Jonathan Brodrick
Jonathan Brodrick
Isn't @varchasgopalaswamy 's suggestion more akin to using a relative rather than an absolute tolerance so it will scale with `t1`, rearranging and renaming: `clip = tnext > (1 -...
Only the fourth option included dependency on `tprev` so we could leave that out. Both ULP and the rtol approach should scale with `t1` meaning that users could manage similar...
Not sure I understand the reason convert to integer. Could we just do this then: `tol = 100.0 * (jnp.nextafter(t1) - jnp.prevbefore(t1))` And keep your current approach? If so, happy...
jax version: 0.6.0 equinox version: 0.12.1 diffrax version: 0.7.0
Thanks, do you know why it depends on adjoint?
Thanks Patrick, I did some digging around and noticed there is some [commentary](https://github.com/jax-ml/jax/blob/35e2657be8308917c7fa407be5a0b53192134890/jax/_src/core.py#L268) about this in `jax._src.core.jaxpr_as_fun`. I'm guessing there's no clever way we could use `jaxpr_as_fun` with the recursive...
The only thing I'm not sure about is how to enforce that `cpu/gpu` is selected, it might be better to have `jax` in core dependencies and have the `cpu` requirements...
What I mean is should it throw an error when we run install `pip install adept` or should this just default to be the equivalent to `pip install adept[cpu]`. I...
All good, feel free to hit squash and merge 🙂
Actually, need to update the file to match your new requirements will see if I have time tonight but feel free to do it on your side if you like.