[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>

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,
             )