diff --git a/vllm/v1/attention/backends/rocm_aiter_fa.py b/vllm/v1/attention/backends/rocm_aiter_fa.py
index a4cdce3f5d..14c473a70a 100644
--- a/vllm/v1/attention/backends/rocm_aiter_fa.py
+++ b/vllm/v1/attention/backends/rocm_aiter_fa.py
@@ -844,9 +844,13 @@ class AiterFlashAttentionImpl(AttentionImpl):
         assert self.num_heads % self.num_kv_heads == 0
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
 
-        if attn_type not in [AttentionType.DECODER, AttentionType.ENCODER_DECODER]:
+        if attn_type != AttentionType.DECODER:
             raise NotImplementedError(
-                "Encoder self-attention is not implemented for AiterFlashAttentionImpl"
+                "Only decoder self-attention is supported for "
+                "AiterFlashAttentionImpl. ENCODER_DECODER is not supported "
+                "because the prefill path uses cu_seqlens_k set to decoder "
+                "query_start_loc with causal=True, which is incorrect for "
+                "cross-attention."
             )
 
     def extend_for_sliding_window(