Created by: yxnal

创建LoDTensor数据结构：tensor place = fluid.CUDAPlace(0) tensor = fluid.LoDTensor() tensor.set([[1]], place) print(tensor) 输出： paddle.fluid.core.LoDTensor object at 0x7fdb480bc110 然后，在该数据地址：0x7fdb480bc110 上通过numba.cuda.cudadrv.devicearray.DeviceNDArray数据结构对tensor重新赋值。 最后，print(np.array(tensor)) 报错：

terminate called after throwing an instance of 'paddle::platform::EnforceNotMet'  what():cudaMemcpy failed in paddle::platform::GpuMemcpySync: an illegal memory access was encountered at [/paddle/paddle/fluid/platform/gpu_info.cc:184]

请问是什么问题导致的？

相关联的issue： https://github.com/PaddlePaddle/Paddle/issues/15221

代码片段：