Apple discreetly open-sources AI tools, including a library for large-scale deep learning on Apple Silicon and a versatile data loading framework

TL;DR:

  • Apple has quietly open-sourced AI tools, including a library for large-scale deep learning models and a machine learning framework for Apple Silicon.
  • MLX, the flagship release, stands out with a unified memory model, simplifying operations across supported devices.
  • Apple’s move is hailed as its most significant step into open-source AI, receiving praise from experts.
  • MLX offers a Python API akin to NumPy, while MLX Data streamlines data loading for various frameworks.
  • These developments foreshadow Apple’s future AI-centric operating systems, creating opportunities for product developers.
  • Apple’s earlier open-sourcing of AXLearn, a library for large-scale deep learning models, further underscores its commitment to AI research.

Main AI News:

In a groundbreaking maneuver, Apple has silently unveiled a suite of open-source AI tools that could revolutionize the landscape of artificial intelligence. This remarkable release encompasses a library designed for “large-scale deep learning models” optimized for execution on the public cloud, along with a framework tailored for machine learning on Apple silicon.

The initial rollout, made available under the MIT license this week, stands out for its native compatibility with Apple Silicon, requiring a single pip installation and devoid of any additional dependencies. This innovative offering is a game-changer for developers seeking to train or fine-tune transformer language models on Apple platforms.

Distinguished by its unified memory model, MLX, as Apple christens it, sets itself apart from other frameworks. Within MLX, arrays reside in shared memory, facilitating operations on these arrays across various supported device types without the cumbersome data transfer processes. This remarkable development signifies Apple’s most substantial foray into open-source AI to date, according to senior NVIDIA research scientist Jim Fan, who expressed his admiration for the design of MLX’s API and its minimalist examples featuring OSS models like Llama, LoRA, Stable Diffusion, and Whisper.

MLX also boasts a Python API closely aligned with the immensely popular NumPy library, ensuring a seamless transition for developers familiar with this ecosystem. Accompanying MLX is the release of MLX Data, a versatile package for data loading, which Apple researcher Awni Hannun described as “framework agnostic, efficient, and flexible.” This package supports PyTorch, Jax, or MLX itself, offering high efficiency while enabling the processing of thousands of images per second and the execution of arbitrary Python transformations on resulting batches.

In Apple’s own words, MLX and MLX Data are creations “designed by machine learning researchers for machine learning researchers.” The primary goal is to strike a balance between user-friendliness and efficiency in training and deploying models, all wrapped in a conceptually straightforward framework. Apple also encourages researchers to contribute and enhance MLX to foster the rapid exploration of new ideas.

This release isn’t just about AI tools; it’s a glimpse into Apple’s evolving strategy, with future Apple operating systems poised to become more AI-centric. Developers building products on iOS, iPadOS, or MacOS should take heed of this development, as it sets the stage for significant advancements in AI integration.

Apple’s commitment to advancing AI research doesn’t stop here. In a relatively understated move, the company earlier open-sourced AXLearn. Released under the permissive Apache 2.0 license in July, AXLearn is a library built atop JAX and XLA, designed to bolster the development of large-scale deep learning models. It excels in the composition of models from reusable building blocks, seamlessly integrating with other libraries like Flax and Hugging Face transformers.

AXLearn is tailored for scalability and is capable of handling models with hundreds of billions of parameters across thousands of accelerators, making it ideal for deployment on public clouds. Its robust configuration system empowers users to manage jobs and data efficiently.

This library builds upon GSPMD, a system introduced by Google in 2021, enabling users to write programs as they would for a single device and parallelize computations based on a few annotations. AXLearn embraces a global computation paradigm, allowing users to describe computation on a virtual global computer rather than focusing on individual accelerators. It’s versatile and adaptable, catering to a wide array of applications, including natural language processing, computer vision, and speech recognition.

Conclusion:

Apple’s strategic release of open-source AI tools, particularly MLX and AXLearn, signifies a substantial shift in the AI landscape. This move not only demonstrates Apple’s commitment to AI research but also sets the stage for AI-centric developments in their future operating systems. Developers and businesses should closely monitor these advancements, as they hold the potential to reshape the market and drive innovation in AI-driven products and services.

Source