Owen L

> I'm not sure. The way the controls are called at the moment is with the t, not the step index. We'd also have to have a way to pass...

> Awesome, I'm looking forward to it! Let's see if we can get the stateful controls done at the same time? I'd like to contain the breaking changes to a...

@pierreguilmin feel free to try out the draft I have currently: https://github.com/patrick-kidger/diffrax/pull/559. This (under some situations) allow a 100x speedup over VBT. For small problems with very small numbers of...

Sounds good, let me know if you encounter any issues. I will try to get it into a more final state by EoW so Patrick can start reviewing (though I...

> Adaptive stepping, I think :p At least the Rößler paper shows that for 1D noise you can do weak adaptive stepping (and it isn't proven to work on anything...

Also looking at the just while loops I see the same thing (i.e. checkpoint being faster) Code + Results ```python import os os.environ["XLA_FLAGS"] = "--xla_cpu_use_thunk_runtime=false" import equinox as eqx import...

Do you have MVC for the different steps? I will say vmap of double while loops (like with adaptive stepping) I have encountered some interesting edge cases with before, but...

0.5.1 is pre the events refactor right?

This hijacking of solver state was also how we had the stateful control initially. The PR does make the terms stateful (now having init/passing state), but currently only for the...

I think stateful terms are definitely a necessary condition, but I'm not sure they are sufficient (in the ideal case). Terms can manage their own state and adding the vf...