diff --git a/python/sglang/srt/layers/attention/attention_registry.py b/python/sglang/srt/layers/attention/attention_registry.py
index 6eabb10b5..9020e5260 100644
--- a/python/sglang/srt/layers/attention/attention_registry.py
+++ b/python/sglang/srt/layers/attention/attention_registry.py
@@ -209,7 +209,8 @@ def attn_backend_wrapper(runner: "ModelRunner", full_attn_backend: "AttentionBac
                 assert (
                     runner.server_args.attention_backend == "triton"
                     or runner.server_args.attention_backend == "trtllm_mha"
-                ), "triton or trtllm_mha backend are the only supported backends on Blackwell GPUs for hybrid GDN models, use --attention-backend triton or --attention-backend trtllm_mha to specify the backend."
+                    or runner.server_args.attention_backend == "fa4"
+                ), "triton or trtllm_mha or fa4 backend are the only supported backends on Blackwell GPUs for hybrid GDN models, use --attention-backend triton or --attention-backend trtllm_mha to specify the backend."
             if is_npu():
                 assert (
                     runner.server_args.attention_backend == "ascend"
diff --git a/python/sglang/srt/layers/attention/flashattention_backend.py b/python/sglang/srt/layers/attention/flashattention_backend.py
index c0adc510e..6c5d67220 100644
--- a/python/sglang/srt/layers/attention/flashattention_backend.py
+++ b/python/sglang/srt/layers/attention/flashattention_backend.py
@@ -1136,6 +1136,14 @@ class FlashAttentionBackend(AttentionBackend):
         if sinks is not None:
             kwargs["sinks"] = sinks
 
+        flash_attn_with_kvcache_base = flash_attn_with_kvcache_fa3
+
+        flash_attn_with_kvcache = (
+            flash_attn_with_kvcache_fa4
+            if self.fa_impl_ver == 4
+            else flash_attn_with_kvcache_base
+        )
+
         k_descale, v_descale = None, None
         # only use kv scaling if: 1) fp8 kv is explicitly enabled, 2) RadixAttention
         # has corresponding quantization method so that layer.k_scale is not None,