Google's OpenXLA ML Compiler Ecosystem Simplifies ML
Created on March 10|Last edited on March 14
Comment
What is OpenXLA?
There's lots of different hardware and lots of different software. Unfortunately, not all the software is compatible with all the hardware and vice versa. Picking the right set or stack of tools for deploying a model (e.g. which framework to use and what hardware) is difficult.
OpenXLA is the Tower of Babel for ML frameworks and serving them on hardware. It's a compiler for TensorFlow, PyTorch, and Jax and it's meant to be a seamless tool for developers in all frameworks to deploy and serve models on a wide variety of hardware.
As shown in the above figure, any of the 3 frameworks is fed into this StableHLO (Stable High Level Operations) block. StableHLO is a portability layer which means it's responsible for things like quantization.
The current, general MLOps, ML, and data spaces are convoluted!
OpenXLA reminds me of the rise of unifying tools, like Ivy, that aim to simplify a lot of this complexity. Whereas Ivy aims to unify the ML frameworks, OpenXLA aims to simplify the deployment process for users of a variety of frameworks and hardwares.
References
Rubin, James. “OpenXLA Is Available Now to Accelerate and Simplify Machine Learning.” Google Open Source Blog, 8 Mar. 2023.
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.