Llm Inference Serving hard
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?
- A Lower max_tokens to 1,024 so each reservation is 8x smaller, and require clients that need longer outputs to send follow-up continuation requests.
- B Allocate KV in small fixed-size blocks on demand through a block table, so memory tracks actual length instead of the reserved maximum.
- C Quantize model weights to 4-bit, which frees enough HBM that the contiguous 8K reservations fit roughly twice as many concurrent requests.
- D Add tensor parallelism across more GPUs so each 8K reservation is sharded, multiplying the concurrent request count by the parallelism degree.
Answer: B. Allocate KV in small fixed-size blocks on demand through a block table, so memory tracks actual length instead of the reserved maximum.
Worst-case contiguous reservation wastes over 90% of KV memory here, and **paged attention** removes that fragmentation at the cost of a small block-table indirection inside the attention kernel. Lowering max_tokens also raises concurrency but truncates long outputs, which the question rules out, and it is a fine choice only when outputs are short by contract. Weight quantization or more GPUs add memory, but the waste ratio stays the same.
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?
- 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?
- 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?