svngoku/unsloth-studio repository overview

⁠Unsloth Studio — GPU image

Unofficial, community-built Docker image for Unsloth Studio⁠ on NVIDIA GPUs.

Why this exists: the official unsloth/unsloth⁠ image has not been updated in months. This image is rebuilt weekly from unsloth main, so it follows upstream instead of rotting.

• Image: svngoku/unsloth-studio:latest (plus dated tags, e.g. :20260811)
• Source: https://github.com/svngoku/unsloth-studio-image⁠
• Base: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04 — devel, so llama.cpp can source-build via nvcc when no prebuilt exists
• Arch: linux/amd64

⁠How it works

The image contains only system deps + a shallow clone of unslothai/unsloth. On first container start, the entrypoint runs the project's own install.sh (creates the Studio venv + base install) followed by studio/setup.sh --local (overlays the cloned unsloth as editable + unsloth-zoo from git main) — because a Docker build has no GPU, but a running container does, and the installer auto-selects the correct torch CUDA wheel, llama.cpp backend, and CUDA arch for the GPU it finds. The install persists on a volume, so it happens once, not per boot.

⁠Run

docker run --gpus all -p 8888:8888 \
  -v unsloth-home:/root/.unsloth \
  -v hf-cache:/root/.cache/huggingface \
  -e HF_TOKEN=$HF_TOKEN \
  svngoku/unsloth-studio

Then open http://<host>:8888.

• First boot: full install (torch, venv, Node, frontend build, llama.cpp) — expect several minutes.
• Later boots: no-op fast path via the unsloth-home volume. Drop the volume and it reinstalls every boot.
• STUDIO_PORT env var changes the port (default 8888).

⚠️ Anyone who can reach the port can run code — put it behind auth/firewall on public instances, or use -H 0.0.0.0 --secure / --cloudflare options from unsloth studio --help.

⁠Update

docker pull svngoku/unsloth-studio:latest
# recreate the container with the same volumes

The weekly rebuild tracks unsloth main; setup.sh's update paths migrate an existing volume install.

⁠Requirements

• NVIDIA GPU + drivers, with the NVIDIA Container Toolkit (--gpus all) installed
• AMD/ROCm and CPU are not covered by this image

⁠Not affiliated with Unsloth AI

This is a wrapper around the official install script, not a fork. Studio itself is AGPL-3.0 — see studio/LICENSE.AGPL-3.0 in upstream.