ov-layers/skills/unsloth-studio/SKILL.md
Unsloth Studio fine-tuning web UI on ports 8888/8000 with vLLM inference. Tier 2 environment-owner meta-layer composing llama-cpp + unsloth, owns pixi.toml. Use when working with Unsloth Studio, the fine-tuning web UI, or the unsloth-studio image.
npx skillsauth add overthinkos/overthink-plugins unsloth-studioInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
| Property | Value |
|----------|-------|
| Dependencies | cuda, supervisord |
| Sub-layers | llama-cpp, unsloth |
| Ports | 8888 (Studio UI), 8000 (vLLM API) |
| Volumes | workspace -> ~/workspace |
| Service | unsloth-studio (supervisord) |
| Install files | layer.yml, pixi.toml |
This layer owns the pixi.toml for the fine-tuning environment and composes two Tier 1 layers via layers: [llama-cpp, unsloth]. Build order: pixi environment → llama-cpp (binaries) → unsloth (vLLM wheel + unsloth pip + patch) → supervisord config.
| Variable | Value |
|----------|-------|
| NVIDIA_PYTHON_PROJECT | ~/.pixi |
| LD_LIBRARY_PATH | /usr/lib64:$HOME/llama.cpp |
Plus from sub-layers: LLAMA_CPP_PATH, UNSLOTH_SKIP_LLAMA_CPP_INSTALL, HF_HOME
Fine-tuning focused ML stack: PyTorch (CUDA 13.0), xformers, transformers, accelerate, vLLM runtime deps, HuggingFace (datasets, tokenizers, sentencepiece), fine-tuning (peft, trl, bitsandbytes, liger-kernel), GGUF tools
Runs pixi run start-studio which executes unsloth studio -H 0.0.0.0 -p 8888. The Studio launches its own vLLM API server on port 8000 for inference and synthetic data generation.
/ov-images:unsloth-studio/ov-layers:llama-cpp — Sub-layer: llama.cpp binaries/ov-layers:unsloth — Sub-layer: vLLM + unsloth pip install + patch/ov-layers:supervisord — Process manager dependency/ov-layers:jupyter-ml — Alternative: ML Jupyter with MCP (same Tier 1 sub-layers)/ov-layers:python-ml — Alternative: core ML without UIUse when the user asks about:
/ov:layer — layer authoring reference (layer.yml schema, task verbs, service declarations)/ov:test — declarative testing (tests: block, ov image test, ov test)tools
OpenCharly CLI (charly) binary installed into container/VM images for in-container use. Use when working with charly binary deployment inside containers, native D-Bus support, or the full charly toolchain (charly binary + virtualization + gocryptfs + socat).
development
Operator CachyOS workstation profile — a kind:local template + target:local deploy that installs the full dev stack (30 candies) onto a CachyOS host via ShellExecutor. Lives in the overthinkos/cachyos submodule. MUST be invoked before editing or applying the charly-cachyos workstation profile.
tools
Fedora box with the full charly toolchain using shared candies. Rootless-first — runs as uid=1000 with passwordless sudo (no root, no cap_add: ALL). Same candy list as charly-arch. Includes NVIDIA GPU runtime. MUST be invoked before building, deploying, configuring, or troubleshooting the charly-fedora box.
tools
Arch Linux box with the full charly toolchain. Rootless-first — runs as uid=1000 with passwordless sudo (no root, no cap_add: ALL). Composes /charly-coder:charly-mcp so the box is reachable as an MCP gateway on port 18765. NVIDIA GPU runtime composed in. MUST be invoked before building, deploying, configuring, or troubleshooting the charly-arch box.