179 questions
No questions match those filters.
← All questions
LLM FoundationsHard· Mixture of Experts· Distributed Training
Why do you need both a per-expert and a per-device bala...
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 plansPer-expert and per-device balancing losses are distinct requirements. Expert-level balance is a model quality concern, ensuring all parameter sets are trained uniformly. Device-level balance is a throughput concern, ensuring uniform GPU utilization. Even if experts are perfectly balanced, a GPU can still become a critical path if all its assigned experts are simultaneously popular. Therefore, both are necessary when using expert parallelism.