stuff

A Gentoo ebuild overlay that ships hard-to-package software as first-class Portage citizens — dependency-resolved, built from source where practical (or packaged from a named upstream binary distribution when that is the maintained packaging route), and managed with emerge like everything else on the system, instead of through manual git builds or conda environments that leave the package manager blind to what's installed.

No single flagship — a curated, multi-niche overlay. The front-door slices:

• 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 — the full pf-kernel patchset, GA-frozen with CVE backports, plus a stable-tracking pf-sources-extended tier.
• DeaDBeeF plugins, TeX Live, a Qt5 revival mirror, and a Python 2 legacy preservation layer.

With app-eselect/eselect-repository installed:

eselect repository enable stuff
emerge --sync stuff

Packages are primarily ~amd64, with growing ~arm64 coverage. Stable users should accept testing keywords only for the packages they intend to install; the setup guide covers the overlay-wide form and how to scope it per package.

See Overlay:Stuff on the Gentoo Wiki for enabling the overlay and a package-area overview. The overlay documentation provides setup, hardware-selection and stack-specific guides.

Mirrors

Auto-mirrored on every push:

• github.com/istitov/stuff (primary)
• gitlab.com/istitov/stuff
• codeberg.org/istitov/stuff

AI & GPU compute

AMD Ryzen-AI / NPU stack

NPU-first LLM tooling for AMD Ryzen AI (XDNA / XDNA2). Application layer plus the driver and runtime it needs:

• sci-ml/fastflowlm — NPU-first LLM runtime (chat, model pull).
• sci-ml/lemonade — AMD Lemonade SDK.
• sci-ml/kokoros — Kokoro TTS server (Rust + Python).
• sci-ml/amd-gaia — AMD GAIA stack with api / audio / eval / image / mcp / talk / ui USE flags.
• dev-libs/xdna-driver, dev-libs/xrt-xdna, dev-util/xrt — NPU driver and the XDNA-extended Xilinx Runtime.

Local LLM tooling

Backend-agnostic servers, model-swap proxy, CLI clients, and web UI. Pairs with fastflowlm / lemonade above as well as vllm and any other OpenAI-compatible endpoint:

• sci-misc/llama-cpp — llama.cpp server / runtime. Bundled web UI provisioned at configure time (default on; disable with USE=-webui).
• sci-ml/ollama — the Ollama model-server (CUDA / ROCm / CPU backends), with acct-user/ollama + acct-group/ollama and systemd / openrc service files.
• sci-misc/llama-swap — Go HTTP proxy that lifecycle-manages multiple inference backends and routes OpenAI/Anthropic-compatible requests to the right one. Optional embedded Svelte UI via USE=ui; vendored Go modules on extra-stuff.
• www-apps/hollama — Minimal chat UI (SvelteKit + Node, browser-localStorage state, no server-side persistence). Talks to Ollama natively and any OpenAI-compatible endpoint. systemd unit + openrc service files; loopback-only by default.
• dev-util/aichat — All-in-one LLM CLI (Chat-REPL, shell assistant, RAG, agents); multi-provider, single Rust binary.
• dev-util/rtk — "Rust Token Killer" CLI proxy that filters dev-command output (cargo, npm, pytest, …) before it reaches your LLM session, cutting token consumption.
• dev-util/argc — Bash CLI framework + Argcfile.sh task runner; infrastructure for sigoden's tooling cluster.
• dev-util/codex — OpenAI's terminal coding agent, built from its vendored Rust dependency set.
• app-misc/qdrant — vector database and similarity-search service for local RAG workloads, with systemd and OpenRC integration.
• sci-ml/lm-eval (lm-evaluation-harness), sci-ml/evalplus (HumanEval+ / MBPP+), and sci-ml/bigcode-eval — harnesses for benchmarking local language / code models.

Speech / audio ML stack

ASR, speaker diarization, and audio DSP packages:

• app-accessibility/whisper-cpp — Whisper.cpp offline ASR.
• sci-ml/sherpa-onnx + sci-ml/sherpa-onnx-bin — Next-gen-Kaldi ONNX ASR / TTS / speaker-id (source build and prebuilt-wheel flavour).
• sci-ml/pyannote-audio plus the pyannote-{pipeline,metrics,database,core} chain and sci-ml/pyannoteai-sdk — speaker-diarization toolkit.
• sci-ml/torch-audiomentations + sci-ml/torch-pitch-shift + sci-ml/asteroid-filterbanks
  • sci-ml/julius — neural-net friendly audio DSP / augmentation building blocks.

PyTorch / ONNX ecosystem

The overlay carries a split PyTorch stack: sci-ml/caffe2 builds and owns the C++/libtorch side plus torch._C, while sci-ml/pytorch installs the matching Python package without rebuilding libtorch. Each supported line is joined by a matching sci-ml/torchvision; CUDA, ROCm and CPU feature sets remain selected through the Caffe2 ebuild.

Other general-purpose ML infrastructure not covered by ::gentoo, pulled in alongside the speech stack above and broadly useful on its own:

• sci-ml/lightning + sci-ml/lightning-utilities — PyTorch Lightning training framework.
• sci-ml/torchmetrics — metric collection for PyTorch training loops.
• sci-ml/pytorch-metric-learning — embedding / metric-learning utilities.
• sci-ml/torchcodec — PyTorch video / audio decoder.
• sci-libs/onnxruntime, sci-libs/dlpack, dev-cpp/safeint — the ONNX inference engine plus its in-memory tensor-exchange / overflow-safe-int building blocks.
• dev-python/optuna — hyperparameter optimization.
• sci-libs/faiss — FAISS: efficient similarity search and clustering of dense vectors (the vector-index building block for embeddings / RAG).

ROCm stack

An unmasked, source-built ROCm / HIP 10.0 cohort ahead of ::gentoo: AMD's LLVM 23 fork (llvm-core/rocm-llvm), the runtime and compiler, math / communication libraries (rocm-core, hip, hipBLAS / rocBLAS, hipFFT / rocFFT, MIOpen, composable-kernel, rccl, …), debugging and profiling tools, media libraries, and rocm-smi / rocminfo. The 7.2.4 line is kept as the last 7.x rollback anchor.

dev-util/therock-bin is a /opt-installed ROCm SDK sliced out of AMD's official ROCm runfile installer for a per-host AMDGPU_TARGETS. It coexists with the split /usr stack above and remains the lower-build-cost alternative because it bundles AMD's compiler and libraries. Nvchecker follows upstream's therock-<major.minor> tag line; the distribution carries official gfx1150/gfx1151 support.

dev-util/zluda — a drop-in CUDA runtime for AMD GPUs (an honorable mention as it is ROCm-bound).

CUDA / generic ML inference

NVIDIA CUDA runtime, Python bindings, and generic ML inference: dev-python/cuda-bindings, dev-python/cuda-python, dev-python/cuda-pathfinder, dev-python/cuda-tile-bin, dev-python/pycuda, dev-util/nvidia-cuda-toolkit, and dev-python/vllm (optional USE=humming pulls dev-python/humming-kernels for MXFP4 quantization).

RAPIDS GPU computing

GPU-accelerated dataframes and distributed compute, carried on two parallel RAPIDS release lines and not in ::gentoo: dev-python/rmm + dev-python/librmm (GPU memory manager), dev-python/dask-cuda (multi-GPU Dask clusters), plus dev-python/rapids-logger and dev-python/rapids-dask-dependency.

Generative-AI applications

• media-gfx/comfyui — node-graph generative-media application, packaged with its frontend, workflow-template family and optional CUDA / ROCm / CPU PyTorch backend.
• sci-ml/unsloth — model fine-tuning stack, with an optional system-packaged Studio backend. sci-ml/unsloth-desktop-bin remains masked because the vendor binary self-bootstraps a private Python/PyTorch environment outside Portage.
• sci-misc/stable-diffusion-cpp and sci-misc/koboldcpp — native local inference frontends for image and language models.

3D generative (Gaussian splatting / mesh)

Image-to-3D and Gaussian-splat / mesh generation — the stack behind the TRELLIS pipeline, fronted by media-gfx/comfyui-if-trellis: dev-python/diff-gaussian-rasterization and dev-python/diffoctreerast (differentiable rasterizers), dev-python/nvdiffrast (differentiable rendering), dev-python/spconv-cu126 (sparse convolution), dev-python/xatlas + dev-python/pymeshfix + dev-python/plyfile (mesh UV-unwrap / repair / I/O), and dev-python/vox2seq + dev-python/utils3d helpers.

Materials science

Electron-microscopy stack

The full HyperSpy stack — core, GUI backends, I/O, and domain extensions (exspy, atomap, pyxem, py4dstem).

sci-visualization/gwyddion + sci-visualization/gwyddion3 — the de-facto standard open-source tool for SPM (AFM / STM) data visualization and analysis.

SANS / SAXS / XAFS analysis

• sci-physics/mantid — SANS reduction and analysis. Installs under /opt/mantid and keeps building against the current ::gentoo by carrying a few local deps (see the Qt5 revival mirror section below).
• sci-physics/sasview + dev-python/sasmodels
  • dev-python/bumps + dev-python/periodictable — SAS modeling and fitting.
• sci-libs/ausaxs + dev-python/pyausaxs — AUSAXS solvent-scattering calculator and its Python bindings.
• sci-physics/bornagain — simulate and fit X-ray / neutron reflectometry and grazing-incidence small-angle scattering (GISAS / GISANS), with its sci-libs/libformfactor + sci-libs/libheinz libraries.
• sci-physics/xraylarch — XAFS analysis; modern replacement for the discontinued ifeffit.
• sci-physics/demeter — classic Athena/Artemis XAFS GUIs (Perl).

Micromagnetism

sci-physics/mumax (GPU finite-difference, Go + CUDA), sci-physics/oommf (Tcl/Tk reference implementation), and sci-physics/vampire (atomistic spin dynamics).

Rietveld refinement / crystallography

X-ray powder diffraction and Rietveld refinement: sci-physics/bgmn (the BGMN refinement engine) and sci-physics/profex (Profex, its Qt6 GUI front-end). Plus crystallography / atomistic tooling: sci-physics/prismatic (STEM image simulation), sci-libs/nexus (NeXus neutron / X-ray data format), and sci-libs/pycifrw (CIF read/write).

TeX Live

Current TeX Live, kept ahead of ::gentoo's stabilized line: the full dev-texlive/* collection set at the TL2025 and TL2026 releases (recent tlpdb revisions), with upstream package additions, removals, and relocations tracked per release — plus dev-tex/{biber, biblatex, latexmk, minted, pgf, tex4ht, …} build tooling.

Qt5 revival mirror

::gentoo last-rited the entire dev-qt:5 set on 2026-05-15 (bug #948836) and has since removed it, but sci-physics/mantid and a few others still need Qt5. This overlay carries the full 23-module dev-qt/* slot:5 set at v5.15.19-lts-lgpl — the KDE Qt5 Patch Collection applied via the local qt5-build.eclass, distfiles mirrored on extra-stuff — plus the revived dev-python/pyqt5 and the last Qt5-compatible x11-libs/qscintilla, all unmasked in profiles/package.unmask so they stay installable.

Drop the mirror once mantid finishes its Qt6 port and the other consumers follow.

Also here

• XMPP clients — net-im/profanity, net-im/stabber, net-im/xmppconsole, dev-libs/libstrophe.
• Collaborative editing — app-editors/gobby, net-libs/libinfinity, acct-{group,user}/infinote.
• Kernel / low-level — sys-kernel/pf-sources, the full pf patchset on a GA-frozen base, with surgical CVE backports where pf's deliberate no-linux-stable policy leaves a gap; sys-kernel/pf-sources-extended, the complementary tier — vanilla plus Gentoo genpatches, so linux-stable is tracked, carrying a curated subset of the pf delta rebased on top; sys-apps/dkms-gentoo, sys-kernel/kernel-cleaner.
• 3D printing — media-gfx/orcaslicer, the open-source slicer (a PrusaSlicer / Bambu Studio fork), with x11-libs/wxGTK 3.3 pulled ahead of ::gentoo's 3.2 (and a local wxwidgets.eclass that accepts the new 3.3-gtk3 slot) to build it.
• SuiteSparse imports — sci-libs/{amd,camd,cholmod,colamd,ccolamd,umfpack,suitesparseconfig}.
• Retro / fun — x11-terms/cool-retro-term, games-roguelike/adom, games-roguelike/dwarftherapist, games-misc/fortune-mod-lorquotes.
• Niche tools — dev-lang/tcc, dev-vcs/fossil, app-office/mytetra, app-misc/tudu, app-misc/mc6 (Midnight Commander fork with panel plugins; installs /usr/bin/mc and therefore blocks app-misc/mc), sys-fs/google-drive-ocamlfuse, app-text/pandoc-crossref-bin, app-portage/portconf (/etc/portage cleaner; forked to istitov and bumped to the 2.x series in this overlay), and app-portage/udept (a Portage analysis toolkit, likewise forked to istitov).
• DeaDBeeF plugins — twenty-six media-plugins/deadbeef-* packages: audio-format support, visualization, playback / session control, file browsing, desktop integration, and output plumbing.
• Python 2 preservation — a small py2 surface kept alive for sci-visualization/gwyddion's pygwy bindings: locally-vendored *_py2 eclasses, py2 forks of core libs (dev-python/numpy-python2, certifi-python2, pycairo-python2, …), and py2-only legacy packages (pygobject-2.28.6, pygtk-2.24.0, unittest-or-fail).
• Masked but kept — legacy, rollback and opt-in packages. Examples include net-misc/ipx-utils (IPX left the kernel in 4.18), app-portage/portopts (upstream dormant since 2014), and sci-ml/unsloth-desktop-bin (self-bootstraps a private runtime outside Portage). Every entry in profiles/package.mask carries a dated comment saying why it is masked and what would lift it.

Repository layout and conventions

• Thin manifests, masters = gentoo only. Every package depends on ::gentoo being enabled.
• Profiles under profiles/ follow standard PMS layout.
• Keywording — packages are primarily ~amd64; ~arm64 keywording is being rolled out across the tree.
• Patches live in <category>/<package>/files/ and are applied via PATCHES=() or src_prepare().
• Commit messages use subject + body form (72-char subject, blank line, rationale hard-wrapped at 72). Single-line messages only for truly trivial edits.
• metadata/pkgcheck.conf documents which checks are suppressed and why (not just that they're suppressed).
• CI gates every pull request and relevant push to master, and runs scheduled repo-wide and staleness scans on top. CONTRIBUTING.md is the canonical description — this bullet deliberately does not restate it, so the two cannot drift apart.
• Documentation — CONTRIBUTING.md (house-style checklist, AI/LLM disclosure expectation), SECURITY.md (vulnerability reporting — GitHub private advisories, or mail the maintainer), CODE_OF_CONDUCT.md.
• News items — eselect news read after emerge --sync stuff surfaces GLEP-42 announcements for migrations, mask windows, and CVE-sensitive notices. Items live in metadata/news/.
• Upstream version tracking — scripts/nvchecker/ holds the generated nvchecker.toml, its generator and audit helpers, and a local cron runner. Pull requests and relevant pushes verify that the generated config is current; a weekly CI job at .github/workflows/nvchecker.yml runs the same config against the tree as baseline, detects entries that return no upstream version, uploads a 30-day artifact and maintains a rolling drift issue.

Credits

Originally created by @megabaks and now maintained by Ivan S. Titov; see the contributors list for everyone who has contributed since. Thank you.

License

Original packaging files — ebuilds, eclasses, metadata.xml, profiles, and news items — are distributed under the GNU General Public License v2, matching ::gentoo's per-file header convention and its Copyright Policy (GLEP 76).

Patches that modify upstream code are derivative works of it and carry that code's own license, as do the upstream sources themselves; the applicable license is the one declared in each ebuild's LICENSE variable — e.g. the dev-python/py4dstem numpy-2 patch is GPL-3 and the media-libs/opencv patches are Apache-2.0, not GPL-2. This mirrors GLEP 76's Certificate of Origin, under which a contribution "based upon previous work" is submitted "under the same free software license" as that prior work.