The abstractions provided by JAX for parallelism are beautiful. JAX is an absolute master-class in programming-interface design and a lesson in the power of providing composable primitive operations and FP inspired design. An astounding amount of complexity is hidden from the user behind primitives like pmap, and the power is exposed in such a simple interface.
Thanks for the kind words! We've been doing a lot more work in this direction too, for both compiler-based automatic parallelization [0] and a work-in-progress pmap successor for 'manual' parallelism (per-device code and explicit collectives) [1] which composes seamlessly with the compiler-based stuff.
That's true, and is a massive part of what I love about JAX, but they also form barriers in weird parts of your code, preventing standard introspection tools, which is the single thing I hate about JAX. The errors are amazingly opaque.
If you have any particular examples in mind, and time to share them on https://github.com/google/jax/issues, we'd love to try to improve them. Improving error messages is a priority.
About introspection tools, at least for runtime value debugging there is to some extent a fundamental challenge: since jax.jit stages computation out of Python (though jax.grad and jax.vmap don't), it means standard Python runtime value inspection tools, like printing and pdb, can't work under a jax.jit as the values aren't available as the Python code is executing. You can always remove the jax.jit while debugging (or use `with jax.disable_jit(): ...`), but that's not always convenient, and we need jax.jit for good performance.
That's handy, and I hadn't seen it before, thanks.
It's been a bit, but I think the most frustrating errors were around mapping pytrees (like this issue https://github.com/google/jax/issues/9928). I'm not sure the exact solution, but the axis juggling and specifications were where I remember a lot of pain, and the docs (though extensive) were unclear. At times it feels like improvements are punted on in the hopes that xmap eventually fixes everything (and xmap has been in experimental for far longer than I expected).
Also the barriers where I couldn't disable jit. IIRC pmap automatically jits, so there was no way to avoid staging that part out. When it came to doing some complex jax.lax.ppermute, it felt more difficult than it needed to be to debug.
Next time I encounter something particularly opaque, I'll share on the github issue tracker.
Thanks! One last thing, since I have your ear. The function transformation aspects of jax seem to make their way into downstream libraries like haiku, resulting in a lot of "magic" that can be difficult to examine and debug. Are there any utils you made to make jax's own transformations more transparent, which you think might be helpful to third party transformations?
Higher order functions are difficult in general, and it would be fantastic to have core patterns or tools for breaking them open.
This actually works with JAX's native transformations. (There's no `equinox.vmap` for example.)
On higher-order functions more generally, Equinox offers a way to control these quite carefully, by making ubiquitous use of callables that are also pytrees. E.g. a neural network is both a callable in that it has a forward pass, and a pytree in that it records its parameters in its tree structure.
You're right that downstream libraries have often tended to introduce magic (some more than others), and moreover one library's magic is typically incompatible with other libraries'. It's something that we're working on but we don't have much to show for it yet. Two avenues are:
1. as you say, exposing patterns and tools for library authors to implement transformations/higher-order primitives using JAX's machinery rather than requiring each library to introduce bespoke magic to do the same;
2. adding JAX core infrastructure which directly solves the common problems that libraries tend to solve independently (and with bespoke magic).
Thanks for the info! And I want to be clear that I make all these comments and questions out of a love for the tool. It's one of my favorite tools that I wish I could use at work. The design and community engagement are both fantastic.