Skip to content

stuff overlay

A Gentoo ebuild overlay that ships hard-to-package software as first-class Portage citizens — dependency-resolved, installed and tracked by Portage, and built from source where upstream permits. Prebuilt components remain explicit in package metadata and these guides; many use -bin names or binary-specific licenses instead of being hidden inside an unmanaged environment.

stuff is a multi-niche overlay. Its primary focuses are: local AI & GPU compute (AMD Ryzen-AI / NPU · AMD ROCm · NVIDIA CUDA, with LLM runtimes and the PyTorch / ONNX ecosystem) · materials science (SAXS / SANS / XAFS / XRD / Rietveld · electron microscopy · SPM · micromagnetism) · pf-sources (curated pf-kernel patchset) · DeaDBeeF plugins · TeX Live · a Qt5 revival mirror · a Python 2 legacy preservation layer.

For readers coming from other distributions, the advantage is that the overlay extends Gentoo's dependency graph instead of bypassing it. Source-capable components build against the machine's selected hardware and libraries, with optional features tailored through USE flags. Prebuilt payloads remain explicit in package metadata and under Portage's control, but their upstream-fixed feature sets naturally offer less build-time tailoring. Portage upgrades, rebuilds, or removes the declared chain through the same interface as the main Gentoo repository.

Every ebuild is build-tested before it lands, using the relevant supported path: amd64 systems cover the ROCm and CUDA build configurations, while ebuilds carrying ~arm64 are validated on arm64. Hardware-dependent runtime checks are performed where matching devices are available.

Get started: enable the overlay and accept its keywords, then pick a guide below. The Gentoo Wiki Overlay:Stuff page is the neutral enable reference.

Guides

Step-by-step walkthroughs for the overlay's headline stacks:

  • The AI / ML stack — orientation map: which runtime for which hardware (NPU / ROCm / CUDA / CPU).
  • Check the GPU hardware version — detect the AMD gfx* target or NVIDIA compute capability and configure Portage's build variables.
  • AMD Ryzen AI NPU — the XDNA driver plus an NPU-first LLM runtime; from a stock install to a model running on the NPU.
  • vLLM — high-throughput serving on CPU, NVIDIA CUDA, or AMD ROCm, with backend-specific build and host setup.
  • llama.cpp, Ollama & llama-swap — the lighter local inference path across CPU, CUDA, ROCm, and Vulkan, plus model routing with llama-swap.
  • Fine-tuning, Unsloth & RAG — QLoRA fine-tuning (bitsandbytes / peft / trl), packaged Unsloth + Studio, and a local Qdrant vector store.
  • ComfyUI + TRELLIS — node-based generative-AI workflows, plus the TRELLIS image-to-3D pipeline (Gaussian-splat / mesh).
  • Qt5 LTS — the 5.15.19 LTS mirror and KDE patch collection, so Qt5 apps keep building after ::gentoo drops the slot.
  • TeX Live — the current TL2025 / TL2026 collections kept ahead of ::gentoo, with the LaTeX build tooling.

What's in the overlay

An overview of the major areas. The README carries the authoritative, current package list; this page is the orientation.

AI & GPU compute

NPU-first LLM tooling for AMD Ryzen AI (fastflowlm, lemonade, the amdxdna driver and XRT runtime); backend-agnostic local LLM serving (llama.cpp, ollama, vllm, koboldcpp, the llama-swap proxy, plus CLI and browser clients); and a fine-tuning + RAG stack (bitsandbytes, peft, trl, unsloth, and the qdrant vector database). Also included: a speech / audio ML stack (Whisper, sherpa-onnx, pyannote diarization); the surrounding PyTorch / ONNX ecosystem (Lightning, torchmetrics, onnxruntime, FAISS); the source-built ROCm / HIP 10.0 line (with 7.2.4 retained as a rollback anchor) ahead of ::gentoo; and NVIDIA CUDA, RAPIDS GPU dataframes, and a 3D-generative (Gaussian-splat / mesh) cluster. Start with the AI / ML stack map, the fine-tuning & RAG guide, or the ComfyUI + TRELLIS guide for image-to-3D.

Materials science

Electron microscopy — the full HyperSpy / 4D-STEM stack, plus gwyddion for SPM (AFM / STM); small-angle and X-ray-absorption analysis (mantid, sasview, bornagain, xraylarch, demeter); micromagnetism (mumax, oommf, vampire); and X-ray powder diffraction / Rietveld refinement (bgmn, profex).

Revival mirrors & extras

A Qt5 5.15.19-LTS mirror (excluding QtWebEngine) with the KDE Qt5 Patch Collection, so Qt5 consumers keep building — see the Qt5 LTS guide; a current TeX Live kept ahead of ::gentoo; a Python 2 preservation layer for gwyddion's pygwy bindings; twenty-six DeaDBeeF plugins; the pf-sources kernel; and a long tail of niche tools (XMPP clients, collaborative editing, 3D-printing, SuiteSparse, retro / fun). The full list lives in the README.

Overlay mirrors

Auto-mirrored on every push:

Versions move quickly. After synchronizing the repositories and refreshing the eix cache, eix --in-overlay stuff shows the packages and versions currently provided by the overlay.