Serverless Optimized Model Initialization With Lazy Loading

via “lazy model loading with automatic weight downloading”

min(DALL·E) is a fast, minimal port of DALL·E Mini to PyTorch

Unique: Implements lazy loading at the MinDalle orchestrator level rather than individual model classes, enabling centralized control over caching policy and device placement. Integrates directly with Hugging Face Hub's model_id resolution (no custom download logic), ensuring compatibility with future model updates and enabling users to override via HF_HOME environment variable.

vs others: Simpler than manual model management (e.g., torch.hub.load) while providing more control than fully automatic frameworks like Hugging Face transformers pipeline; lazy loading reduces cold-start time by 50-70% vs eager loading all three models.