Llm Inference Serving easy
A summarization endpoint takes a 3,000-token prompt and returns a 600-token summary, and p50 latency is about 15 seconds. Product wants it faster without changing hardware. Which change is likely to cut end-to-end latency the most?
- A Ask for a shorter summary of about 150 tokens, since each output token needs its own sequential forward pass.
- B Trim the prompt to 1,500 tokens, since the input is five times larger than the output and dominates the total work.
- C Double the server batch size, since processing more requests together amortizes per-step overhead for everyone.
- D Switch to streaming responses, since sending tokens as they are generated shortens the total generation time.
Answer: A. Ask for a shorter summary of about 150 tokens, since each output token needs its own sequential forward pass.
Prefill processes the 3,000 prompt tokens **in parallel** in a fraction of a second, while 600 output tokens need 600 **sequential** decode steps, which is where the 15 seconds goes. Trimming the prompt is tempting because it is the larger number, but it only shaves TTFT. It becomes the right lever when prompts are very long (tens of thousands of tokens) and outputs are short. Bigger batches raise latency, and streaming changes perceived latency, not total latency.
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?