Викторина по квантованию

Ответы и пояснения появляются по мере прохождения теста.

1. In a weight-focused quantization scheme, what representation property is changed?
   • The mandatory number of network layers
   • The identity of every original training document
   • The numerical levels or precision used for selected weights
   • Only the displayed model title
2. Ignoring metadata and other working memory, what fraction of raw packed-weight bits remains when moving the same weights from 16-bit to 4-bit?
   • The same amount
   • Twice as much
   • One-hundredth in every case
   • One-quarter
3. Каков основной недостаток агрессивного квантования?
   • Ошибка округления может снизить точность модели.
   • Модель больше не может быть загружена.
   • Он навсегда удаляет данные обучения
   • Это всегда делает модель медленнее
4. Why do the cited GPTQ and AWQ methods inspect calibration inputs?
   • To guarantee identical performance on every future input
   • To guide their respective weight-quantization choices and error management
   • To translate the vocabulary into another language
   • To add a required number of new transformer layers
5. Why must runtime speed be measured rather than inferred from a lower weight bitwidth alone?
   • A bitwidth label guarantees an identical speedup on every device
   • All quantized values remove the need for computation
   • Every compressed model must be slower by definition
   • Kernels, hardware, and conversion overhead also affect runtime
6. What does activation-aware mean in the cited AWQ weight-quantization method?
   • Every activation must be stored at the same low bitwidth as every weight
   • The weights are never quantized
   • Activation observations help choose salient weight scaling
   • The method uses only parameter names without input observations
7. Using the stated reconstruction rule scale times code minus zero point with subtraction inside, what value comes from code six, scale 0.5, and zero point two?
   • Четыре
   • Два
   • Три
   • Шесть
8. How does quantization-aware training differ in timing from post-training quantization?
   • Training incorporates quantization effects so learned values can adapt
   • It is defined as changing only a fitted model after every training stage ends
   • It forbids all training examples
   • It guarantees no rounding or approximation error in any representation