Dependency on jaxlib makes TFF 0.18+ impossible to install on Windows

[rstoki]

In commit e481167279f60ad7f3d773264b8c0afab113d5dc, the dependency on jaxlib was added. This dependency makes version 0.18 impossible to install on Windows machine, because jaxlib is not compiled/provided for Windows. The previous version of TFF 0.17 works fine also on Windows.

Is the jaxlib really essential for Tensor Flow Federated? Could the dependency be made "optional"?

It would be a pity to drop the Windows support altogether "only" because of employing some optimization tool.

[CoderTylor]

I have met the same issue, is there any solutions for it

[ammary-mo]

Yeah, I had the same problem today

[rutgerfick]

I'm having the same problem. Is there any solution?

[amitport]

I like tff and I am using it for research, but it should be a lot more clear on the website about the very low priority for supporting windows. Windows builds break very often and are just not tested by the core team (correct me if I'm wrong).

(note that in this case jax is being imported by tff even if not used, so this isn't just a dependency issue)

[Jasperhino]

I am having the same problem

[ayberkuckun]

Same here. Moreover, TFF 0.17 requires TF 2.3 and I can't use the versions of TF below 2.5 since I am using an RTX 30XX GPU and it requires CUDA 11.1 and above.

[michaelreneer]

@amitport you are correct we currently do not test on Windows.

@rstoki

It would be a pity to drop the Windows support altogether "only" because of employing some optimization tool.

TFF has never supported Windows so I don't think we have dropped anything. We did not support Windows before taking the dependency on jaxlib for a few reasons, so our current set of dependencies is really just one of many reasons.

[michaelreneer]

Maybe we can leave this issue open to track the feature request to support Windows as a platform?