Adrian Seyboldt

> There is no fusion when inputs broadcast together. True, I didn't think that through... > Edit: I like the speedups, maybe it's an okay trade-off? Yeah, you gotta make...

> Ah just something dumb: output_bc_patterns = tuple([out.type.broadcastable for out in node.inputs]) Let me guess, AI autocomplete? :-)

Yes, `overload` calls njit on the returned function, and uses `jit_options` for that njit call.

I'm also pretty confused... There was [this](https://github.com/PyO3/maturin/pull/1965) PR in maturin recently (and released in 1.5 2 weeks ago), which changed metadata in source distributions. Maybe that is somehow related? But...

I'll try to upload it locally, maybe that gives me some more info.

I uploaded the artifacts manually, so the release is now available. With a bit of luck this will be fixed somewhere when we make the next release...

@jessegrabowski I think we could just put that code in `linker/numba/dispatch/linalg.py`. We don't need to override the scipy functions (that would change the behavior of completely unrelated code, just because...

erfcx works for me, that was added here: #46

logsumexp is currently build from other parts:  Maybe we could improve performance of this further, but for now I think this is fine.

Would be great to have a second implementation of this! > In a first window we sample using the identity as mass matrix and adapt the step size using the...