From d11da2403ce2ab00b0248872e95c8f1d8d0a4ac0 Mon Sep 17 00:00:00 2001
From: Liangsheng Yin <hnyls2002@gmail.com>
Date: Sat, 11 Apr 2026 00:12:10 -0700
Subject: [PATCH] Add hisparse staging + decode offload guards to
 is_fully_idle() (#22577)

---
 python/sglang/srt/managers/scheduler.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/python/sglang/srt/managers/scheduler.py b/python/sglang/srt/managers/scheduler.py
index 5493c0aa4..1ea27ba24 100644
--- a/python/sglang/srt/managers/scheduler.py
+++ b/python/sglang/srt/managers/scheduler.py
@@ -3013,6 +3013,12 @@ class Scheduler(
             if self.disaggregation_mode == DisaggregationMode.DECODE:
                 idle &= len(self.disagg_decode_prealloc_queue.queue) == 0
                 idle &= len(self.disagg_decode_transfer_queue.queue) == 0
+                if self.decode_offload_manager is not None:
+                    idle &= len(self.decode_offload_manager.ongoing_offload) == 0
+
+            # HiSparse: staging requests transitioning prefill -> decode
+            if self.enable_hisparse:
+                idle &= not self.hisparse_coordinator.has_ongoing_staging()
 
             # HiCache: in-flight async ops (GPU↔Host↔L3) must drain before
             # destructive operations like attach/detach/flush_cache.