vLLM for one person: serving a small model

I use vLLM to serve a small Japanese-optimized model with continuous batching to achieve high throughput for embeddings and light inference tasks.