How do you scale a LLM microservice if it has 10min cold start up time? How to load balance LLM service? How to scale up high servicec load? Whats the difference between ... and rerank? Explain why and how NN activation maps in consume GPU memory. Explain activation functions and its function in FFNs. What is LORA? How do you decrease model size (Quantization, etc)? How do you decide to shift to more complex model architecture (how do you evaluate that the simpler model reached its max capability)? More questions I don't remember