Google unveils open source projects for generative AI

Google at Google Cloud Subsequent 24 unveiled three open supply initiatives for constructing and operating generative AI fashions. The corporate additionally launched new massive language fashions to its MaxText mission of JAX-built LLMs.

The brand new LLM fashions in MaxText embrace Gemma, GPT-3, Llama 2, and Mistral, that are supported throughout each Google Cloud TPUs and Nvidia GPUs, the corporate stated. 

The newly unveiled open supply initiatives are MaxDiffusion, JetStream, and Optimum-TPU.

MaxDiffusion is a group of high-performance and scalable reference implementations for diffusion fashions akin to Secure Diffusion. Just like the MaxText fashions, the MaxDiffusion fashions are constructed on JAX, which is a framework for high-performance numerical computing and large-scale machine studying. 

JAX in flip is built-in with the OpenXLA compiler, which optimizes numerical features and delivers wonderful efficiency at scale, permitting mannequin builders to deal with the mathematics and let the software program drive the simplest implementation. 

“We’ve closely optimized JAX and OpenXLA efficiency on Cloud TPU and partnered carefully with Nvidia to optimize OpenXLA efficiency on massive Cloud GPU clusters,” Google stated.

The corporate additionally launched Jetstream, which is an open supply optimized LLM inference engine supporting XLA compilers. 

“As prospects carry their AI workloads to manufacturing, there’s an growing demand for a cost-efficient inference stack that delivers excessive efficiency. JetStream helps with this want and affords assist for fashions educated with each JAX and PyTorch/XLA, and contains optimizations for common open fashions akin to Llama 2 and Gemma,” Mark Lohmeyer, common supervisor of compute and ML infrastructure at Google Cloud, stated.

Lastly, Google’s open supply bulletins included the launch of Optimum-TPU for PyTorch customers within the Hugging Face group. Optimum-TPU brings Google Cloud TPU efficiency optimizations for each coaching and inference. It helps the Gemma 2b mannequin now and Llama and Mistral quickly, Google stated.