记录pytorch跑模型时遇到的奇怪bug

def rotate_half(x):
    x = rearrange(x, "... (j d) -> ... j d", j=2)
    x = x.contiguous()
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1).cuda()

以上代码在GPU运行时，报如下错误：
RuntimeError: transform: failed to synchronize: cudaErrorIllegalAddress: an illegal memory access was encountered

但是将模型放到CPU运行时，能正常运行，试了网上各种方法都不行，最后将这部分内容放CPU计算就可以了

def rotate_half(x):
    x = x.to("cpu")
    x = rearrange(x, "... (j d) -> ... j d", j=2)
    x = x.contiguous()
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1).cuda()