diff --git a/xtuner/v1/ops/moe/npu/group_gemm.py b/xtuner/v1/ops/moe/npu/group_gemm.py
index f09ceaa5a..4ce113023 100644
--- a/xtuner/v1/ops/moe/npu/group_gemm.py
+++ b/xtuner/v1/ops/moe/npu/group_gemm.py
@@ -1,9 +1,8 @@
 import torch
+from mindspeed.core.fusions.grouped_matmul import Ops
 
 
 def npu_group_gemm(x: torch.Tensor, weights: torch.Tensor, split_sizes: torch.Tensor) -> torch.Tensor:
-    from mindspeed.core.fusions.grouped_matmul import Ops
-
     weights = weights.transpose(1, 2)
 
     out = Ops.gmm(x, weights, split_sizes, trans_b=False)