A recent blog co-authored by Hwalsuk Lee at Upstage announced that Upstage’s Solar foundation model is now available on Amazon SageMaker JumpStart. The Solar language model has been pre-trained, offering improved functionality across languages, domains, and tasks.
The Solar Mini Chat and Solar Mini Chat – Quant models are now accessible via SageMaker JumpStart. Amazon SageMaker JumpStart provides machine learning solutions, enabling bespoke applications and the deployment of foundation models. Amazon SageMaker Studio allows users to discover and deploy the Solar model.
Solar is designed for English and Korean languages, and is tailored for multi-turn chat applications. It can handle extended conversations effectively, making it ideal for interactive applications. The model uses a scaling method termed depth up-scaling, enabling an efficient enlargement of smaller models.
In December 2023, Solar 10.7B topped the Open Language Large Model Leaderboard of Hugging Face, with fewer parameters but comparable response delivery to GPT-3.5, and twice the speed. The model allows customers to deploy the Solar Mini Chat model, capable of adeptly understanding Korean language nuances, thereby improving user interactions in chat environments.
To utilize the Solar models, users turn to SageMaker JumpStart, which can deploy pre-built models into a hosted environment ready for production. SageMaker Studio provides an integrated development environment where users can access necessary tools for ML development, from building, training, and deploying ML models, to data preparation.
The blog post guides users step-by-step on how to deploy the Solar Mini Chat model, list interactions with AWS Marketplace subscriptions, and deployment resources. Successful deployment allows the user to test the model through various SageMaker environments. The blog also provides a Python example, depicting how the model operates in the GitHub repository, alongside details on how the model can be used for request/response payloads compatible with OpenAI’s Chat completion endpoint.
Once users have trialed the endpoint, they can delete it and the model to avoid incurring costs. They can also shut down the SageMaker Studio resources that are no longer needed.
Solar models are pre-trained which reduces training costs and infrastructure, and allows for customizations for generative AI applications. They can be tested on SageMaker JumpStart console or SageMaker Studio console.
