Your question is Optimizer Trade-offs for Large Models. Take a moment with it on the right.
Talk me through your thinking if you like. When you're confident, submit your answer and I'll grade it like a real screen (7/10 or better passes).
What are the trade-offs of using different optimizer variants, such as AdamW versus SGD with momentum, when training large-scale models?