Jimmy MA
Jimmy MA
share 本身就是随机数，没有很好的可读性支持，如果有需求可以参考[这个 issue](https://github.com/secretflow/spu/issues/458) 修改 SPU 源码打印 share 到 log 里看
Thx for the reply. So it will be released in torch-xla v2.2.1 ?
@fionser would you mind taking a look?
Hi @deevashwer , we have encountered a some similar [issue](https://github.com/secretflow/spu/issues/421) (it's in chinese) before due to a potential memory leak problem in glibc. Maybe you can have a try with...
see https://github.com/secretflow/spu/blob/main/libspu/spu.proto. Protobuf compiler will generate these files during building.
从你描述的场景来看似乎不需要用到 spu，就是分布式训练，只是每一方训练数据不是原始数据，而是秘密分享
你可以使用 secretflow，把其中涉及 mpc 协议的部分用 spu 处理
如果想最大程度自主控制秘密分享，肯定还是修改 spu 运行时的 c++ 代码二次开发最为直接
beaver server 的接口在 commit 7b45d886007d5dc2aff1f63952fa18c6b3da2cf7 之后发生了变更，运行互联互通的示例请尝试从 SPU 该 commit 及其之前的源码编译