Launch a GPU workspace with PyTorch, TensorFlow, or the CUDA stack ready to go — then open it in JupyterLab, VS Code, or a terminal. Pick a template and you're training in minutes.
Save a GPU workspace as a template and hand it to a labmate — everyone starts from the same PyTorch, the same CUDA, the same packages. Your conda env and dotfiles persist across every restart.
A web console for the shared GPU in the corner — launch your PyTorch or CUDA environment, keep it across restarts, and see who's on the card right now.
Templates ship a CUDA / cuDNN base with Conda and JupyterLab. Bring the rest with pip or conda — or build your own image.
A workspace is a deep-learning environment you launch in a click, keep across restarts, and hand to the whole lab — bounded by fair CPU, GPU, and memory limits.
Pick a PyTorch, TensorFlow, or CUDA template and it launches GPU-ready — open it in JupyterLab, VS Code, or a terminal. Bring or build your own image whenever you need to.
Per-workspace home keeps conda envs, dotfiles, pip installs, and shell history across every stop and start. /work survives re-creation; the HF cache downloads once.
Export a workspace template to a labmate, mount shared read-only datasets, and expose your own Gradio or Streamlit through a named LabPod port.
Give a user a whole GPU, or split one card by memory and compute so several share it at once — on any NVIDIA card.
Per-user CPU, memory, and GPU quotas, enforced by cgroups — not an honor system. Disk and image storage are tracked and shown. LabPod decides how much; you decide what to run.
Rootless Podman per Linux account — no Docker group, no root-equivalent access. Capability drop, no-new-privileges, localhost bind.
Start simple with whole-GPU passthrough; add fractional sharing when the lab grows. The same workspace, different boundaries.
One or more GPUs assigned to a user, with locks so nothing double-books. Single-GPU training and multi-GPU DDP both work out of the box.
Share a single GPU between several researchers by capping memory and compute per workspace — on any NVIDIA card. Where the hardware supports partitioning, slices are fully isolated.
Who holds which GPU, how much memory is live, which workspaces look idle, how much image storage each user takes — on one page.
Research images often have a shell but no web terminal — or none of the usual launchers at all. LabPod injects a minimal, read-only runtime so a browser terminal and live monitoring just work.
One idempotent script sets up Podman and the LabPod service. Then your lab logs in and starts launching workspaces.
$ curl -fsSL labpod.ai/install.sh | sudo bash