Llm Inference Serving hard
A model needs 16 GPUs to hold its weights plus a useful KV cache. You have two 8-GPU nodes with fast intra-node links and a much slower inter-node network. Traffic is interactive with moderate concurrency. Which layout is most defensible?
- A 16-way tensor parallelism across both nodes, since splitting every layer 16 ways gives the lowest per-token latency regardless of interconnect speed.
- B 16-stage pipeline parallelism with one stage per GPU, since stages only pass activations and therefore never contend for interconnect bandwidth.
- C Two independent 8-GPU replicas with 4-bit weights, since quantization is effectively free and it avoids multi-node parallelism entirely.
- D 8-way tensor parallelism inside each node and 2-stage pipelining across nodes, with enough concurrent sequences to keep both stages busy.
Answer: D. 8-way tensor parallelism inside each node and 2-stage pipelining across nodes, with enough concurrent sequences to keep both stages busy.
Tensor parallelism needs all-reduces in every layer for every token, so it belongs on fast intra-node links, while pipeline parallelism sends activations only once per stage boundary and tolerates the slow network. The trade-off is that pipelining does not cut single-sequence latency and leaves **bubbles** unless enough sequences are in flight, which 16 stages would make much worse. Quantizing to fit one node is valid only after quality evaluation, not on the assumption that it is free.
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?