Grouped-Query Attention Quiz

Answers and explanations appear as you take the quiz.

1. In Grouped-Query Attention, what is shared among a group of query heads?
   • The key and value heads
   • The output projection only
   • The input embeddings
   • The learning rate
2. GQA is best described as a middle ground between which two extremes?
   • Convolution and attention
   • Full multi-head attention and multi-query attention
   • FP32 and FP8 precision
   • Encoder-only and decoder-only models
3. What part of inference does GQA primarily make cheaper?
   • Tokenization of the prompt
   • The training loss computation
   • The KV cache memory and its bandwidth during generation
   • The final softmax over the vocabulary
4. If a model has 32 query heads and 8 KV groups, the KV cache is reduced by roughly what factor?
   • About 32x
   • About 2x
   • No reduction
   • About 4x
5. What quality/speed tradeoff did the GQA paper report in its experiments?
   • Quality close to MHA with inference speed closer to MQA
   • Lower quality and slower inference than both
   • Exactly the same latency and quality for every model
   • A result only on image classification
6. How are existing multi-head models often converted to GQA?
   • By deleting all attention layers
   • By a short 'uptraining' step from the existing checkpoint
   • By switching to a different tokenizer
   • They cannot be converted at all
7. Which task families did the GQA paper evaluate?
   • Image segmentation and object detection
   • Speech synthesis and speaker verification
   • Language modeling and machine translation
   • Molecular generation and toxicity prediction
8. How much of the original pretraining compute did the GQA paper report using for its uptraining recipe?
   • About 50%
   • About 100%
   • No additional training compute
   • About 5%