follow dev70_2, referring to PyQCU/examples/pyqcu/conftest.*-v2026050*.log, similar to dev70_2, the multigrid part is done, the single iteration time in torch is about 0.074112 s (16*16*16*32, complex64), the single iteration time in cuda cpp is about 0.00487 s (16*16*16*32, complex64), the single iteration time in torch (with cuda qcu) is about 0.005856 s (16*16*16*32, complex64), however, the multgrid here, like bistabcg in dev70_2, suffers from an abnormal number of iterations. The initial hypothesis is that this is because the CUDA C++ version uses the SU(3) feature of the gauge (PyQCU/cpp/cuda/qcu/include/define.h:#define get_u(U, tmp_U, lat_xyzt)), while the Torch version does not. Each time the calculation results are compared, there is always a relative error of approximately one part per million.