Llm Inference Serving easy
A coding assistant runs at low concurrency, its GPUs have spare compute during decode, and users want faster per-token output. A teammate proposes speculative decoding with a small draft model. Why can this help?
- A The draft model replaces the large model on easy tokens, so output quality depends on the draft for those.
- B It shrinks the KV cache, because only the small draft model keeps attention state during generation.
- C It reduces memory bandwidth per step by loading only half of the large model's layers for each token.
- D The large model verifies several drafted tokens in one forward pass, cutting the number of sequential steps.
Answer: D. The large model verifies several drafted tokens in one forward pass, cutting the number of sequential steps.
**Speculative decoding** trades extra compute (running the draft model and discarding rejected tokens) for fewer sequential passes through the large model. Verification keeps the large model's output distribution unchanged. The tempting misconception is that the draft model replaces the large model on easy tokens, but it only proposes tokens. The technique stops paying off when acceptance rates are low or the GPUs are already saturated with large batches.
One system design question a day, with spaced repetition so the ones you miss come back.
Get the appRelated questions
- Your 70B-class model servers each handle about 40 concurrent chats with 4K-token contexts. Product raises the maximum context to 32K tokens, and many users start using it. What is the main capacity consequence?
- A 70B-class model (80 layers, grouped-query attention with 8 KV heads of dimension 128) is served in fp16 on one 8-GPU node with about 640 GB of total HBM. Product wants 256 concurrent sessions, each holding a full 32K-token context. What is the binding constraint, and what does it imply?
- Your server reserves a contiguous KV region sized for max_tokens = 8,192 for every request at admission, but the average sequence ends near 600 tokens. The node reports out-of-KV-memory at about 20 concurrent requests while GPU compute sits at 30%. Which change best raises concurrency without restricting legitimate long outputs?
- You must summarize 10 million documents overnight, and no user is waiting on any single result. How should you configure the serving stack compared with your interactive chat deployment?
- You serve a 7B-parameter model in fp16 (about 14 GB of weights) for a chat app. Users say single-conversation token generation is too slow, and you are choosing new GPU hardware. Which hardware spec most directly increases per-sequence decode speed?
- Your RAG workload sends 8K-token prompts and gets ~150-token answers from a 70B-class model. You plan to move prefill and decode onto separate GPU pools so long prefills stop stalling decode. What new resource must you size carefully?