Llm Inference Serving easy
Your model's fp16 weights are 140 GB, and quality requirements rule out quantization. Each GPU has 80 GB of memory. What must you do to serve this model?
- A Use tensor parallelism across several GPUs, splitting each layer's weights so the model and KV cache fit.
- B Run data-parallel replicas on two separate GPUs, so each one handles half of the incoming request traffic.
- C Increase the batch size, so the weight memory is amortized across more requests and the footprint shrinks.
- D Enable paged attention, so weights are paged into GPU memory on demand as each layer runs its forward pass.
Answer: A. Use tensor parallelism across several GPUs, splitting each layer's weights so the model and KV cache fit.
**Tensor parallelism** shards each layer's weights across GPUs, so a model larger than one GPU can run. The cost is inter-GPU communication on every layer, which calls for fast interconnects. Data-parallel replicas are tempting, but every replica needs a full copy of the model, so they only help once a single copy already fits. Paged attention manages **KV cache** memory, not weights.
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?