jyafn
is a project for enabling MLOps by using Computational graphs that compile to machine code, with a convenient and familiar Python interface.
💡 Don't forget to check out the docs for more in-depth info.
Look at this little innocent piece of code:
import jyafn as fn
import numpy as np
@fn.func
def reduce_sum(mat: fn.tensor[2, 2]) -> fn.scalar:
return np.sum(mat)
I know: it looks a bit funny but what if I told you that
- This compiles to machine code.
- You can still call it as a regular Python function.
- You can export it, load it and call it from Go (or Rust, or Python, or C!)
Neat, huh? It's basically tf.function
+ onnx
in a single package!
Let's write a silly function in jyafn
:
@fn.func
def a_fun(a: fn.scalar, b: fn.scalar) -> fn.scalar:
return 2.0 * a + b + 1.0
It's so silly that if you call it like you normally would, a_fun(2, 3)
, you get what you expect, 8
. But that is not the fun part. The fun part is that you can export this function to a file:
with open("a_fun.jyafn", "wb") as f:
f.write(a_fun.dump())
And now you can pass this file anywhere and it will work. Let's call it, for example, from Go:
// Read exported data:
code, err := os.ReadFile("a_fun.jyafn")
if err != nil {
log.Fatal(err)
}
// Load the function:
fn, err := jyafn.LoadFunction(code)
if err != nil {
log.Fatal(err)
}
// Call the function:
result, err := jyafn.Call[float64](
fn,
struct {
a float64
b float64
}{a: 2.0, b: 3.0},
)
if err != nil {
log.Fatal(err)
}
fmt.Println(result, "==", 8.0)
For all cases, unfortuately you will need GNU's binutils
(or equivalent) installed (it is not a build dependency!), since we need an assembler and a linker to finish QBE's job. In most computers, it's most likely already installed (as part of gcc
or Python). However, this is a detail that you need to be aware when, e.g., building a Docker image. Also, jyafn
is guaranteed not to work in Windows. For your specific programming environment, see below:
This is the most convenient way of getting jyafn
:
pip install jyafn
Clone the repo, then
make install
This should do the trick. You can set the specific target Python version like so:
make install py=3.12
The default version is 3.11 at the moment.
At the moment, the Python version depends on the Rust compiler to work. It will compile jyafn
from source. As such, you will need cargo
, Rust's package manager, as well as maturin
, the tool for building Python packages from Rust code. Maturin can be easily installed with pip
:
pip install maturin
You can use this as a Go module:
import "github.com/viodotcom/jyafn/jyafn-go/pkg/jyafn"
You will also need to install the libjyafn
shared object in your system, which is available for your platform in the GitHub latest release.
You can get the latest version of the jyafn
crate, the basis of this project, from crates.io directly:
cargo add jyafn
Jyafn is available to be used directly from C via the libjyafn
shared object that is available in the GitHub latest release. Please check the Rust interface for details on how to use the available functions.
Yes, there is definitely something going on. What you see is basically a mini-JIT (just-in-time compiler). Your Python instructions (add this! multiply that!) are recorded in a computational graph, which is then compiled to machine code thanks to QBE, which does all the heavy-lifting, "compilery stuff". This code is exposed as a function pointer in the running program.
Yes, but the code produced by jyafn
is far from arbitrary. First, it's pure: it doesn't cause any outside mutations of any kind. Second, since it is based on a computational graph, which is acyclic, it is guaranteed to finish (and the size of the code limits how much time it takes to run). Third, no machine code is exchanged: that code is only valid for the duration of the process in which it resides. It's the computational graph and not the code that is exchanged.
No, far from it:
- Your python function can only have constant-sized loops.
- The libraries you are using must be able to work with generic Python objects (i.e., support the magic dunder methods,
__add__
,__sub__
, etc...). Thankfully,numpy
does just that for itsndarray
s. if-else
is still a bit wonky (you can use thechoose
method instead). This can be solved in the future, if there is demand.- By now, support is focused on DS uses. So, floats and bools are supported, but not other primitive types.
- Once in a while, a new function out of a litany of them, might not be supported. However, implementing it is very easy.
You bet! There is a benchmark in ./jyafn-python/tests/simple_graph.py
at which you can take a look. One example showed a 10x speedup over vanilla CPython.
By now, Go, Rust, Python and C. You can use the cjyafn
library to port jyafn
to your language. Compiled shared objects are available as GitHub releases for this repo.
It's mature enough for a test. Guinea pigs wanted!