Google at Google Cloud Next 24 revealed three open source jobs for structure and running generative AI models. The company also introduced brand-new big language designs to its MaxText job of JAX-built LLMs.The brand-new LLM designs in MaxText include Gemma, GPT-3, Llama 2, and Mistral, which are supported throughout both Google Cloud TPUs and Nvidia GPUs, the company said.
The freshly revealed open source jobs are MaxDiffusion, JetStream, and Optimum-TPU.
MaxDiffusion is a collection of high-performance and scalable referral applications for diffusion designs such as Steady Diffusion. Like the MaxText models, the MaxDiffusion designs are developed on JAX, which is a framework for high-performance mathematical computing and massive machine learning.
JAX in turn is incorporated with the OpenXLA compiler, which optimizes mathematical functions and delivers outstanding performance at scale, allowing design home builders to focus on the math and let the software drive the most reliable execution.
“We’ve greatly enhanced JAX and OpenXLA performance on Cloud TPU and partnered carefully with Nvidia to enhance OpenXLA performance on big Cloud GPU clusters,” Google stated.
The company also introduced Jetstream, which is an open source enhanced LLM reasoning engine supporting XLA compilers.
“As clients bring their AI workloads to production, there’s an increasing need for a cost-effective reasoning stack that provides high performance. JetStream helps with this requirement and offers support for designs trained with both JAX and PyTorch/XLA, and consists of optimizations for popular open models such as Llama 2 and Gemma,” Mark Lohmeyer, general supervisor of calculate and ML infrastructure at Google Cloud, stated.
Finally, Google’s open source statements consisted of the launch of Optimum-TPU for PyTorch users in the Hugging Face neighborhood. Optimum-TPU brings Google Cloud TPU efficiency optimizations for both training and inference. It supports the Gemma 2b model now and Llama and Mistral soon, Google said.
Copyright © 2024 IDG Communications, Inc.
