# Booking the GPUs (2 x RTX 5090)

You get a Linux VM with both GPUs for a limited time. Only one booking runs at a time.
Base URL: https://gpu.atibba.com. Send your token with every request:
`Authorization: Bearer YOUR_TOKEN`.

## 1. Create an SSH key for this booking

    ssh-keygen -t ed25519 -N "" -f gpukey

## 2. Book

    curl -X POST https://gpu.atibba.com/api/bookings \
      -H "Authorization: Bearer YOUR_TOKEN" -H "Content-Type: application/json" \
      -d "{\"hours\": 3, \"purpose\": \"train model\", \"ssh_public_key\": \"$(cat gpukey.pub)\"}"

`hours` is 0.25 to 8. The limit is 16 hours per 24 hours and 2 open bookings per agent.
You get back an `id`, for example `b-1a2b3c4d`. Your time starts when the VM is ready.

## 3. Check every 30 seconds until status is `ready`

    curl https://gpu.atibba.com/api/bookings/BOOKING_ID -H "Authorization: Bearer YOUR_TOKEN"

| status | meaning |
|---|---|
| queued | waiting; `position` shows your place |
| waiting_approval | the owner is using the machine and must approve |
| starting | machine is being prepared, usually 1 to 3 minutes |
| ready | connect now |
| ending, done | finished |
| failed, cancelled | see `error`; book again if needed |

## 4. Connect

Save `certificate` from the answer as `cert.pub`, then run the `ssh` command it gives you,
with `gpukey` in place of `YOUR_KEY`:

    ssh -p 6050 -i gpukey -o CertificateFile=cert.pub agent@85.10.200.130

Check the GPUs with `nvidia-smi`.

## 5. Working in the VM

- You are the normal user `agent`, without root or sudo. Use Python venv, pip or conda in your
  home folder. Run `export TMPDIR=~/tmp` first if a build needs lots of temp space.
- GPU containers work with rootless podman:
  `podman run --rm --device nvidia.com/gpu=all ubuntu nvidia-smi -L`
- Disk: about 190 GB, plus 200 GB in `/shared`. RAM: about 100 GB. Internet is allowed; outgoing mail (port 25) is not.
- Internal networks cannot be reached.

## Keeping files between bookings

Everything in the VM is deleted when a booking ends, except these two folders:

| Folder | Who sees it | Use it for |
|---|---|---|
| `/shared/my-work` | Only you; this is your own folder | Your code, data, checkpoints and results |
| `/shared/common` | All agents; any agent can change or delete files here | Large models and download caches |

- **Save your work only in `/shared/my-work`.** Do not put anything private in `/shared/common`.
- Put large models in `/shared/common`, so you and other agents do not download them again.
  Check there first. `HF_HOME`, `TORCH_HOME`, `PIP_CACHE_DIR`, `UV_CACHE_DIR` and
  `CONDA_PKGS_DIRS` already point to it.
- Build venvs in `/shared/my-work` if you want to keep them. Your home folder is deleted.
- The two folders share 200 GB. Delete what you no longer need.
- **There is no backup.** Copy important results off the VM as well.
- Never store passwords or API keys in either folder.

## Public ports for your services

Four ports are open to the internet, TCP and UDP: **6051, 6052, 6053, 6054**. Start a service in
the VM on one of them, listening on `0.0.0.0`, and it is reachable worldwide at
`85.10.200.130:PORT`. Example: `python3 -m http.server 6051 --bind 0.0.0.0`

- There is no login or HTTPS in front of these ports. Protect your service yourself.
- All clients appear with the IP 10.10.11.1, so you cannot see or filter real client IPs.
- The ports work only during your booking, and stop when it ends.
- For browser microphone or camera access, or anything else that needs HTTPS, use your own TLS.

**Everything outside `/shared` is deleted when the booking ends.** Copy results out like this:

    scp -P 6050 -i gpukey -o CertificateFile=cert.pub -r agent@85.10.200.130:results .

## 6. Ending or extending

- Finished early: `curl -X DELETE https://gpu.atibba.com/api/bookings/BOOKING_ID -H "Authorization: Bearer YOUR_TOKEN"`
- More time:

      curl -X POST https://gpu.atibba.com/api/bookings/BOOKING_ID/extend \
        -H "Authorization: Bearer YOUR_TOKEN" -H "Content-Type: application/json" -d '{"hours": 1}'

  This works when nobody is waiting. Save the new `certificate` as `cert.pub` afterwards.

You see a warning in your terminal 10 minutes before the end. `seconds_left` in the booking
shows the time left. At the end your login stops and the VM resets.