Llm Inference Serving medium
Token streaming works when you hit the inference server directly, but through your production reverse proxy users see nothing for 20 seconds and then the whole response at once. Server logs show tokens flushed every ~50 ms. What is the most likely cause and fix?
- A SSE cannot traverse reverse proxies reliably; migrate the endpoint to WebSockets so frames pass through unbuffered.
- B The model decodes too slowly under load; add GPU replicas so that tokens are produced fast enough to stream.
- C The proxy is buffering or compressing the response; disable buffering and gzip for the text/event-stream route.
- D The client's HTTP library waits for Content-Length; send the full length in headers before streaming starts.
Answer: C. The proxy is buffering or compressing the response; disable buffering and gzip for the text/event-stream route.
Reverse proxies often **buffer upstream responses**, and compression waits to fill blocks, which collapses a stream into one delivery. Disabling both for event-stream routes restores incremental delivery and gives up compression savings in exchange for perceived latency. SSE is plain HTTP and passes through proxies once buffering is off (A is wrong), so WebSockets are worth it only for bidirectional needs. Content-Length (D) cannot be known in advance for a stream, which uses chunked transfer instead.
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?