Hardware

Setup	GPU	Notes
Hosted API (OpenAI or another provider)	none	Any OS. Cost per run is shown before and recorded after.
Local vllm serve with a 4–8B model	16–24 GB VRAM, Linux, NVIDIA driver R580+	Free per run. compose.yaml sets it up.
Ollama / llama.cpp	optional	Works on CPU but slowly; use 1–2 parallel requests and a long timeout.
LoRA training	8 GB+ VRAM for 4-bit QLoRA of a 4B model	One training at a time per machine. Tiny models train on CPU (tests do).

The web app itself needs little: about 1 GB of RAM plus room for the documents and run folders.