diff --git a/python/sglang/srt/lora/marlin_lora_temp/moe_runner.py b/python/sglang/srt/lora/marlin_lora_temp/moe_runner.py
index 57e4148f8..3cdac75f4 100644
--- a/python/sglang/srt/lora/marlin_lora_temp/moe_runner.py
+++ b/python/sglang/srt/lora/marlin_lora_temp/moe_runner.py
@@ -24,6 +24,9 @@ _is_cuda = is_cuda()
 if _is_cuda:
     from sgl_kernel import silu_and_mul
 
+    from sglang.kernels.ops.moe.fused_moe_triton_kernels import (
+        moe_sum_reduce_triton,
+    )
     from sglang.kernels.ops.moe.moe_wna16_marlin import moe_wna16_marlin_gemm
     from sglang.kernels.ops.moe.trtllm_lora_temp.virtual_experts import (
         _align_block_size_jit as moe_align_block_size,
@@ -31,9 +34,6 @@ if _is_cuda:
     from sglang.srt.layers.moe.fused_moe_triton.fused_marlin_moe import (
         get_scalar_type,
     )
-    from sglang.srt.layers.moe.moe_runner.triton_utils.fused_moe_triton_kernels import (
-        moe_sum_reduce_triton,
-    )
     from sglang.srt.layers.quantization.marlin_utils import marlin_make_workspace
     from sglang.srt.lora.marlin_lora_temp.activation import silu_and_mul_add_delta
     from sglang.srt.lora.marlin_lora_temp.direct_decode import (