Patrick Kidger
Patrick Kidger
Hmm, I think I'm inclined not to do this just to avoid complicating what is currently a nicely simple build system. I definitely see the value, I just don't think...
Does this seem to be an OOM during runtime or an OOM during compilation? If the latter then it might be due to closing over a very large constant. If...
This is Hyrum's law at its finest 😄 So I wonder if the correct choice here would be for you to override the metaclass instead? As it is you'll still...
I think probably not I'm afraid – that this niche enough that it's better for users to provide their own implementation. :)
Guesing you mean the metaclass - in this case then `type(eqx.Module)` might be what you're after?
So I think this is expected. What is going on here is that `cauchy_termination=False` corresponds to a preference for failing a solve, rather than performing an expensive solve. In particular...
Hey there! The main reason it's not implemented is the work required to review (not just to create) a PR for adding support; plus the maintenance cost. So I'm afraid...
Take a look at `eqx.nn.Lambda`, which is a wrapper that exists for this purpose. I decided against inspecting signatures as 'too much magic'. I no longer recall why there is...
Thanks for the report! I've just pushed a commit, so this should be fixed on the latest HEAD.
Hi @littlebaker! First of all, your description of how things work at the moment is absolutely spot-on. Indeed `RecursiveCheckpointAdjoint` only records the values at the end of each numerical step,...