The BYOW system simplifies both external and internal weight management. The idea is to provide API that enables automatic downloading of model weights from an external source (HF, internal post-training service, S3, etc) into the TokenFactory weights management platform, which then validates it and handles all necessary replication across regions.This system not only allows users to bring their own models, but establishes next generation of internal weights management. Eventually, we should stop copying weights around manually and worry about their location.The system supports both full weights and LoRAs upload. Right now, only full weights can be used with dedicated endpoints.Model ArtifactsWe have UI for operating model artifacts, but API is also available. You can find examples in the ipynb.Currently, base models can be uploaded from HF or the internal post-train service. Other ways to upload models are WIP - feel free to ask for the new ones.https://gitlab.nebius.dev/project-ai-studio-be/ai-studio-service/-/blob/main/inference_playground/tools/base_models_and_byow.ipynb?ref_type=heads
