Add type-stability checks via `DispatchDoctor.jl`

LuxDL / Lux.jl

Elegant & Performant Scientific Machine Learning in Julia

https://lux.csail.mit.edu/

MIT License

446 stars 50 forks source link

Open avik-pal opened 1 month ago

avik-pal commented 1 month ago

We can't apply it to the entire Lux module. Some of the code exists for debugging and is intentionally type unstable.
I like the default of erroring, but that will be too breaking. Maybe we can consider that for v0.6. We will warn if the code is type unstable for now.

avik-pal commented 4 days ago

We can do it at multiple levels:

[ ] apply: this is the most aggressive one https://github.com/LuxDL/LuxCore.jl/pull/28
[x] In the LuxLib _impl functions: It makes sense to guarantee these ones to be type stable https://github.com/LuxDL/LuxLib.jl/pull/79
[ ] In Lux applychain, applyparallel, ... functions which get hit pretty commonly.
[ ] function to set all the preferences