Training
A GPU sandbox is a sandbox with a whole GPU of its own, in the EU. It starts with PyTorch and the
usual training libraries installed, so a training run starts right away. You drive it with the same
nido commands as any other sandbox, see AI Sandboxes.
Start a GPU sandbox
nido create train --gpu l4 --network github.com --network release-assets.githubusercontent.com
nido exec train -- nvidia-smi
--gpu l4 gives the sandbox one NVIDIA L4 with 24 GB of memory. If no GPU is free at that moment,
nido waits and shows how long. A sandbox on a machine that just started needs about three to five
minutes; most of that is loading the image.
--network opens the hosts your training needs; nothing else is reachable. Give one
--network per host:
| For | Hosts |
|---|---|
| Pretrained weights for YOLO | github.com, release-assets.githubusercontent.com |
| Hugging Face models and datasets | huggingface.co, **.hf.co |
| Python packages | pypi.org, files.pythonhosted.org |
*.example.com opens one more label, such as cdn.example.com; **.example.com opens any
number, such as eu.cdn.example.com. Hosts are set when the sandbox is created: to change them,
remove the sandbox and create it again. When a download fails, nido logs train names each host that was blocked: look for
is not allowed by any policy.
What is installed
| What | Included |
|---|---|
| Python | 3.12, in /sandbox/.venv, active for every command |
| PyTorch | torch 2.14 with CUDA 13.0, torchvision |
| Hugging Face | transformers, datasets, accelerate, peft, trl, bitsandbytes, huggingface_hub, safetensors |
| Data and vision | numpy, pandas, scikit-learn, matplotlib, opencv-python-headless, ultralytics |
| Notebooks | ipykernel |
| Tools | ffmpeg, git, git-lfs, build-essential, nvtop, tmux, uv |
| Agents | Claude Code, Codex, OpenCode, charta, kyte |
Add what is missing with uv pip install or pip install, for example nido exec train -- uv pip install wandb, with PyPI under --network. It installs into /sandbox/.venv, keeps what is
already installed and can also replace a version. Unsloth and Axolotl keep the installed torch.
vLLM brings its own torch version. The command-line tools of the libraries are on the path too:
hf, torchrun, accelerate, yolo.
With --image a GPU sandbox starts from your own image instead.
Get data in and results out
nido cp ./dataset train:/sandbox/dataset
nido cp train:/sandbox/runs ./runs
Paths inside the sandbox must be below /sandbox. Public datasets and models can also come
straight from Hugging Face, with its hosts under --network. Storage volumes cannot be attached
to GPU sandboxes yet.
Examples
| Page | What it shows |
|---|---|
| Vision model | Fine-tune YOLO on your own images and export it to ONNX |
| Language model | Fine-tune Qwen2.5 with LoRA from a chat dataset |
| Long runs | Keep a run going in the background, or run it as a task file |
| Pricing and limits | How GPU sandboxes are billed, and the limits today |