Llm Inference Serving hard
Under continuous batching, running sequences keep growing until KV cache blocks run out mid-decode. Long-running requests that are close to finishing are the most latency-sensitive. Which preemption policy is best?
- A Preempt the newest sequences, free their blocks, and restore them later by recomputing via prefill, or by swapping KV to host memory if long.
- B Reserve worst-case KV for every request at admission, so preemption is never needed and no sequence can run out of memory during decode.
- C Preempt the oldest running sequences first, since they hold the most KV blocks and freeing them reclaims the largest amount of memory fastest.
- D Truncate the oldest tokens in every sequence's KV cache to a sliding window, so all sequences keep decoding and none of them are paused.
Answer: A. Preempt the newest sequences, free their blocks, and restore them later by recomputing via prefill, or by swapping KV to host memory if long.
The newest sequences have made the least progress and hold the smallest KV, so they are the **cheapest to evict and rebuild**. Recompute avoids PCIe traffic and suits short contexts, while swapping wins when re-prefilling a long context would cost more than the transfer. Worst-case reservation brings back the fragmentation that kills concurrency and fits only strict-latency, low-concurrency tiers, and sliding-window truncation changes outputs for models not trained with it.
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?