Account, Billing & Settings
TeachFORGE.ai
Account, Billing & Settings / Your own GPU
Connect your own GPU
Your browser talks to your GPU directly. Nothing goes through TeachForge's servers.
Your hardware, your model, your data, no bill.
What it does
Your own GPU points TeachForge at an OpenAI-compatible endpoint you run — vLLM, Ollama or similar. The panel is specific about the arrangement: your browser calls it directly, the GPU never leaves your machine, and no request goes through TeachForge's servers.
That is the privacy property worth understanding. The prompt goes from your browser to your own machine. TeachForge sees the result you choose to save, not the traffic.
It needs Standard or above, and an administrator must have enabled bring-your-own AI for the site. The panel says which of the two is missing when it is unavailable.
Setting it up
Set up vLLM for browser access links a step-by-step guide for each operating system — Ubuntu desktop, Windows and macOS. Follow the one for your machine; the browser-access part is the bit that catches people out and the guides cover it.
Then, on this tab: switch Enable on, enter the GPU base URL and the Model, and press Test connection. Success reads Connected — n model(s) found, with the models listed.
Two details the panel spells out and people still miss. The URL must end in /v1. And when the GPU runs on the same computer, use the loopback address — it works even on the live secure site, because loopback is exempt from mixed-content blocking. A machine name or a local network address will not be.
Inference overrides
An optional Inference overrides section takes Temperature, Max tokens, Top P, Context window, Timeout (seconds) and Thinking budget (tokens), plus Enforce JSON schema (guided decoding).
Leave them alone to start with. The two worth reaching for are the context window, if long prompts are being truncated, and the timeout, if a large generation is being cut off before it finishes.
A VRAM & context — rough guide disclosure gives bands from about 8 GB upwards, which is the fastest way to judge whether a model will fit before you download it.
While it runs
Your browser is what talks to the GPU, so the tab has to stay open for the whole generation. The AI panels say so, and it is the difference between this mode and the platform-hosted one.
Everything else is identical: the same compiled prompt, the same task defaults, the same validation of what comes back.