Джамба хибриден трансформатор-Мамба тест за модели

Отговорите и обясненията се появяват, докато отговаряте на теста.

1. Кои два типа основни слоеве преплита Jamba?
   • Конволюционни и обединяващи слоеве
   • RNN и CNN слоеве
   • Трансформатор на вниманието и слоеве на пространството на състоянието на Mamba
   • Слоеве на дифузия и внимание
2. Каква допълнителна техника използва Jamba за увеличаване на капацитета, като същевременно поддържа ниски активни параметри?
   • Смес от експерти (MoE)
   • Дестилация на знания
   • Обучение за квантуване
   • Подсилващо обучение от човешка обратна връзка
3. How does a Mamba layer's recurrent representation differ from an attention layer's KV cache?
   • It removes every model parameter from memory
   • It skips all tokens beyond the first one
   • It stores an independent full copy of every earlier attention matrix
   • It maintains a fixed-size recurrent state instead of retaining a growing set of past keys and values
4. Какъв контекстен прозорец поддържаше първият модел на Jamba?
   • 4K токени
   • 256K токена
   • 32K токена
   • 8 милиона токена
5. What is a reason to retain attention components alongside state-space layers in a Jamba-style hybrid?
   • To combine different access and representation mechanisms while evaluating their quality-efficiency tradeoff
   • To guarantee perfect retrieval of every token on every task
   • To make all attention computation independent of sequence length
   • To eliminate the need for any model evaluation
6. How should Jamba's reported long-context memory benefit be evaluated for deployment?
   • Assume every 256K request fits any single GPU
   • Count only the recurrent state while ignoring weights and attention cache
   • Measure the actual model, precision, batch size, context, and hardware configuration
   • Treat every hybrid model as having zero sequence-state cost
7. Does original Jamba's use of many Mamba layers make the whole model's sequence memory constant as input length grows?
   • Yes; every hybrid attention cache has a fixed one-token size
   • No; its retained attention layers still store a growing KV cache
   • Yes; model weights disappear once the sequence becomes long
   • Only the tokenizer's file size changes
8. Why are total MoE parameters and active parameters per token different quantities in Jamba?
   • Every stored expert is necessarily deleted when inactive
   • A token must run all experts regardless of routing
   • The model's supported context directly sets its vocabulary size
   • Routing can activate only a subset of stored experts for each token