179 questions
No questions match those filters.
← All questions
Inference & ServingHard· Attention Mechanisms· KV Cache
Explain how MLA differs from GQA as a strategy for redu...
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 plansWhile GQA reduces the number of KV head pairs from N to K, cutting cache size by a factor of G, Multi-Head Latent Attention (MLA) keeps all N heads but projects them into a C-dimensional latent representation. This can reduce cache size by over 30x in specific configurations. The trade-off is that MLA requires additional up-projection compute at query time to reconstruct the necessary information, which is offset by the massive reduction in HBM traffic.