diff --git a/vllm/patches/deepseek_v4_attention.py b/vllm/patches/deepseek_v4_attention.py
index 8286a288..252d9234 100644
--- a/vllm/patches/deepseek_v4_attention.py
+++ b/vllm/patches/deepseek_v4_attention.py
@@ -14,7 +14,12 @@ import torch.nn.functional as F
 from transformers import DeepseekV2Config, DeepseekV3Config
 
 import vllm.envs as envs
-from vllm.compilation.breakable_cudagraph import eager_break_during_capture
+try:
+    from vllm.compilation.breakable_cudagraph import eager_break_during_capture
+except ImportError:
+    # Older vLLM versions don't have this module; use identity decorator
+    def eager_break_during_capture(fn):
+        return fn
 from vllm.model_executor.layers.linear import (
     ReplicatedLinear,
 )