Llm Inference Serving hard
Responses stream to clients over SSE. A replica crashes after sending 400 of about 1,000 tokens. The client library automatically retries the original request on disconnect, and users now see duplicated paragraphs. What is the best fix?
- A Rely on the SSE Last-Event-ID header so the reconnected stream resumes at token 401 on whichever replica happens to receive the retry.
- B Buffer the full response on the server and send it only when generation completes, so a crash never exposes partial output to clients.
- C Have the client resend the prompt plus the 400 received tokens as a prefix to a healthy replica, which re-prefills them and continues generating.
- D Switch from SSE to WebSocket, whose bidirectional acknowledgments let a new replica replay exactly the tokens the client never acknowledged.
Answer: C. Have the client resend the prompt plus the 400 received tokens as a prefix to a healthy replica, which re-prefills them and continues generating.
Streaming **lowers perceived latency but makes partial output a failure mode**, and the KV cache and sampler state died with the replica. Last-Event-ID or WebSocket acks only help if some server still holds that state, so continuation requires re-prefilling prompt plus emitted tokens, at the cost of extra prefill and a possible shift in style. Buffering is the right choice for machine-consumed outputs such as strict JSON, where partial text is useless anyway.
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?