Llm Inference Serving medium
Your multi-tenant inference API limits each tenant to 60 requests per minute. One tenant sends 40 requests per minute, each with a 100K-token prompt, and other tenants' TTFT degrades badly even though everyone is within limits. How should you rework the limits?
- A Lower the request limit for every tenant so total requests per minute stay under the cluster's measured capacity.
- B Cap each tenant at a fixed number of concurrent streams, since open connections are what consume GPU resources.
- C Meter tokens per minute, counting prompt and output tokens, because GPU cost scales with tokens, not requests.
- D Route the heavy tenant to replicas with larger batch sizes so its long prompts are processed more efficiently.
Answer: C. Meter tokens per minute, counting prompt and output tokens, because GPU cost scales with tokens, not requests.
Prefill compute grows with prompt length and KV memory grows with context, so one 100K-token request can cost as much as hundreds of short ones. **Token budgets** align quotas with real cost, at the price of more complex accounting, because output length is unknown at admission. Concurrency caps (B) bound KV use but still let a few huge prompts monopolize prefill. Request-based limits are adequate only when request sizes are roughly uniform.
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?