launch_args = ["--chunked-prefill-size", "128"] + cls.other_args
-        with envs.SGLANG_TEST_RETRACT.override(True):
+        with envs.SGLANG_TEST_RETRACT.override(
+            True
+        ), envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(1):
             cls.process = popen_launch_server(
                 cls.model,
                 cls.base_url,