Llm Inference Serving medium
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?
- A Network bandwidth between pools, since each request's KV cache, at hundreds of KB per token, must move to a decode GPU.
- B FLOPs on the decode pool, since decode becomes compute-bound once it no longer shares GPUs with prefill work.
- C CPU memory on the prefill pool, since prefill nodes must hold every finished request's output tokens for retries.
- D Storage throughput for weights, since both pools must reload the model from disk for every request handoff.
Answer: A. Network bandwidth between pools, since each request's KV cache, at hundreds of KB per token, must move to a decode GPU.
An 8K-token prompt at a few hundred KB per token produces **gigabytes of KV per request**, which must cross the network fast (RDMA-class links) or the handoff erases the TTFT gain. Disaggregation trades isolation and independent tuning for transfer cost and operational complexity. Decode stays bandwidth-bound after the split (B is wrong). When prompts are short, keeping prefill and decode on the same GPUs with chunked scheduling is often simpler and just as good.
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?
- Speculative decoding with a small draft model cut inter-token latency by about 2x at low traffic. At peak, with about 200 sequences batched per replica, aggregate throughput is 15% lower than with speculation disabled. What explains this, and what is the fix?