I’m still not sold that we need a new framework in this space. A lot of these articles talk about the amazing features of Jax but not the sharp edges it has.
Frankly, if we were to change anything I would prefer to build on Candle, so if I need to drop down I can write Rust instead of C++. That would be a much bigger win than chasing after a pure functional dream land.
Just cursory looking at Jax, it does feel like >95% of the "everyone excited about Jax" people would get >95% of the performance win for 5% of the increased complexity by using Cython instead...
7 comments
[ 4.3 ms ] story [ 34.0 ms ] threadFrankly, if we were to change anything I would prefer to build on Candle, so if I need to drop down I can write Rust instead of C++. That would be a much bigger win than chasing after a pure functional dream land.
I used TF, Torch and Jax to implement a monte carlo sampling thingy and Jax was by far the most readable and performant.
I think the Cython use-case is kinda subsumed by the pypy jit right?