diff --git a/python/sglang/srt/layers/attention/deepseek_v4_backend.py b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
index f85b3bfc6..402d3baa0 100644
--- a/python/sglang/srt/layers/attention/deepseek_v4_backend.py
+++ b/python/sglang/srt/layers/attention/deepseek_v4_backend.py
@@ -1780,7 +1780,10 @@ class DeepseekV4AttnBackend(
         indices. Chunk-invariant scaffolding lives in
         ``self.forward_metadata.sparse_prefill_cache``.
         """
-        from sgl_kernel.flash_mla import flash_mla_sparse_fwd
+        if _is_xpu:
+            from sgl_kernel import flash_mla_sparse_fwd
+        else:
+            from sgl_kernel.flash_mla import flash_mla_sparse_fwd
 
         # q is (b, 1, h_q, d_qk); flash_mla_sparse_fwd takes (s_q, h_q, d_qk).
         q_flat = q.squeeze(1)