Owen L

(just linking other relevant PDE issues: https://github.com/patrick-kidger/diffrax/issues/394, https://github.com/patrick-kidger/diffrax/issues/177)

I would say this behavior is expected (whether not it is wanted maybe another question). Since in general numpy arrays are not hashable, and making things static is to set...

Hmmm I see it yea I misread it, it's a int64 class from numpy. That would be an mis usage of the `is_array` then, because np.int64's are hashable. I can...

Having a clearer set of guidelines/tools for more complex auxiliary states is of interest to me as well. I will note that in the above MVC, I think if there...

My approach would be to define a filter spec for that set of parameters then just define the optimizer over that specific set. Similar to https://docs.kidger.site/equinox/tricks/#custom-per-parameter-behaviour I would first try...

Your points are generally true I think, I was just highlighting the general framework I would go about doing. 1. If you know what layers you want, I think it...

So in that case, it would be like ```python exclude = lambda x: isinstance(x, eqx.nn.LayerNorm) leaf = lambda x: hasattr(x, "weight") and not exclude(x) get_weights = lambda m: [x.weight for...

Awesome, I think I have 3 open PRs right now 😅 , so that should give some time to get things into a presentable shape (and think on solver representation...

Closing to open more specific issues as we work on getting these solvers in (e.g. https://github.com/patrick-kidger/diffrax/issues/481)

Exploring more with `ForwardMode` and seeing different errors depending on the solver (all of this feels unexpected to me, since it seems like a very simple modification to the reverse...