Daytona Sandbox GPU Image

Dockerfile defines daytonaio/sandbox-gpu, a GPU-enabled sandbox image for x86 GPU hosts (e.g. H100).

It is a superset of the default daytonaio/sandbox image — it builds FROM daytonaio/sandbox and layers the GPU stack on top, so everything in the standard sandbox (Python, Node, language servers, the computer-use/VNC tooling, the default Python/agent packages, the daytona user) is present, plus GPU support.

NOTE

This image is amd64-only — there is no arm64 variant.

Added on top of the base image

CUDA / GPU runtime:

• CUDA 13 toolkit (nvcc), installed via the NVIDIA runfile
• PyTorch (CUDA 13 build) — torch, torchvision, torchaudio
• vLLM, with FlashInfer kernels pre-staged so first-serve cold-start stays fast
• FlashAttention (bundled with vLLM), cuDNN

GPU / ML / fine-tuning:

• accelerate
• datasets
• safetensors
• peft
• bitsandbytes
• trl
• einops
• sentencepiece
• sentence-transformers
• nvitop

Experiment tracking:

• wandb
• tensorboard

Interactive:

• jupyterlab
• ipykernel
• ipywidgets