Llm Inference Serving medium
You must choose between two open-weight models of similar size and quality for a long-context, high-concurrency service. Model X uses multi-head attention with 64 KV heads. Model Y uses grouped-query attention with 8 KV heads. How should the attention design affect your choice?
- A It should not matter much, because the KV cache is small next to the weights, which dominate GPU memory.
- B Prefer Model X, because more KV heads split more evenly under tensor parallelism and therefore prefill faster.
- C Prefer Model Y: its KV per token is about 8x smaller, so more sequences fit and decode reads fewer bytes.
- D Prefer Model Y, because grouped-query attention mostly shrinks the weight files so the model fits on fewer GPUs.
Answer: C. Prefer Model Y: its KV per token is about 8x smaller, so more sequences fit and decode reads fewer bytes.
KV bytes per token scale with the number of KV heads, so **GQA** with 8 heads cuts KV memory and per-step KV reads by about 8x. That directly raises the concurrency ceiling at long context, in exchange for a usually negligible quality difference. Option A is the common misconception: the KV cache is small only at short contexts and low batch sizes, which is the one regime where the choice barely matters.
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?