[Bugfix][XPU] Run GDN attention as eager break under breakable cudagraph (#51928)
Signed-off-by: Huanxing <huanxing.shen@intel.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Huanxing committed
Aug 12, 2026 at 16:23 UTC
aeece10c061b8ef708b1962c175f5600f05c1933
1 file changed
+2
-1
vllm/_xpu_ops.py
+2
-1
@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING
7
import torch
8
from vllm_xpu_kernels.flash_attn_interface import flash_attn_varlen_func
9
10
+from vllm.compilation.breakable_cudagraph import eager_break_during_capture
11
from vllm.logger import init_logger
12
from vllm.platforms import current_platform
13
from vllm.triton_utils import tl, triton
@@ -1257,7 +1258,7 @@ class xpu_ops:
1258
1259
direct_register_custom_op(
1260
op_name="gdn_attention_core_xpu",
1260
- op_func=_gdn_attention_core_xpu_impl,
1261
+ op_func=eager_break_during_capture(_gdn_attention_core_xpu_impl),
1262
mutates_args=["core_attn_out", "z"],
1263
fake_impl=_gdn_attention_core_xpu_fake,
1264
)