bug caused by lazy malloc of tensor memory. while batchsize large than
1, output tensor is divided(shared memory), and each divided tensor do malloc.