diff --git a/atom/plugin/vllm/attention_backend/attention_gdn.py b/atom/plugin/vllm/attention_backend/attention_gdn.py
index 619964abf..bc1230665 100644
--- a/atom/plugin/vllm/attention_backend/attention_gdn.py
+++ b/atom/plugin/vllm/attention_backend/attention_gdn.py
@@ -358,12 +358,16 @@ def forward(
 
         # 2.2: Process the remaining part
         if attn_metadata.num_prefills > 0:
+            from aiter.ops.triton.gated_delta_net.gated_delta_rule import (
+                chunk_gated_delta_rule_opt_vk,
+            )
+
             initial_state = ssm_state[non_spec_state_indices_tensor].contiguous()
             initial_state[~has_initial_state, ...] = 0
             (
                 core_attn_out_non_spec,
                 last_recurrent_state,
-            ) = self.chunk_gated_delta_rule(
+            ) = chunk_gated_delta_rule_opt_vk(
                 q=query_non_spec,
                 k=key_non_spec,
                 v=value_non_spec,