From 6cbdfec45308759009511894cccd8654c893392b Mon Sep 17 00:00:00 2001 From: Kurkur <102506892+litmei@users.noreply.github.com> Date: Mon, 28 Sep 2026 09:32:16 +0800 Subject: [PATCH] [Fix][NPU] Fix performance degradation caused by serial execution of two-stage ACL op calls on graph (#40814) --- .../hardware_backend/npu/graph_runner/npu_cudagraph_backend.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/python/sglang/srt/hardware_backend/npu/graph_runner/npu_cudagraph_backend.py b/python/sglang/srt/hardware_backend/npu/graph_runner/npu_cudagraph_backend.py index 9767e9dafb..ef6fdcc839 100644 --- a/python/sglang/srt/hardware_backend/npu/graph_runner/npu_cudagraph_backend.py +++ b/python/sglang/srt/hardware_backend/npu/graph_runner/npu_cudagraph_backend.py @@ -178,8 +178,8 @@ class NPUCudaGraphBackend(BaseCudaGraphBackend): update_future = self._update_executor.submit( graph.update, cpu_update_input=cpu_update_input ) - update_future.result() graph.replay() + update_future.result() return self._outputs[shape_key] def cleanup(self) -> None: