Equinox has great idioms — really pioneered the Pytree perspective. Penzai is also great.
JAX feels close to achieving a sort of high-level GPGPU ecosystem. Super fledgling — but I keep finding more little libraries that build on JAX (and can be compositionally used with other JAX libraries because of it).
Only problem is that lots of compositional usage leads to big code and therefore big compile times for XLA.
Personally a big fan of Flax. The way it separates the params and the compute graph is - imo - the right (TM) way. Saying this after many years of soing ML :)
I've enjoyed using Equinox and Diffrax for performing ODE simulations. To my knowledge the only other peer library with similar capabilities is the Julia DifferentialEquations.jl package.
I wish Jax had everything I need to experiment with DL models built in, natively - like Pytorch. Instead there are many third party libraries (flax, trax, haiku, this one, etc). I have no idea which one to use. This was the case when I first played with jax 5 years ago, and it’s still the case today (even worse it seems). This makes it a non starter for me.
So, “do as I say, not as I do”? I just looked at Equinox and it does look a little better than flax, but flax seems to be more widely used. Both haiku and trax seem to be on the way out. All four were intended to do exactly the same thing. What a mess of an ecosystem.
But it crashes Python when I try to run a compiled function. And that's only after discovering I need an older specific version of jax-metal, because newer versions apparently don't work with M1 anymore (only M2/M3) -- they don't even report the GPU as existing. And even if you get it running, it's missing support for complex numbers.
I'm not clear whether it's Google or Apple who is building/maintaining support for Apple M chips though.
JAX works perfectly in CPU mode though on my MBA, so at least I can use it for development and debugging.
Pretty sure it's Apple building it, and they're using JAX in-house so I imagine it will get better over time. Though they do love to drop support for old things so maybe M1 will never work again...
I think they’re likely using MLX in house now, no? (Probably not everyone, ofc - but seems likely that many will just use the native array framework designed explicitly for MX chips)
JAX feels close to achieving a sort of high-level GPGPU ecosystem. Super fledgling — but I keep finding more little libraries that build on JAX (and can be compositionally used with other JAX libraries because of it).
Only problem is that lots of compositional usage leads to big code and therefore big compile times for XLA.
reply