Hacker News new | ask | show | jobs
by BadInformatics 1957 days ago
Pretty disappointing to see it using XLA. Anything Tensorflow is notoriously difficult to build on a local machine and generally considered a dependency nightmare. I would've thought libtorch bindings would be much easier.
1 comments
josevalim 1957 days ago
Our compiler backends are pluggable. We went with XLA because that seemed the most accessible to us 3 months ago but you should be able to bring in libtorch or any other tensor compiler.

That's actually one of the things I am really looking forward to: see what other compilers people will integrate with. I am aware some of the neural network libs have pluggable compilers but it will be interesting to see it done at the tensor level. :)

link
BadInformatics 1957 days ago
That's great to hear. I'd love to see the Bazel monstrosity that is Tensorflow XLA get easier to build because of Nx. Not holding my breath though!
link
josevalim 1957 days ago
Yeah, I would love if we could depend only on XLA. I think other communities like Julia could benefit from it too.
link