diff --git a/python/sglang/srt/layers/rotary_embedding.py b/python/sglang/srt/layers/rotary_embedding.py
index 48f564f83..d48dc1b88 100644
--- a/python/sglang/srt/layers/rotary_embedding.py
+++ b/python/sglang/srt/layers/rotary_embedding.py
@@ -312,10 +312,20 @@ class RotaryEmbedding(CustomOp):
         query: torch.Tensor,
         key: torch.Tensor,
         offsets: Optional[torch.Tensor] = None,
+        fused_set_kv_buffer_arg: Optional[FusedSetKVBufferArg] = None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
-        # TODO: make a wrapper, and XPU will implement this kernel later.
-        self.cos_sin_cache = self.cos_sin_cache.to(query.device)
-        return self.forward_native(positions, query, key, offsets)
+        assert (
+            fused_set_kv_buffer_arg is None
+        ), "fused_set_kv_buffer_arg is not supported for xpu implementation"
+        positions = torch.add(positions, offsets) if offsets is not None else positions
+        return torch.ops.sgl_kernel.rotary_embedding(
+            positions,
+            query,
+            key,
+            self.head_size,
+            self.cos_sin_cache,
+            self.is_neox_style,
+        )
 
 
 class LinearScalingRotaryEmbedding(RotaryEmbedding):