diff --git a/python/sglang/srt/managers/scheduler_update_weights_mixin.py b/python/sglang/srt/managers/scheduler_update_weights_mixin.py index 9bed7030d..fafdd48f4 100644 --- a/python/sglang/srt/managers/scheduler_update_weights_mixin.py +++ b/python/sglang/srt/managers/scheduler_update_weights_mixin.py @@ -132,7 +132,7 @@ class SchedulerUpdateWeightsMixin: if GPU_MEMORY_TYPE_CUDA_GRAPH in tags: self.memory_saver_adapter.pause(GPU_MEMORY_TYPE_CUDA_GRAPH) - torch.cuda.synchronize() + torch.get_device_module().synchronize() return ReleaseMemoryOccupationReqOutput()