The Anyscale team shared how you can achieve considerable speedups for model loading in production with examples on the Llama 2 variants.