From aeece10c061b8ef708b1962c175f5600f05c1933 Mon Sep 17 00:00:00 2001 From: Huanxing Date: Wed, 12 Aug 2026 16:23:39 +0800 Subject: [PATCH] [Bugfix][XPU] Run GDN attention as eager break under breakable cudagraph (#51928) Signed-off-by: Huanxing Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> --- vllm/_xpu_ops.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/vllm/_xpu_ops.py b/vllm/_xpu_ops.py index 288dbdae9b..aea1785897 100644 --- a/vllm/_xpu_ops.py +++ b/vllm/_xpu_ops.py @@ -7,6 +7,7 @@ from typing import TYPE_CHECKING import torch from vllm_xpu_kernels.flash_attn_interface import flash_attn_varlen_func +from vllm.compilation.breakable_cudagraph import eager_break_during_capture from vllm.logger import init_logger from vllm.platforms import current_platform from vllm.triton_utils import tl, triton @@ -1257,7 +1258,7 @@ class xpu_ops: direct_register_custom_op( op_name="gdn_attention_core_xpu", - op_func=_gdn_attention_core_xpu_impl, + op_func=eager_break_during_capture(_gdn_attention_core_xpu_impl), mutates_args=["core_attn_out", "z"], fake_impl=_gdn_attention_core_xpu_fake, )