Patrick Kidger

I don't think that'll be enough. If that's used, and someone does: ``` pip install signatory==1.2.4.1.6.0 pip install torch==1.6.0 ``` Then the first `pip` will grab Signatory, look for PyTorch,...

So Signatory doesn't currently offer prebuilt binaries for PyTorch 1.10 on Windows I'm afraid. (I need to update the build scripts with every PyTorch release.) Your options are: - Switch...

Whoops, sorry for the radio silence on getting back to you on this one. Anyway, it looks like 1.2.5 didn't get created properly on PyPI, which is why things failed....

To start off by answering your question, yes: ```python fnet = net.__call__ params = net ``` which obviously looks a bit trivial. Any model is already a PyTree of parameters....

Yep, `jit-grad-vmap` computes gradients in the "normal" way e.g. like PyTorch.

Alright, LGTM! I'll merge this along with all the other ready-to-go PRs and do a new release shortly.

Hmm, I'd need to think about the cleanest abstraction (it feels a bit weird to me to pass a term into a solver at init time) but something like your...

So there's basically two main patterns here. 1. Never use `static_field`. Always filter things out using `eqx.filter_{jit, grad, ...}`. 2. Use `static_field` on every leaf that isn't a floating-point JAX...

So this is deliberately non-static because the forward pass of `nn.Dropout` is fine with having that as a tracer. In some sense the same is true of `nn.Linear.in_features` etc. as...

Sounds good!