New Colossal-AI Released: Offers Hardware Savings Up To 46x for AI-Generated Content and Boasts Novell Automatic Parallelism

Dover, DE - January 18, 2023 - Colossal-AI, the industry leading open source system for maximizing the speed and scale of training, inference and fine-tuning of large deep learning models, today unveiled a new product release version 0.2.0.

This new version includes two pre-configured recipes, one for Stable Diffusion 2.0 and the other for BLOOM. The recipes are designed to support instant training and inference of models for AI-Generated Content (AIGC) and allow for significantly reduced hardware costs, of up to 46 times. In addition, the new release also features automatic parallelism, which is not offered by other solutions and enables instant distributed training on multiple computing devices with a single line of code.

Colossal-AI Delivers Exceptional Cost-Savings for AIGC
AIGC has recently risen to be one of the hottest topics in AI (artificial intelligence), because it rewrites the rules of human creativity and content creation. Methods that train and fine-tune AIGC models in a faster and cheaper manner have become extremely sought after for the commercialization and application of AIGC.

AI model training is the process of teaching a machine learning model to recognize patterns in data by adjusting the model's parameters based on data. In inference, a trained deep learning model is being used to make predictions on new, unseen data. Fine-tuning allows a pre-trained model to adapt to a new task and improve its performance, by adjusting the model's parameters based on new data.

The new Colossal-AI release includes new deep learning optimization recipes significantly lowering the cost for two AIGC models. The Stable Diffusion 2.0 model recipe enables low-cost training and inference, while also reducing GPU memory consumption by up to 5.6 times and hardware costs by up to 46 times. This allows users to meet their requirements with consumer-grade graphics cards like the RTX 3060 12 GB GPU, which costs $329, rather than the more expensive A100 80GB GPU, which costs $14,999, resulting in hardware cost savings of up to 46 times per GPU. The ready-made inclusion of the 175 billion parameter BLOOM model features stand-alone inference with a 4-fold reduction in GPU memory consumption and hardware cost savings of more than 10 times.

Deep learning optimization recipes are pre-configured model architectures and hyperparameters that are designed to perform well on a specific task or dataset. They can save time and resources by providing a starting point for AI developers and data scientists, rather than requiring them to design and test their own model architectures from scratch. Colossal-AI's new optimization recipes make it easier for users to get started with deep learning and achieve high cost savings with minimal effort.

Colossal-AI Introduces Automatic Parallelism for Simplified Distributed Training
Colossal-AI's world-first technology for automatic parallelism makes distributed training easier and more accessible for AI developers. Distributed training systems often require manual configuration of complex parallel policies, which can be challenging for developers without expert knowledge in system engineering and configuration. Colossal-AI simplifies the process with just one line of code, requiring only cluster information and model configurations from the user. It can find a unique parallelism method for each operation and potentially discover a better strategy than what human experts could provide.

Colossal-AI's automatic parallelization saves AI developers and engineers a significant amount of time and effort compared to traditional approaches that require configuring a cluster of machines, installing and configuring software and libraries, setting up communication infrastructure between the machines in the cluster, and writing code to parallelize the training process, including partitioning the data and model across the machines in the cluster and coordinating the training process. With the new Colossal-AI, they can easily search for the best parallelism strategy with just one line of code, making distributed training faster and easier than ever before.

"Our customers are always looking for ways to optimize their deep learning infrastructure to reduce costs, and with our latest release, we've given them two important integrations into Stable Diffusion and BLOOM as well as automatic parallelism, a world-first technology, making our product the most efficient solution available on the market today for AIGC and other large AI models on the market today," says Prof. James Demmel, Co-founder and CSO at HPC-AI Tech, and professor at UC Berkeley. "By making our innovations available as open-source software, we are bringing the latest advances in AI within reach for everyone worldwide at lower costs, leading to rapid and wide adoption of Colossal-AI."

For more information please visit https://www.hpc-ai.tech/blog/colossal-ai-0-2-0.

Find the latest source code of Colossal-AI on GitHub at https://github.com/hpcaitech/ColossalAI.

HPC-AI Tech
+1 6693070917
contact@hpc-ai.tech
www.hpc-ai.tech

HPC-AI Tech is a global company focusing on High Performance Computing and Artificial Intelligence that has developed an efficient large AI model training, inference and fine-tuning system, called Colossal-AI. Colossal-AI integrates with many advanced technologies such as multi-dimensional tensor parallelism, sequence parallelism, heterogeneous memory management, large-scale optimization and adaptive task scheduling. It helps users to efficiently and quickly deploy state-of-the-art AI applications with budget savings of up to 46x and go to market speeds of 7-10 times faster than comparable solutions.

Colossal-AI is available under an open-source license on GitHub and has ranked first as a trending project multiple times with about 8,000 Stars. Plus, Colossal-AI natively supports popular AI frameworks such as PyTorch Lightning, Hugging Face, or Timm and it supports the training of any size model such as Stable Diffusion, BLOOM, AlphaFold or GPT-3 on any GPU. This means AI developers and data scientists can use their favorite tools with ease to utilize any of the over 20,000 available deep learning models.

The company was founded by Dr. Yang You, Founder and Chairman and Prof. James Demmel, Chief Strategy Officer and professor at UC Berkeley. The company is based out of Delaware and Singapore. Users worldwide include AWS, Meta, BioMap, Hugging Face, and Lightning AI. The company is backed by BlueRun Ventures, Sinovation Ventures and ZhenFund. For more information, visit www.hpc-ai.tech.

This release was published on openPR.