Llm Inference Serving hard
A static-batching server groups 32 requests and runs each batch until every sequence finishes. Output lengths range from 20 to 2,000 tokens. Throughput is poor, and short requests wait behind long ones. Which change best fixes both problems?
- A Bucket requests by prompt length before batching, so sequences within a batch finish at similar times and padding waste is minimized.
- B Reduce the static batch size to 8, so each batch completes sooner and short requests spend less time stuck behind the long-running ones.
- C Schedule at iteration level: after each decode step, remove finished sequences and admit queued requests into the freed slots.
- D Sort the queue by requested max_tokens and run shortest-job-first batches, so short requests never share a batch with long ones.
Answer: C. Schedule at iteration level: after each decode step, remove finished sequences and admit queued requests into the freed slots.
**Continuous batching** keeps GPU slots full as sequences finish at different times, and it lets short requests leave immediately. Sorting by max_tokens or prompt length fails because max_tokens is only a ceiling and prompt length does not predict output length. Smaller static batches cut waiting but lower throughput, and static batching is defensible only for offline jobs with uniform, known output lengths.
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?