From ff0b64e1e6b89b7517ebf65c7b253960113dbce7 Mon Sep 17 00:00:00 2001
From: fzyzcjy <5236035+fzyzcjy@users.noreply.github.com>
Date: Tue, 4 Nov 2025 18:01:27 +0800
Subject: [PATCH] Ensure GPU work is finished when release memory occupation
 call is finished (#12592)

---
 python/sglang/srt/managers/scheduler_update_weights_mixin.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/python/sglang/srt/managers/scheduler_update_weights_mixin.py b/python/sglang/srt/managers/scheduler_update_weights_mixin.py
index 1fde62ff1..9bed7030d 100644
--- a/python/sglang/srt/managers/scheduler_update_weights_mixin.py
+++ b/python/sglang/srt/managers/scheduler_update_weights_mixin.py
@@ -132,6 +132,8 @@ class SchedulerUpdateWeightsMixin:
         if GPU_MEMORY_TYPE_CUDA_GRAPH in tags:
             self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_CUDA_GRAPH)
 
+        torch.cuda.synchronize()
+
         return ReleaseMemoryOccupationReqOutput()
 
     def resume_memory_occupation(