François Chollet
François Chollet
So what shall we do with this PR?
Are you doing ok with the test fixes across the different backends? I see that TF is passing at least, that's good!
> For Torch backend I am not doing test fixes for complex dtype as torch doesn't support it. Then you should explicitly exclude torch from these unit tests via a...
Thanks for the update. > This branch has conflicts that must be resolved Can you please solve the conflicts?
> Its installed from conda-forge, but the test fails seems to pass :/ You mean installing with pip fails but installing with conda works? I can reproduce the issue on...
Not sure what's going on with conda, but my hunch is that the bug with torch 2.3 + compile + keras is real, not a build issue.
> How can i test that torch + compile is being invoked? The test does that, there's nothing in your environment that it depends on. You can make sure that...
If you need a SPMD API, I 100% recommend JAX. It's more mature and better optimized than DTensor. You can use it via the easy `keras.distribution` API with the JAX...
> @fchollet any thoughts on what to use as fallback - numpy or jax. Sorry, just checking this now. I would say that JAX is preferable as fallback since numpy...
It is certainly feasible. It would need to get added not just in `PyDataset`, but also in `DataAdapter` and then called in `EpochIterator` as well as all `EpochIterator` subclasses that...