diff --git a/vllm/models/deepseek_v32/attention.py b/vllm/models/deepseek_v32/attention.py
index e6fb00434a..f7d0e76e54 100644
--- a/vllm/models/deepseek_v32/attention.py
+++ b/vllm/models/deepseek_v32/attention.py
@@ -495,6 +495,7 @@ class DeepseekV32Attention(MLAAttention):
                     self._vllm_config.parallel_config.cp_kv_cache_interleave_size
                 ),
                 skip_topk_buffer_clear=True,
+                topk_backend=self.indexer.indexer_op.topk_backend,
             )
         self.impl.record_logical_topk_ready()  # type: ignore[attr-defined]