colehaus

Hi, thanks for looking at it. I haven't looked into the issue at all since opening it. So the notebook is still the most up-to-date reflection of the situation.

Unless I'm misunderstanding you: There's a comment in the source describing the problem there which is that `betainc` doesn't have all grads defined: https://github.com/tensorflow/probability/issues/655#issuecomment-558236514.

Ah, that's good news! And that sounds like a reasonable plan. I probably won't be able to think about doing it for a few weeks though.

This approach works for my use case. Thanks! Feel free to close.

I turned on event logging and finally got some possibly useful info. Copilot.el is no longer proposing any completions in the buffer, but it looks like it is still communicating...

I import `Text.Pandoc` myself and wrap everything up with an `unsafeCompiler`: ```haskell -- Run pandoc in IO instead of purely so that we can work with org mode file includes...

FYI, I don't think the Travis failure has much to do with the changes here. The message is `Unknown value flushCrumbs` in `purescript-spec`.

One workaround is to add the jQuery `require` in whichever application relies on jQuery. For example: Main.ps ```purescript foreign import jQuery :: Foreign ... ``` Main.js ```js "use strict"; var...

Something like this seems to roughly work: ```py @quax.register(lax.scan_p) def _( *args: Union[quax.ArrayValue, ArrayLike], reverse: bool, length: int, jaxpr, num_consts: int, num_carry: int, linear, unroll: int = 1, _split_transpose: Optional[bool]...

And here's an implementation of `remat` that seems to work (important for use with scan: https://jax.readthedocs.io/en/latest/notebooks/autodiff_remat.html#practical-notes): ```python @quax.register(jax._src.ad_checkpoint.remat_p) def _(*args, jaxpr, prevent_cse, differentiated, policy): del prevent_cse, differentiated, policy # `jaxpr_as_fun`...