Llm Inference Serving easy
Many users of your chat app click 'stop' or close the tab mid-response, but the server keeps generating until it reaches max_tokens. What is the impact of propagating these cancellations to the inference engine?
- A It frees decode slots and KV memory sooner, raising effective capacity at the cost of extra plumbing.
- B It has little impact, because a batch is fixed once formed and cannot drop requests mid-generation.
- C It mostly saves network bandwidth, since KV cache is released only once the whole batch finishes.
- D It only lowers TTFT for the cancelled user, since cancelling never affects others in the same batch.
Answer: A. It frees decode slots and KV memory sooner, raising effective capacity at the cost of extra plumbing.
With **continuous batching**, a cancelled sequence leaves the batch at the next iteration. Its KV cache is freed and its slot goes to a queued request, so wasted decode work turns into capacity. The cost is wiring disconnect signals through the gateway and server. The 'little impact' answer would hold under **static batching**, where freed slots sit idle until the whole batch finishes.
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?