diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py
index e71c051a4..a99e230e2 100644
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -187,7 +187,9 @@ class AiterAttnBackend(AttentionBackend):
 
             # current persist a16w16 mla_decode kernel does not support head_num = 128
             # need to fall back to non-persist
-            if self.kv_cache_dtype is not fp8_dtype and self.enable_dp_attention:
+            # only use mla_ps_kernel when fp8 kv_cache
+            # for non-fp8 kv_cache, use non-persist kernel to avoid performance degradation
+            if self.kv_cache_dtype is not fp8_dtype:
                 _use_mla_ps_kernel = False
                 fast_mode = False
                 intra_batch_mode = False