Re:Makie, a modern and fast plotting library for Julia
by the_svd_doctor on Tuesday July 04, 2023 @03:46PM
(#36591699)
I mean, yes but also no. JAX will JIT on accelerators OOTB and has outstanding (imho) support for multi node. Julia does not have that OOTB AFAIK.
Now I agree that Julia, with its JIT, is well positioned to do all this. But I don’t think it does.
Re:Makie, a modern and fast plotting library for Julia
by ssivark on Tuesday July 04, 2023 @03:53PM
(#36591774)
Ah, I see. There’s XLA.jl for TPUs and a bunch of stuff from the JuliaGPU organization for GPUs — I wonder how they compare with TF/PyTorch/JAX.
I’m curious to get perspectives on what remains to be done to solve the missing piece, and what are the biggest challenges to overcome (assuming one gets access to the appropriate low level APIs on top of the hardware — like OpenAI’s Triton).
Re:Makie, a modern and fast plotting library for Julia
by patagurbon on Tuesday July 04, 2023 @05:05PM
(#36592501)
XLA.jl is essentially abandoned, no funding IIRC. Coil.jl and maybe some future hypothetical work on MLIR.jl could bridge the gap pretty easily.
Re:Makie, a modern and fast plotting library for Julia
by nextos on Tuesday July 04, 2023 @05:02PM
(#36592468)
That's right, but at the same time Flux, the default choice is not mature at all compared to JAX. So that advantage is only on paper.
I would really prefer to use Julia, I actually dislike Python, but it's unrealistic to implement large architectures in Flux. It is buggy, lots of gradient calculations are unreliable.
See some discussion here for instance: https://kidger.site/thoughts/jax-vs-julia
What do you mean by that? AFAICT Julia is already somewhat like the best that Jax could ever hope to be — since there is language level support for JIT compilation, gradients, etc — given how well libraries compose without having to pick different incompatible subsets of the language.