Google Open Sources Tools to Aid AI Model Development

Google’s Cloud Next conference this year showcased a departure from its usual lineup of managed and closed-source products by introducing several open-source tools aimed at supporting generative AI projects and infrastructure. Among these offerings, MaxDiffusion stands out as a collection of reference implementations of diffusion models, such as the Stable Diffusion image generator, optimized to run on XLA devices like Google’s tensor processing units (TPUs) and recent Nvidia GPUs. XLA, or Accelerated Linear Algebra, accelerates specific AI workloads, including fine-tuning and serving.

JetStream, another new tool, is an engine designed to run text-generating AI models like Gemma 7B and Meta’s Llama 2. Initially supporting TPUs with GPU compatibility planned for the future, JetStream promises up to 3x higher performance per dollar, although specifics regarding the benchmarking and comparisons behind this claim are not entirely clear yet.

Additionally, Google expanded its MaxText collection, which targets TPUs and Nvidia GPUs, to include models like GPT-3, Gemma 7B, Llama 2, and models from Mistral. These models are optimized for performance on TPUs and large GPU clusters, aiming to maximize energy efficiency and cost optimization.

Overall, these open-source contributions from Google reflect a shift towards fostering developer engagement and supporting AI infrastructure in a more accessible and collaborative manner. By providing reference implementations, optimized engines, and expanded model collections, Google aims to empower developers to leverage generative AI technologies more effectively while driving innovation in the field.