Llm Inference Serving medium
A model that needs 4 GPUs runs on a single NVLink-connected node using 4-stage pipeline parallelism. At low concurrency, per-token latency is poor and each GPU sits idle about 75% of the time. What explains this, and what is the better layout?
- A KV cache transfers between stages saturate NVLink; keep pipeline parallelism but quantize the KV cache.
- B A token passes through the stages in turn, leaving 3 GPUs idle; use tensor parallelism within the node.
- C Each stage is compute-bound on its own layers; add a fifth GPU to shorten every stage and lower latency.
- D Pipeline stages split memory bandwidth unevenly; rebalance the layers so each GPU holds an equal share.
Answer: B. A token passes through the stages in turn, leaving 3 GPUs idle; use tensor parallelism within the node.
With pipeline parallelism, one sequence's decode step crosses the stages serially, so without many micro-batches in flight, **pipeline bubbles** leave most GPUs idle. **Tensor parallelism** has all 4 GPUs read their weight shards at once, cutting per-token time at the cost of per-layer all-reduces, which are cheap over NVLink. Stages pass activations, not KV cache (A is wrong). Pipeline parallelism is the right choice across nodes with slow interconnects, where frequent all-reduces would dominate.
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?