Bria.ai is building a fully copyrighted generative AI platform for developers. AI and engineering teams can use this platform to accelerate visual content creation, increase productivity, and safely develop AI products at scale.
Part of the Bria platform is providing foundation models in the field of text-to-image. The models are trained from scratch using licensed data from the „Fair Diffusion“ program. For that, Bria uses AWS cloud infrastructure for the entire training lifecycle. This includes Amazon SageMaker for cloud orchestration, Amazon S3 data stream, and third-party libraries like Hugging Face Accelerate.
With this challenge in mind, Bria.ai contacted Automat-it and engaged the experienced MLOps team. Working closely with Bria.ai, Automat-it developed a custom container that worked with Amazon SageMaker DDP policy. This enabled Bria to seamlessly replace the current training stack from Hugging Face and accelerate to MosaicML’s Composer framework, saving GPU costs by using new optimizations that were not available before.
Bria.ai now runs Stable Diffusion with a MosaicML Composer trainer integrated with Amazon SageMaker’s state-of-the-art PyTorch Distributed Data-Parallel implementation.
Benefits:
- Reached production grade in weeks rather than months.
- 40% reduction in training cost
- A secure, robust, customized container framework enables Bria.ai to run the same workloads locally and on Amazon SageMaker.
- Secure backup implementation.
- Swift implementation of new functionalities
