Inference founder course labs

These labs run without a GPU. They let you exercise the control plane locally, then point the gateway at a real OpenAI-compatible vLLM endpoint.

1. Start the gateway

python3 gateway.py --port 8080

Set BACKEND_URL to proxy to a real endpoint. Without it, the gateway returns a deterministic streamed mock response.

2. Run a benchmark

python3 benchmark.py --url http://127.0.0.1:8080/v1/chat/completions --concurrency 1,4,8,16 --requests 40

3. Calculate unit economics

python3 unit_economics.py --gpu-hour 6.50 --tokens-per-second 1400 --utilization .55 --price-per-million 1.20

Treat every default as a placeholder. Replace it with a dated measured value.