Enzyme VS Zygote.jl

Automatic differentiation is done using https://enzyme.mit.edu/

lattner left google and was the primary reason they chose swift, so they lost interest.

if you're asking from an ML perspective, i believe the original motivation was to incorporate automatic differentiation in the swift compiler. i believe enzyme is the spiritual successor.

https://github.com/EnzymeAD/Enzyme

https://ispc.github.io/ispc.html

For the auto-differentiation when I need performance or memory, I currently use tapenade ( http://tapenade.inria.fr:8080/tapenade/index.jsp ) and/or manually written gradient when I need to fuse some kernel, but Enzyme ( https://enzyme.mit.edu/ ) is also very promising.

MPI for parallelization across machines.

For autodiff, we are currently working again on publishing a new Enzyme (https://enzyme.mit.edu) Frontend for Rust which can also handle pure Rust types, first version should be done in ~ a week.

Have you taken a look at enzymeAD? There is a group porting it to rust.

Enzyme dev here, so take everything I say as being a bit biased:

While, by design Enzyme is able to run very fast by operating within the compiler (see https://proceedings.neurips.cc/paper/2020/file/9332c513ef44b... for details) -- it aggressively prioritizes correctness. Of course that doesn't mean that there aren't bugs (we're only human and its a large codebase [https://github.com/EnzymeAD/Enzyme], especially if you're trying out newly-added features).

Notably, this is where the current rough edges for Julia users are -- Enzyme will throw an error saying it couldn't prove correctness, rather than running (there is a flag for "making a best guess, but that's off by default"). The exception to this is garbage collection, for which you can either run a static analysis, or stick to the "officially supported" subset of Julia that Enzyme specifies.

Incidentally, this is also where being a cross-language tool is really nice -- namely we can see edge cases/bug reports from any LLVM-based language (C/C++, Fortran, Swift, Rust, Python, Julia, etc). So far the biggest code we've handled (and verified correctness for) was O(1million) lines of LLVM from some C++ template hell.

I will also add that while I absolutely love (and will do everything I can to support) Enzyme being used throughout arbitrary Julia code: in addition to exposing a nice user-facing interface for custom rules in the Enzyme Julia bindings like Chris mentioned, some Julia-specific features (such as full garbage collection support) also need handling in Enzyme.jl, before Enzyme can be considered an "all Julia AD" framework. We are of course working on all of these things (and the more the merrier), but there's only a finite amount of time in the day. [^]

[^] Incidentally, this is in contrast to say C++/Fortran/Swift/etc, where Enzyme has much closer to whole-language coverage than Julia -- this isn't anything against GC/Julia/etc, but we just have things on our todo list.

Idk, Enzyme is pretty next gen, all the way down to LLVM code.

https://github.com/EnzymeAD/Enzyme

I'm working on merging my build-tool for (oxide)-enzyme into Enzyme itself. Also looking into improving the documentation.

that seems one of the points of enzyme[1], which was mentioned in the article.

[1] - https://enzyme.mit.edu/

being able in effect do interprocedural cross language analysis seems awesome.

If you look at Julia open source projects you'll see that the projects tend to have a lot more contributors than the Python counterparts, even over smaller time periods. A package for defining statistical distributions has had 202 contributors (https://github.com/JuliaStats/Distributions.jl), etc. Julia Base even has had over 1,300 contributors (https://github.com/JuliaLang/julia) which is quite a lot for a core language, and that's mostly because the majority of the core is in Julia itself.

This is one of the things that was noted quite a bit at this SIAM CSE conference, that Julia development tends to have a lot more code reuse than other ecosystems like Python. For example, the various machine learning libraries like Flux.jl and Lux.jl share a lot of layer intrinsics in NNlib.jl (https://github.com/FluxML/NNlib.jl), the same GPU libraries (https://github.com/JuliaGPU/CUDA.jl), the same automatic differentiation library (https://github.com/FluxML/Zygote.jl), and of course the same JIT compiler (Julia itself). These two libraries are far enough apart that people say "Flux is to PyTorch as Lux is to JAX/flax", but while in the Python world those share almost 0 code or implementation, in the Julia world they share >90% of the core internals but have different higher levels APIs.

If one hasn't participated in this space it's a bit hard to fathom how much code reuse goes on and how that is influenced by the design of multiple dispatch. This is one of the reasons there is so much cohesion in the community since it doesn't matter if one person is an ecologist and the other is a financial engineer, you may both be contributing to the same library like Distances.jl just adding a distance function which is then used in thousands of places. With the Python ecosystem you tend to have a lot more "megapackages", PyTorch, SciPy, etc. where the barrier to entry is generally a lot higher (and sometimes requires handling the build systems, fun times). But in the Julia ecosystem you have a lot of core development happening in "small" but central libraries, like Distances.jl or Distributions.jl, which are simple enough for an undergrad to get productive in a week but is then used everywhere (Distributions.jl for example is used in every statistics package, and definitions of prior distributions for Turing.jl's probabilistic programming language, etc.).

I think julia has a lot going for it. I feel like autograd is one of the bigger ones given that it's a language feature basically (https://github.com/FluxML/Zygote.jl for reference). I think the ecosystem is a bit of an uphill battle though.

Also check out https://github.com/FluxML/Zygote.jl which is the AD engine

> There's sadly no performant autodiff system for general purpose Python.

Like there is for general purpose Julia? (https://github.com/FluxML/Zygote.jl)

Thanks again. Like you said it is fun to dream (ask the "Scheme Machine" guys sometime about how they would go about it now), but practically with technology like Julia's Zygote:

https://github.com/FluxML/Zygote.jl

the efficiency of autodiff might be similar to that of an opcode anyway.

So, how did DEC do on the Alpha processor? I always heard good things about it--IIRC it was based on the VAX, but 64 bit. I learned PDP-11 assembler at RPI, during their college program for high school students in about 1984. We hand assembled code and really got to know the architecture.

and make that computable. It seems like line 88 --> 90 of this file in Zygote does that: https://github.com/FluxML/Zygote.jl/blob/master/src/compiler/chainrules.jl

It's super powerful

For example Zygote.jl (https://github.com/FluxML/Zygote.jl) implements reverse mode automatic differentiation, by defining a function that is a generated transformation of the function being differentiated.