Disaggregated Prefill and Decode Serving Quiz

Answers and explanations appear as you take the quiz.

1. What is the core hardware reason for separating prefill and decode onto different GPU pools?
   • Prefill is compute-bound while decode is memory-bandwidth-bound
   • They use different programming languages
   • Decode requires more electricity than prefill
   • Prefill cannot run on GPUs at all
2. What data structure must be transferred from prefill workers to decode workers?
   • The model weights
   • The KV (key-value) cache
   • The tokenizer vocabulary
   • The optimizer state
3. Which problem does disaggregation specifically reduce in a shared-GPU setup?
   • GPU driver crashes
   • Network packet loss on the internet
   • Head-of-line blocking where a long prefill stalls others' decode
   • Model hallucination
4. Why can prefill be batched aggressively but decode benefits from different tuning?
   • Decode never uses the GPU
   • Prefill only runs once per day
   • Batching is impossible during prefill
   • Prefill tokens are read in parallel, maximizing compute utilization per pass
5. Which interconnects are typically used to move the KV cache between disaggregated pools?
   • NVLink or InfiniBand
   • Dial-up modems
   • USB 2.0 cables
   • Bluetooth
6. Which named systems are associated with disaggregated prefill/decode serving?
   • TCP, UDP, and HTTP
   • DistServe, Splitwise, and Mooncake
   • Word, Excel, and PowerPoint
   • JPEG, PNG, and GIF
7. Why must the KV cache produced during prefill be available to the decode workers?
   • It contains the model’s optimizer state
   • It stores the user’s GPU allocation
   • Decode uses cached prompt keys and values to generate subsequent tokens without recomputing the whole prompt
   • It is needed only to calculate validation accuracy
8. What is a resource-allocation benefit of separating prefill and decode pools?
   • Both phases must use identical batch sizes
   • The model no longer needs a KV cache
   • The scheduler can eliminate all network communication
   • Each phase can be sized and parallelized for its own compute or memory bottleneck