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(