fix(mini-lb): forward the flush_cache timeout param to workers (#36150)

sgl-model-gateway/bindings/python/src/sglang_router/mini_lb.py

@@ -258,12 +258,18 @@ async def health_generate():
 
 
 @app.post("/flush_cache")
-async def flush_cache():
+async def flush_cache(timeout: Optional[float] = None):
+    # `timeout` must reach the workers. The scheduler treats a missing or
+    # non-positive timeout as "flush now, skip the idle check", so dropping it
+    # here frees KV buffers while a PD KV transfer is still reading them: the
+    # transfer then fails for real and the peer session gets blacklisted.
+    # Forwarding it keeps the scheduler on its deferred, drain-first path.
+    params = None if timeout is None else {"timeout": timeout}
     async with aiohttp.ClientSession() as session:
         # Create the tasks
         tasks = []
         for server in chain(lb.prefill_urls, lb.decode_urls):
-            tasks.append(session.post(f"{server}/flush_cache"))
+            tasks.append(session.post(f"{server}/flush_cache", params=params))
         for i, response in enumerate(asyncio.as_completed(tasks)):
             await response
     return Response(status_code=200)