How would you design a distributed inference pipeline for a generative AI model with over 100 billion parameters at Meta Platforms?
