Llm Inference Serving medium
Your scheduler admits a request only if KV memory can hold its prompt plus max_tokens, so admitted requests are never preempted. The API's default max_tokens is 8,192, while the median reply is 300 tokens. The scheduler reports memory full, but measured KV usage is under 20%. What is the best fix?
- A Add GPUs to the pool, since the scheduler's memory accounting shows the pool is genuinely out of KV capacity.
- B Increase the maximum batch size in the scheduler config so it packs more sequences into each forward pass.
- C Remove max_tokens entirely and let every request generate until the model emits an end-of-sequence token.
- D Lower the default to fit observed lengths, or admit on expected length and accept occasional preemption.
Answer: D. Lower the default to fit observed lengths, or admit on expected length and accept occasional preemption.
**Worst-case reservation** guarantees zero preemption but strands memory reserved for tokens that are never generated. Realistic defaults or optimistic admission reclaim that capacity at the cost of occasional swap or recompute. Adding GPUs (A) only buys more stranded memory, and raising the batch limit (B) does nothing while admission is blocked by reservations. Worst-case reservation is right when outputs routinely approach the cap or preemption spikes would break a strict SLO.
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?