diff --git a/python/sglang/kernels/ops/attention/dsv4/elementwise.py b/python/sglang/kernels/ops/attention/dsv4/elementwise.py
index dac07d218..053eb01c6 100644
--- a/python/sglang/kernels/ops/attention/dsv4/elementwise.py
+++ b/python/sglang/kernels/ops/attention/dsv4/elementwise.py
@@ -17,6 +17,7 @@ _is_xpu = is_xpu()
 
 if _is_xpu:
     from sgl_kernel import fused_k_norm_rope_flashmla as fused_k_norm_rope_flashmla_xpu
+    from sgl_kernel import fused_q_norm_rope as fused_q_norm_rope_xpu
 
 
 @cache_once
@@ -150,8 +151,11 @@ def fused_q_norm_rope(
     freqs_real = torch.view_as_real(freqs_cis).flatten(-2)
     head_dim = q_input.shape[-1]
     rope_dim = freqs_real.shape[-1]
-    module = _jit_main_q_norm_rope_module(q_input.dtype, head_dim, rope_dim)
-    module.forward(q_input, q_output, freqs_real, positions, eps)
+    if _is_xpu:
+        fused_q_norm_rope_xpu(q_input, q_output, freqs_real, positions, eps)
+    else:
+        module = _jit_main_q_norm_rope_module(q_input.dtype, head_dim, rope_dim)
+        module.forward(q_input, q_output, freqs_real, positions, eps)
 
 
 def fused_q_indexer_rope_hadamard_quant(