Add Support for torch==1.5.x

[madhavajay]

Is there any chance torch 1.5.x can be supported?

Then we can achieve our current support matrix in Syft / Duet: https://github.com/OpenMined/pysyft of torch 1.5.x+ as we need csprng in SyMPC: https://github.com/OpenMined/SyMPC/

[pbelevich]

I don't think it's possible, because support of custom RNGs was added in PT 1.6.0. The only possible workaround for 1.5.0 is to have something like
torchcsprng.random(t) instead of t.random_(generator=csprng_rng)
torchcsprng.normal(t) instead of t.normal_(generator=csprng_rng)
...

[madhavajay]

@pbelevich Good idea, lets see if thats sufficient for our use case and then we can just wait out the deprecation of 1.5.x.

[pbelevich]

@madhavajay you closed this because you don't need it anymore?

[madhavajay]

@pbelevich Oh sorry, I thought you were suggesting that this only required changing the use of t.x to torchcsprng.x(t) but I guess to prevent torch version conflicts there might still need to be a change on torchcsprng?

[pbelevich]

@madhavajay how long do you think you will need to support 1.5.x? Or in other words when are you going to stop supporting 1.5.x?

[madhavajay]

@pbelevich good question. I would imagine we would probably deprecate it when 1.8.x rolls around but we currently had to re-add 1.4.0 for another use case the other day so its hard to say exactly when. Is there a technical reason the random number generator won't work or is it just that there is a lot of related work to support the older code base and different platforms / targets?

[turian]

@madhavajay https://pytorch.org/blog/pytorch-1.8-released/

[madhavajay]

Yes, we have stopped testing 1.5.x and below in our CI now so we can probably just close this now.