Gate the idle-loop tree-cache sanity check behind a default-off env (#36205)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: hzh0425 <hzh0425@apache.org>

python/sglang/srt/environ.py

@@ -37,6 +37,11 @@ def _default_cache_subdir(name: str) -> str:
     return os.path.join(os.path.expanduser(envs.SGLANG_CACHE_DIR.get()), name)
 
 
+def _default_tree_cache_sanity_check() -> bool:
+    """Enable the expensive tree-cache sanity check by default in CI."""
+    return envs.SGLANG_IS_IN_CI.get()
+
+
 class EnvField:
     _allow_set_name = True
 
@@ -464,6 +469,9 @@ class Envs:
     SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK = EnvBool(True)
     SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY = EnvInt(0)
     SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_IDLE = EnvBool(True)
+    # The explicit environment variable still takes precedence over this CI
+    # default, so production remains opt-in and CI remains opt-out if needed.
+    SGLANG_ENABLE_TREE_CACHE_SANITY_CHECK = EnvBool(_default_tree_cache_sanity_check)
     # Physical KV-page checks: committed<=allocated + no page alias.
     SGLANG_CHECK_KV_PAGE_INVARIANTS = EnvBool(False)
     SGLANG_TBO_DEBUG = EnvBool(False)

python/sglang/srt/managers/scheduler_components/invariant_checker.py

@@ -462,6 +462,8 @@ class SchedulerInvariantChecker:
         return has_leak, messages
 
     def _check_tree_cache(self):
+        if not envs.SGLANG_ENABLE_TREE_CACHE_SANITY_CHECK.get():
+            return
         if (
             self.tree_cache.is_tree_cache()
             and (self.is_hybrid_swa and self.tree_cache.supports_swa())