Llm Inference Serving easy
Your autoscaler adds GPU nodes when queue depth rises, but each new node takes several minutes to start serving because it must download and load about 140 GB of weights. Traffic spikes cause errors before the new capacity arrives. What is the most reasonable mitigation?
- A Scale on GPU utilization instead of queue depth, since utilization reacts sooner to incoming spikes.
- B Keep a warm buffer of loaded replicas and cache weights on local disk, paying for some idle GPU time.
- C Scale to zero between spikes, since idle GPUs dominate cost and cold starts happen only occasionally.
- D Use larger batches on existing nodes, since this lowers latency for queued requests during each spike.
Answer: B. Keep a warm buffer of loaded replicas and cache weights on local disk, paying for some idle GPU time.
When cold starts take minutes, the only way to absorb a spike is capacity that is **already loaded**. A warm buffer and locally cached weights trade idle GPU cost for spike tolerance. Switching to a utilization signal is tempting, but GPU utilization sits near 100% whenever any decode is running, so it does not rise ahead of a spike, and no signal shortens the minutes-long weight load. Scale-to-zero fits only workloads that can tolerate minutes of startup delay.
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?