Llm Inference Serving easy
During a traffic spike, your LLM API's request queue grows without bound. Requests wait about 60 seconds before prefill, and most clients give up at 30 seconds, even though the GPUs are fully busy. What is the best fix?
- A Pack more requests into each batch, so the GPUs work through the backlog faster and queueing delay falls.
- B Remove the queue entirely and let every request start decoding immediately inside the running batch.
- C Raise the client timeout to 120 seconds, so queued requests eventually complete instead of failing.
- D Cap queue depth or wait time and reject excess requests fast with 429, so admitted ones meet their SLO.
Answer: D. Cap queue depth or wait time and reject excess requests fast with 429, so admitted ones meet their SLO.
**Admission control** sheds load early, so the GPUs stop spending work on requests whose clients have already timed out. The tradeoff is that some users get a fast rejection, which is better than nearly everyone getting a slow failure. Raising the timeout is tempting, but it only makes the backlog bigger. It is acceptable only for offline or batch callers that tolerate long waits. Admitting everything into the batch exhausts KV memory and slows every request.
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?