A cemetery-plot office helper must demo on a staff laptop with no GPU this weekend, then later call the county NIM endpoint. How should the app be structured?
Select an answer to reveal the explanation.
Short Explanation
Demo on a laptop this weekend, county NIM later. Keep one prompt-building path and swap only the client: a local Hugging Face pipeline now, HTTP to NIM in production. Do not rewrite prompts per host, stand up NCCL for a dry run, or demand a video GPU.
Full Explanation
Local Hugging Face inference and a remote NVIDIA microservice are interchangeable clients behind one app interface. Keep prompt construction stable and swap only the caller: pipeline on the laptop, HTTP to NIM in production. Rewriting prompts per host, standing up an NCCL cluster for a weekend demo, or demanding a video GPU all miss that client-swap design.