179 questions
No questions match those filters.
← All questions
LLM FoundationsMedium· Model Architecture· Scaling Laws· Inference Performance
You have two candidate architectures at the same parame...
This is one of the questions in the full AI/ML interview bank. Pro unlocks all 1789 questions; Premium includes the same bank plus the highest daily Practice limit.
See plansAccording to Kaplan et al., model shape has a minimal impact on loss at a fixed parameter count. Therefore, the decision should be driven by operational requirements. The deeper model (48 layers) is 42% more expensive in terms of KV cache and introduces more serialization during decoding. However, if the deployment environment utilizes pipeline parallelism, the deeper model may be preferred as it partitions more cleanly across hardware.