sci-ml/strata
Inference engine and terminal for Qwen3.8-Flash-Next
-
strata-0.2.15~amd64+cuda hip +terminal +vision python_single_target_python3_12 python_single_target_python3_13 python_single_target_python3_14 +amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1030 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 amdgpu_targets_gfx803 amdgpu_targets_gfx900 amdgpu_targets_gfx906 amdgpu_targets_gfx940 amdgpu_targets_gfx941 amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1031 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151
View
Download
Browse License: MIT Overlay: ruizhi-overlay -
strata-0.2.14~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.13~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.12~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.11-r1~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.11~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.10~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.9-r1~amd64+vision amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.9~amd64amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.8~amd64amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.7~amd64amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.6-r1~amd64amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers -
strata-0.2.6~amd64amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1030 amdgpu_targets_gfx1031 amdgpu_targets_gfx1032 amdgpu_targets_gfx1034 amdgpu_targets_gfx1100 amdgpu_targets_gfx1101 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 amdgpu_targets_gfx1152 amdgpu_targets_gfx1153 amdgpu_targets_gfx1200 amdgpu_targets_gfx1201
View
Download
Browse License: MIT Overlay: bennypowers
ChangeLog
commit b94328195d760aefc0b2b888ea7036357ce5b1c5
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Mon Nov 16 23:20:51 2026 +0800
sci-ml/strata: bump to 0.2.15
The new release builds the engine without llama.cpp's MMQ prompt kernels
(STRATA_MMQ defaults to OFF), so pass that value explicitly as upstream's
release packaging does. The picture reader and the engine still build from
the pinned llama.cpp tree.
The portage-managed guard needed its update_cmd hunks redrawn: run() gained a
nested fail() helper and the component-dir logic moved. Re-derive both hunks
against the 0.2.15 tarball and add the new draft_vocab_latin.bin data file.
commit 53277b67e88e9adeac543025f357c4f10903fcf5
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Wed Nov 11 23:19:46 2026 +0800
sci-ml/strata: drop the kvmem-llama-cpp references in -r1
sci-misc/kvmem-llama-cpp has been removed from the overlay, so the elog
message and files/README.gentoo no longer name it. The published revision
moves into -r1 with those text changes. The build and the install layout
are unchanged, and the redaction patch still applies in prepare.
The CI install skip list gets an entry for strata, since its install needs
cmake, ninja and, with the default +cuda USE, the CUDA toolkit.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit dfcc93423102b994f8e3061ec4476d81e456661e
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Tue Nov 10 20:16:24 2026 +0800
sci-ml/strata: build the CUDA engine with the CUDA host GCC
With a CUDA build, nvcc's host compiler (CUDAHOSTCXX) is the newest GCC
that CUDA supports, which can be older than the default GCC. CMake adds
that GCC's library directory to the link line, but the C and C++ objects
were compiled with the default GCC's headers, so the engine did not link:
std::__detail::__notify_impl and __wait_impl come from the newer
libstdc++ only.
Export CC and CXX from the same GCC directory as CUDAHOSTCXX, both for
the engine's configure and for the picture reader's configure in
src_compile, so every object and the link use one GCC. HIP builds are
unchanged.
Verified on the build host (CUDA 12.9.86, default GCC 16, GTX 1080 Ti):
the engine and the CUDA picture reader compile and link, and
package_engine.py writes the component folder with its manifest. The
architecture was set to sm_75 only in a test copy of the ebuild, because
the host GPU is Pascal and upstream refuses sm_61. The packaged binaries
run `strata --version` on this host. Inference on a GPU was not run.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit fb492ed964abef4d4d24f25836b27aac0c9ef6f6
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Tue Nov 10 20:01:55 2026 +0800
overlay: bump packages from issue #36
Bump the packages reported in issue #36 to their current upstream
versions:
- media-sound/sidra-bin 1.2.1 -> 1.3.0
- app-misc/chatgpt-desktop 26.1103.32239 -> 26.1103.60933
- app-misc/claude-desktop 4.23118.0 -> 4.23118.1
- sci-ml/strata 0.2.12 -> 0.2.13.1
The sidra and desktop app distfiles match the SHA256 digests published
for them (the Debian package indexes and the GitHub asset digests), and
their install phases were checked on amd64. The RDEPEND lists already
cover the dependencies of the new packages.
Strata follows the newest upstream tag, 0.2.13.1, rather than the 0.2.13
release, because the release list does not show the patch tag; Strata
versions follow the tags (see AGENTS.md). Its ggml pin, 3cf03257, is
unchanged. The redaction patch was re-derived against the new tree: the
same policy text, its hunks at their new line positions, and it applies
with --fuzz=0 at zero offset.
Not verified: the Strata engine does not link on this build host. It was
configured with a test-only sm_75 architecture, because the host GPU is
Pascal and upstream refuses sm_61. The link pulls GCC 14's libstdc++
(CUDA 12.9's host compiler reaches the link through CMake's CUDA implicit
link directories) while the objects are compiled with GCC 16 headers, so
std::__detail::__notify_impl and __wait_impl are undefined. The ebuild's
CUDA compiler setup is unchanged from 0.2.12, and 0.2.12 uses the same
atomic wait/notify code, so this predates this bump.
Closes #36
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit 53c42444a316899b153e86b6c68c9e6247e2fc54
Author: user.email <ruizhi.zhong@outlook.com>
Date: Sat Nov 7 17:45:07 2026 +0800
overlay: bump Strata and ChatGPT Desktop from issue #34
sci-ml/strata 0.2.11 -> 0.2.12
app-misc/chatgpt-desktop 26.1103.31037 -> 26.1103.32239
Strata 0.2.12 is the newest tag as well as the newest release, so no tag-only
successor was missed. third_party/ggml/VERSION.txt still pins 3cf0325, so
LLAMA_COMMIT and the llama.cpp distfile are unchanged, and every interface the
ebuild drives (package_engine.py, stage_tuning.py, the CMake options for the
engine and tools/vision, tools/hip) is still present. The redaction patch
applies at fuzz=0 and still guards every writer of the component folder: the
0.2.12 cli only exposes read-only components subcommands, and the writing ones
go through update_cmd.
One file-set change needed a fix. 0.2.12 adds data/dflash2-tuned/ (the
Apache-2.0 text and the "modified by Strata" notice for the tuned drafter),
and strata_control/drafting_fetch.provide_licence() resolves it from the
install root to copy beside the drafter it fetches. The ebuild installs data/
file by file, so the directory was left out and the call would have silently
returned False, dropping the attribution required when redistributing that
Apache-2.0 drafter. It is installed now; upstream's own package_terminal.py
PAYLOAD omits it as well, so this is deliberately wider than their archive.
The new tools/force_check.py is a developer tool no installed module imports,
so TOOLS_SUBSET stays as it is.
ChatGPT Desktop only changes its version: the new deb still ships both Qt
shims, codex-launcher, the apparmor profile naming /usr/lib/chatgpt/ChatGPT,
and Exec=chatgpt %U, which the existing sed still rewrites.
commit 2fd413bdc5768f584aef9424b32fb54f54b54898
Author: user.email <ruizhi.zhong@outlook.com>
Date: Thu Nov 5 21:44:07 2026 +0800
sci-ml/strata: bump to 0.2.11
Engine 0.1.57, and the llama.cpp pin is unchanged
(third_party/ggml/VERSION.txt still 3cf03257), so LLAMA_COMMIT and the ggml
distfile stay as they are.
The portage-managed patch needed re-deriving. Upstream moved run() in
strata_cli/update_cmd.py and added source_store()/default_components_dir()
above it, so both hunks there no longer applied. Regenerated against 0.2.11:
the guard now sits after the fail() closure and reads the component folder
through default_components_dir(), which is what run() itself uses, and
`update --check` still passes through while `update`, `components
update|rollback|install-file|use|remove` refuse. The cli.py hunk only shifted.
The model catalogue grew from 17 to 23 selections (Q8_0 in six shards, the
Gyro TQ quants, more UD sizes); every one is still Qwen3.8-Flash-Next, so the
description and the note in README.gentoo stand.
Checked on amd64 with CUDA 13.4 and an RTX 5080: the patch applies with
--fuzz=0, the patched modules parse, `strata install` and `strata update`
refuse while `update --check` and `components engines` are allowed and
`components use --components-dir <user dir>` reaches its own logic, the engine
and the picture reader compile, package_engine.py accepts the real binaries,
and the component passes the terminal's full SHA-256 verification and runs.
commit ad2921156d1e2649dd9de4c927e44a2e30628913
Author: Ruizhi Zhong <ruizhi.zhong@outlook.com>
Date: Tue Nov 3 21:31:29 2026 +0800
sci-ml/strata: let update --check through the Portage guard
The 0.2.10 re-derivation of redact-portage-managed.patch put the update refusal
in front of everything, including `strata update --check` and `strata
components update --check`, which only read the signed feed and write nothing.
The patch header already said read-only operations were left alone, so the
guard and the documentation disagreed.
`--check` is now exempt like `components engines`; the writes in the component
folder stay refused (`update`, `rollback`, `install-file`, `use`, `remove`)
unless --components-dir names a folder the calling user can write. Checked by
calling run() with STRATA_PACKAGE_MANAGED=1: both update forms exit 5, both
--check forms exit 0 and print the feed, and the component-folder writers
refuse unless given an owned directory.
AGENTS.md now records why the patch is kept (the component folder belongs to
Portage, upstream's STRATA_NO_ACQUIRE does not cover the writers), that it is
gated by a launcher-set variable rather than being a sandbox, and that every
future command writing there must be added to it.
commit 2362000c96a1b8f7ad451e7aadb809c4e8171102
Author: Ruizhi Zhong <ruizhi.zhong@outlook.com>
Date: Tue Nov 3 21:13:21 2026 +0800
overlay: bump quarto-bin, chatgpt-desktop and strata
app-text/quarto-bin 1.11.9: straight version bump of the prebuilt tarball.
The layout is unchanged: bin/quarto and share/man/quarto-man.man are still
where the ebuild installs them.
app-misc/chatgpt-desktop 26.1027.51404: straight version bump. The deb still
carries usr/lib/chatgpt with the Qt5/Qt6 shims, codex-launcher and
usr/bin/chatgpt, the AppArmor profile still names /usr/lib/chatgpt/ChatGPT,
and the desktop entry still starts with Exec=chatgpt, so src_prepare's edits
and the install paths need no change. The prebuilds loop still keeps the
linux-x64 node modules and drops every other platform, including the musl
ones the next find deletes.
sci-ml/strata 0.2.10:
- third_party/ggml/VERSION.txt still pins llama.cpp 3cf03257, so LLAMA_COMMIT
stays put; the regenerated Manifest confirms that distfile is unchanged.
- redact-portage-managed.patch is re-derived against the new tarball: 0.2.10
rewrote the end of _print_check in strata_cli/update_cmd.py, which is where
the helper is inserted. `patch --dry-run --fuzz=0` applies cleanly and the
patched modules compile.
- 0.2.10 adds `strata components use|remove`, which switch the active engine
inside the component folder and delete engine packages from it. The patch
treats them like rollback/install-file: refused under
STRATA_PACKAGE_MANAGED=1 unless --components-dir names a folder the calling
user can write, because the system component folder belongs to Portage.
The read-only `engines` listing and `--check` are unaffected.
- requirements.txt, requirements-monitor.txt, the packaging helpers
(package_engine.py, package_terminal.py, stage_tuning.py) and the terminal
file set package_terminal.py stages are unchanged, as are the data files the
ebuild installs, its tools subset and the docs it names. Every CMake switch
and target the ebuild uses still exists; the engine itself is now 0.1.56,
which the ebuild reads from CMakeLists.txt.
Manifests regenerated with pkgdev manifest. pkgcheck reports no new findings;
xmllint and git diff --check are clean. No build was run locally.
commit ff40a925b41d0d8c868a227c9b81d0fec73160b6
Author: user.email <ruizhi.zhong@outlook.com>
Date: Sat Oct 31 20:56:48 2026 +0800
overlay: bump strata, datagrip, prism-llama-cpp and kvmem-llama-cpp
sci-ml/strata 0.2.9: the ggml pin in third_party/ggml/VERSION.txt is
unchanged, so LLAMA_COMMIT stays put. redact-portage-managed.patch is
re-derived against the new tarball; it applied with a nine-line offset and
now applies cleanly. requirements.txt and requirements-monitor.txt did not
change, and the terminal file set package_terminal.py stages is the same.
dev-util/datagrip 2026.2.6.1: straight version bump, every path the ebuild
sets permissions on still exists in the new tarball.
sci-misc/prism-llama-cpp 12128: upstream tag prism-b12128-d7b2896.
sci-misc/kvmem-llama-cpp 0.19.1:
- Upstream moved to kvmem/kvmem, so HOMEPAGE, the SRC_URI, metadata.xml
and the update checker follow.
- backends/versions.json now pins llama.cpp 620d222 and replaces the single
kvmem patch with an ordered five-patch series (kvmem, pq2, ptq1,
ptq1-performance, checkpoint-policy). scripts/prepare-backends.py needs the
submodule's git metadata to verify the series, which a tarball build does
not have, so the pinned tree is unpacked and the same series, in the same
order, is applied with eapply. All five apply with fuzz=0.
- KVMEM_ENABLE_NVME is gone: the NVMe store was retired upstream in favour
of the shared RAM/SSD tiers. Its build switch and its test target are
dropped; host_kv_storage_test and tiered_kv_storage_test cover the tiers
that replaced it.
- New backport: monitor_hardware.hpp includes <nvml.h>, whose one-argument
DEPRECATED(ver) clobbers llama.cpp's two-argument DEPRECATED(func, hint)
and breaks mtmd.h in llama-kvmem-server.cpp. The patch restores the
llama.cpp spelling after NVML is done with the name.
- README.gentoo: the old example passed --kvmem-gen-reserve 16384, which is
now documented as reusable generation workspace rather than an output cap.
It shows --kvmem-host-mib and the opt-in SSD pair instead.
Validated locally on amd64 with CUDA 13.4 and sm_120a: kvmem 0.19.1 builds
llama-kvmem-cli and llama-kvmem-server and its six model-free tests pass;
strata 0.2.9 builds its engine; prism 12128 configures. Manifest regenerated
with pkgdev manifest; pkgcheck reports no new findings.
commit 0f9b1b8cef4122d833faa92d604c370d8d617fe8
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 22:05:32 2026 +0800
sci-ml/strata: move the published packaging fixes into -r1
0.2.8 has been pushed, so the eleven commits since then are changes to a
published ebuild: they must not ride on the same PV. Create -r1 instead.
The revision matters more here than the wording suggests. With FEATURES
holding buildpkg and getbinpkg, a PV that was already built is served from
$PKGDIR/sci-ml/strata/strata-0.2.8-1.gpkg.tar, and Portage reported the
package as a binary reinstall - the corrected src_prepare, the CUDA
architecture handling and the model note would never have been built or shown.
Under -r1 the same emerge now reports an upgrade and compiles the corrected
ebuild.
PV is unchanged, so SRC_URI, the tarball name, S and the pinned llama.cpp tree
are untouched; only PF, and with it the component's build id, gains the
revision. The Manifest needs no change.
commit 35c1c14ff6fa2da6fb88fd7e0cc8b549628777f5
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 22:04:36 2026 +0800
sci-ml/strata: say which model the engine actually runs
The description claimed "Qwen3.8-Flash-Next and other GGUF models", and the
long description added DeepSeek-V4.1-Flash. Neither is true of this package.
The catalogue in data/gguf_fingerprints.json holds 17 selections under the
qwen, swift, coder, unsloth and gyro keys and all of them are
Qwen3.8-Flash-Next (qwen35moe) in GSQ-RCO, Q2_0 or UD quantizations, split
into two shards with exact upstream names; `strata models import` refuses
anything else, including other Qwen3.8 sizes. DeepSeek needs
STRATA_BUILD_DSV41=ON, which the ebuild does not build.
Verified against an installed terminal with a local GGUF library: a
Qwen3.8-27B GSQ-RCO IQ3_S file (arch qwen35) is refused with "this model type
is not supported by the Strata engine (add it as a Custom model API with
llama-server instead)", and so is the MTP drafter beside it, while the
catalogue's own entries are all Flash-Next.
Correct DESCRIPTION and metadata.xml, and point users at
sci-misc/kvmem-llama-cpp, sci-misc/prism-llama-cpp or guru's
sci-misc/llama-cpp for anything else in README.gentoo and pkg_postinst. Record
the constraint in AGENTS.md so a later bump does not reintroduce the claim.
commit d2382c0150bc798b2d00412679c25340a5354217
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:55:50 2026 +0800
sci-ml/strata: fix a comment naming an array as a function
commit 8ca83a04275a6f9bdebcd5b35f4b90e7d50245e1
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:55:43 2026 +0800
sci-ml/strata: bring the ebuild in line with Gentoo practice
A standards pass over the ebuild, rather than another build fix.
Declare dev-python/rich. The terminal imports rich in twenty places and it
arrived only as textual's dependency; an ebuild must name what its code
imports, or a textual change silently breaks it.
Compute the packaging architecture list in src_configure instead of
src_install. src_install ran nvidia-smi to work out what to write into
BUILD.json, which is a build decision made in the wrong phase and repeated a
third time. It now reuses the list that was already derived, and the fallback
notice fires once instead of per call.
Drop ~arm64 from KEYWORDS. The engine, the reader and the platform key all
support linux-aarch64, but nothing has merged on that hardware here, and a
keyword claims testing that has not happened.
Checked that the four architecture settings still resolve to one list shared by
CMake, BUILD.json's archs and vision_archs, that the fallback warning is
emitted once and goes to stderr, and that emergence still resolves.
commit 688d2bde3447fd8058469ad50fc676d45798d2af
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:53:27 2026 +0800
sci-ml/strata: keep the launcher's shell shebang
python_fix_shebang was run on /usr/bin/strata, which is a POSIX shell script
that finds an interpreter and execs strata.py. The helper only rewrites Python
shebangs and dies on anything else when a file is named explicitly, so the
install phase aborted after the engine had been packaged.
Drop the call; the shell shebang is what the launcher needs, and the modules
under /usr/share/strata are still byte-compiled by python_optimize. Checked
that the installed launcher keeps #!/bin/sh, stays executable, and reports the
engine build it was merged with.
commit 7e10a34ceee74bd558643e1c67e1b7ce0b33a147
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:43:32 2026 +0800
sci-ml/strata: allow the CUDA runtime soname when packaging
The allow-list derived from each binary's DT_NEEDED named the AMD sonames and
libcublas/libcublasLt, but not libcudart, so package_engine.py refused every
CUDA package: strata and strata-vision both need libcudart.so.13. My earlier
checks passed only because they spelled --allow-lib by hand, which is exactly
what the ebuild must not depend on.
Match the rest of the CUDA runtime sonames too (cufft, curand, cusparse,
cusolver, nvrtc, nvJitLink) so a future backend does not reintroduce this.
These stay system libraries: nvidia-cuda-toolkit, already in RDEPEND, puts
them on LDPATH through /etc/env.d.
Checked by deriving the arguments with the ebuild's own loop from the real
binaries and packaging with no hand-written --allow-lib - it succeeds, and the
package's strata and strata-vision both run from the unpacked component
folder. Without the allow-list the packager still refuses, so the check has
not been widened into uselessness.
commit 9a81447c982c3093a1cdb2fc07cef6e526c14b43
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:35:08 2026 +0800
sci-ml/strata: stop the HIP copy from failing a CUDA install
package_engine copied strata-device with
use hip && cp "$/strata-device" "${stage}/" || die
which bash associates left to right, so the failure of `use hip` ran the die.
Every CUDA install aborted in src_install, after the whole engine had compiled.
Group the copy under an if, so `use hip` being false skips the copy instead of
being read as its failure.
Only that line had the shape: the other three `use X && command` lines run a
function that dies on its own, so nothing was hidden.
Checked by reproducing the expression with USE lacking hip (the old form calls
die, the new one returns 0), and by running package_engine.py over a stage
built from this release with a CUDA engine and reader: it packages and the
HIP branch is skipped.
commit dbda4f6a96821433c22f46e851d58488997afe25
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:26:43 2026 +0800
sci-ml/strata: add the missing src_prepare
The ebuild declared PATCHES but never defined src_prepare(), so Portage
aborted the prepare phase with "eapply_user (or default) must be called in
src_prepare()". cmake_src_prepare() is what applies PATCHES and then calls
eapply_user, and without it the portage-managed patch was never applied
either: the build would have installed a terminal whose `strata install` and
`strata update` still try to write into /usr.
Add src_prepare() calling cmake_src_prepare(), and check the phases against a
real source tree instead of reading the ebuild: PATCHES applies with --fuzz=0,
eapply_user runs, the patched modules still parse, `strata update` refuses
with STRATA_PACKAGE_MANAGED=1, and the engine and the picture reader both
build from the patched tree with the paths src_install copies from.
commit 74684f53a47f4bddfe2b55a8b30743db2c8066f0
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:14:22 2026 +0800
sci-ml/strata: make the CUDA default portable and derive the platform
The build was tuned to one machine in two places.
With no architecture configured the ebuild fell back to 120-real, which is
this box's RTX 5080. On any other card that produces an engine that cannot
start, and the fallback was silently chosen. Read the cards from the driver
instead (nvidia-smi compute_cap, giving 86-real, 120-real, ...) so an
unconfigured build is right on whatever host it runs on, and refuse with a
usable message when the driver cannot be asked. The value exposed is always
concrete, never CMake's "native", because BUILD.json has to name the archs.
Naming CMAKE_CUDA_ARCHITECTURES or CUDAARCHS still wins, for building for
another machine or one engine for several cards.
The component's platform key was hardcoded to linux-x86_64 while upstream also
ships linux-aarch64 (the DGX Spark), so derive it from tc-arch and add ~arm64
to KEYWORDS.
README.gentoo described the old behaviour, including an architecture default
the build never had; describe the driver probe instead. Also state the HIP
architecture list accurately: upstream validates gfx1100 and gfx1201, has
community reports for gfx1101 and gfx1200, and refuses anything off its list.
Checked on this host: no configuration yields 120-real and a cubin that runs;
a stubbed nvidia-smi reporting 8.6 yields 86-real; one that fails aborts with
the message; explicit lists pass through unchanged; and 86-real, 120-real,
89-real;120-real and native all resolve consistently into BUILD.json archs.
commit e2e7d6e810328cb72447200f6266465b33b25f96
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:11:26 2026 +0800
sci-ml/strata: describe the architectures that were built, not a guess
src_install computed the arch list in BUILD.json from
$, a hardcoded fallback that ignored
CUDAARCHS. A host with CUDAARCHS=120 and no CMAKE_CUDA_ARCHITECTURES compiles
the engine and the reader for sm_120 alone, while the manifest claimed sm_86,
sm_89 and sm_120 - a superset, so package_engine.py's check passed and nothing
failed until a reader met a card it never had code for.
Take the list from cuda_architectures(), the same call src_configure uses, and
split on ';' as well as space so a multi-card list stays a list. BUILD.json's
archs and vision_archs now state what was compiled.
Checked for 120, 120-real, 89-real;120-real, 86;120 and
sm_86;compute_89;120-virtual: the value handed to CMake, the archs written to
BUILD.json and vision_archs agree in every case, and CUDAARCHS=120 alone
yields [sm_120].
commit f9614d13e642519bf33da9fe2d933348d7a8dc78
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:09:14 2026 +0800
sci-ml/strata: settle the CUDA architecture list in one place
Two problems, both in how the architecture list was chosen.
The ebuild read only CMAKE_CUDA_ARCHITECTURES, so a CUDAARCHS set in make.conf
- what CUDA's own tooling reads, and what this host has - reached CMake by its
own environment lookup rather than through the build. That worked by accident.
Both are read now, CMAKE_CUDA_ARCHITECTURES first as the more specific one, and
one cuda_architectures() supplies the engine and the picture reader alike so a
reader can never be built for a different card than the engine (#3669).
The default was also wrong in effect. Upstream defaults to a bare 120, which
CMake compiles to sm_120 SASS plus compute_120 PTX. PTX is only ever
JIT-compiled for a card newer than the one it was made for, so on the card the
build targets it is dead weight: an extra code path through the driver on first
launch, plus binary size. Builds with nothing configured now ask for
120-real, which is the SASS alone; users who want one binary for several cards
still list them, and a trailing bare entry keeps the PTX for unreleased cards.
Checked on an RTX 5080 (compute 12.0): 120 yields t.1.sm_120.cubin and
t.1.sm_120.ptx, 120-real yields the cubin only, both run, and the two hosts'
settings resolve as intended (CUDAARCHS=120 -> 120, nothing set -> 120-real,
CMAKE_CUDA_ARCHITECTURES wins over CUDAARCHS).
commit 86c561c0de8376bd4181c5ffd0923a9921f9c439
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:05:26 2026 +0800
sci-ml/strata: honour CUDAARCHS and correct the default arch note
The ebuild read only CMAKE_CUDA_ARCHITECTURES, so a CUDAARCHS set in
make.conf reached CMake by its own environment lookup rather than through the
build. Read both, and use the same list for the engine and the picture reader
so a reader can never be built for a different card than the engine.
The note also claimed upstream defaults to sm_86, sm_89 and the sm_120 PTX.
It defaults to a bare 120, which CMake compiles to sm_120 SASS plus
compute_120 PTX: that runs on an RTX 50 card and on nothing older. Say that
instead of naming architectures the build does not produce.
Checked on an RTX 5080 (compute 12.0): -DCMAKE_CUDA_ARCHITECTURES=120 emits
t.1.sm_120.cubin and t.1.sm_120.ptx, and the resulting binary runs.
commit 365b2d2e9f052c7704f7850242a75a5f3063a81c
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 20:37:37 2026 +0800
Bump packages reported by the update checker (#28)
Release bumps for the packages issue #28 reported:
app-text/quarto-bin 1.10.19 -> 1.10.20
sci-ml/strata 0.2.7.1 -> 0.2.8
sci-ml/torchaudio 2.11.0-r2 -> 2.11.1
sci-misc/prism-llama-cpp 10826 -> 10831
app-misc/chatgpt-desktop 26.1023.51421 -> 26.1027.30824
app-misc/claude-desktop 4.29.0 -> 4.4212.0
torchaudio 2.11.1 is a packaging and bug-fix release that still supports
PyTorch 2.11 and later, so the existing dependency and PYTHON_COMPAT stand.
Prism moves to prism-b10831-9d9efea. Strata keeps its pinned llama.cpp tree
(third_party/ggml/VERSION.txt is unchanged), and its terminal file set,
package_engine.py/stage_tuning.py interfaces and CMake options are unchanged,
so only the version moves; the Portage redaction patch still applies at fuzz=0.
The remaining two entries needed no ebuild change, but the checker was wrong:
* app-misc/bcompare already tracks 5.2.7.32913. The scooter lookup failed
because the site answers HTTP 500 to curl's default User-Agent; request the
page as a browser does.
* sci-misc/kvmem-llama-cpp stays at 0.18.0. kvmem publishes v0.19.0-rc1 with
the prerelease box unchecked, so get_latest_stable() returned a release
candidate. Filter the version string the way get_latest_tag() already does.
Also correct the launcher comment naming sci-misc/strata instead of sci-ml/strata.
commit d381011ce7e312bc4d83a61d0fb3616dc045f114
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:12:20 2026 +0800
sci-ml/strata: bump to 0.2.7.1
0.2.7.1 is a tag, not a GitHub release: upstream folded its notes into the
v0.2.7 release, and the tag is 127 commits ahead of the v0.2.7 tag. It is the
newest code, so the package follows the tag.
What the bump changes for the package: nothing structural. The llama.cpp pin
in third_party/ggml/VERSION.txt, the engine version (0.1.53), requirements.txt,
the terminal PAYLOAD/TOOLS_SUBSET, the HIP architecture lists and the
hipBLASLt header lookup are all unchanged; package_engine.py and
stage_tuning.py are byte-identical, and package_terminal.py only gains
linux-aarch64 support. files/redact-portage-managed.patch still applies with
--fuzz=0 (two hunks, offset 9).
Verified on amd64 with CUDA 13.4 and an RTX 5080: engine and reader compile,
package_engine.py accepts the real binaries, the component passes the
terminal's full SHA-256 verification, the packaged engine self-reports 0.1.53
and runs, `strata update` still refuses under Portage, and the vision reader
resolves at its installed path.
AGENTS: correct the version note. The update checker reads releases, so it
goes quiet once a release has a newer tag-only successor. Bumps follow tags.
commit 889a103311c4058519b65093bb133b6934b40089
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:05:27 2026 +0800
AGENTS: record the Strata packaging constraints
Collect what governs a Strata bump so it is not rediscovered each time: the
sci-ml placement, the pinned llama.cpp commit and why STRATA_GGML_DIR must be
set (an unset one makes CMake fetch over the network, which the build sandbox
refuses), why no system-ggml flag belongs here (the engine links ggml
statically and installs no headers, so it already coexists with sci-ml/ggml
and sci-misc/llama-cpp), the component manifest and RESTRICT=strip, the vision
reader's architecture rule and its two HIP configure runs, the flat
hipBLASLt version header, the architecture lists that must not be probed from
the build host, the terminal's staged file set, and the launcher's component
and data folders.
Also install docs/TERMINAL_LAUNCH.md: upstream's terminal payload carries it
beside MANAGEMENT_PROTOCOL.md, and the package installed only the latter.
commit 48838b2dd76227b2696be59b49d3b0acc66169a5
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:03:15 2026 +0800
sci-ml/strata: build the picture reader, and fix the install layout
Pictures
USE="vision" (on by default) builds tools/vision into strata-vision, so a
prompt can carry images. It is a second CMake project because it links
llama.cpp's mtmd beside the engine's ggml.
Both are built from the same architecture list, and the ebuild records that
list as vision_archs in BUILD.json: package_engine.py refuses a CUDA package
whose reader does not cover the engine's architectures, and a reader carried
over from another build dies with "device kernel image is invalid" on a card
the engine itself handles (#3669).
On HIP the project renames its single target to strata-vision-hip, so the
two readers upstream ships come from two configure runs of the project, the
way packaging/build_hip_linux.sh does it: the CPU reader strata-vision, which
starts even where the HIP runtime does not load, and strata-vision-hip for
the card.
Install layout
The component folder was created under /usr/$(get_libdir)/strata while a
dosym turned the same path into a link to /usr/share/strata, which cannot
both exist. The engine now lives in a real /usr/lib/strata/components and the
terminal's code stays in /usr/share/strata, reached through
/usr/lib/strata/terminal.
Nothing is installed under /usr/include or /usr/$(get_libdir) beside ggml or
llama, and the engine links both statically: a build never calls
cmake --install on the pinned llama.cpp tree, finds no system ggml or llama,
and the packaged binaries need no libggml or libllama. sci-ml/ggml (shared,
include/ggml*.h) and sci-misc/llama-cpp (include/llama.cpp) can be installed
beside this package.
Verified on amd64 with CUDA 13.4 and an RTX 5080: the engine and the reader
compile, the packaging helper accepts the real binaries and refuses a
deliberately mismatched vision_archs, the component passes the terminal's full
SHA-256 verification, and strata_control.vision_reader resolves the reader at
its installed path. The HIP branch is not built here (no ROCm on this host).
commit 0cab938e0f4a884431823a4eb0f517348ff62f5e
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 00:03:25 2026 +0800
sci-ml/strata: add 0.2.7
Strata is an inference engine for Qwen3.8-Flash-Next and other GGUF models
on consumer NVIDIA and AMD hardware, with an OpenAI/Anthropic-compatible API
and a terminal that picks, verifies and starts the engine.
The package follows the sci-ml category rather than the sci-misc the two
public ebuilds in the wild use: sci-ml is defined for machine learning and
natural language processing, sci-misc is the fallback for packages that fit
no other sci-* category, and the sibling inference engines (sci-ml/ollama,
sci-ml/ggml) already live there.
Engine
Built for exactly one backend, cuda or hip. ggml comes from the llama.cpp
commit pinned in third_party/ggml/VERSION.txt as a distfile, so a build
never downloads; upstream's own ladder of CMake options is used instead of
a patch. The binary is installed as the component package the terminal
resolves and verifies (engine-component.json, SHA-256 per file), which is
why RESTRICT=strip. For HIP the ebuild reads the hipBLASLt version from the
Gentoo-installed header -- hipBLASLt installs it flat as
include/hipblaslt-version.h, while upstream's stage_tuning.py looks under
include/hipblaslt/ -- and passes the number with --version.
Terminal
Installed under /usr/share/strata with the module subset upstream's
package_terminal.py stages, so `strata` and the API server import tools.*
from the install root. A launcher in /usr/bin points the component folder
at /usr/lib/strata/components and the model and run-config folders at the
caller's XDG data directory, which Strata would otherwise put in a
root-owned sibling.
files/redact-portage-managed.patch makes `strata install` and `strata
update` refuse under Portage and name emerge instead, while rollback and
install-file keep working for a user-owned --components-dir.
The picture reader (strata-vision) is not built, so Strata answers text only;
pkg_postinst says so.
Verified on amd64 with CUDA 13.4 and an RTX 5080: the engine compiles, the
packaging helper accepts the real binary, the component passes the terminal's
full SHA-256 verification, the packaged engine reports its version, and
strata/strata update/strata components behave as above from the installed
tree. Patch verified with patch --dry-run --fuzz=0 against the 0.2.7 tarball.
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Mon Nov 16 23:20:51 2026 +0800
sci-ml/strata: bump to 0.2.15
The new release builds the engine without llama.cpp's MMQ prompt kernels
(STRATA_MMQ defaults to OFF), so pass that value explicitly as upstream's
release packaging does. The picture reader and the engine still build from
the pinned llama.cpp tree.
The portage-managed guard needed its update_cmd hunks redrawn: run() gained a
nested fail() helper and the component-dir logic moved. Re-derive both hunks
against the 0.2.15 tarball and add the new draft_vocab_latin.bin data file.
commit 53277b67e88e9adeac543025f357c4f10903fcf5
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Wed Nov 11 23:19:46 2026 +0800
sci-ml/strata: drop the kvmem-llama-cpp references in -r1
sci-misc/kvmem-llama-cpp has been removed from the overlay, so the elog
message and files/README.gentoo no longer name it. The published revision
moves into -r1 with those text changes. The build and the install layout
are unchanged, and the redaction patch still applies in prepare.
The CI install skip list gets an entry for strata, since its install needs
cmake, ninja and, with the default +cuda USE, the CUDA toolkit.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit dfcc93423102b994f8e3061ec4476d81e456661e
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Tue Nov 10 20:16:24 2026 +0800
sci-ml/strata: build the CUDA engine with the CUDA host GCC
With a CUDA build, nvcc's host compiler (CUDAHOSTCXX) is the newest GCC
that CUDA supports, which can be older than the default GCC. CMake adds
that GCC's library directory to the link line, but the C and C++ objects
were compiled with the default GCC's headers, so the engine did not link:
std::__detail::__notify_impl and __wait_impl come from the newer
libstdc++ only.
Export CC and CXX from the same GCC directory as CUDAHOSTCXX, both for
the engine's configure and for the picture reader's configure in
src_compile, so every object and the link use one GCC. HIP builds are
unchanged.
Verified on the build host (CUDA 12.9.86, default GCC 16, GTX 1080 Ti):
the engine and the CUDA picture reader compile and link, and
package_engine.py writes the component folder with its manifest. The
architecture was set to sm_75 only in a test copy of the ebuild, because
the host GPU is Pascal and upstream refuses sm_61. The packaged binaries
run `strata --version` on this host. Inference on a GPU was not run.
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit fb492ed964abef4d4d24f25836b27aac0c9ef6f6
Author: Ruizhi Zhong <ruizhi.zhong88@gmail.com>
Date: Tue Nov 10 20:01:55 2026 +0800
overlay: bump packages from issue #36
Bump the packages reported in issue #36 to their current upstream
versions:
- media-sound/sidra-bin 1.2.1 -> 1.3.0
- app-misc/chatgpt-desktop 26.1103.32239 -> 26.1103.60933
- app-misc/claude-desktop 4.23118.0 -> 4.23118.1
- sci-ml/strata 0.2.12 -> 0.2.13.1
The sidra and desktop app distfiles match the SHA256 digests published
for them (the Debian package indexes and the GitHub asset digests), and
their install phases were checked on amd64. The RDEPEND lists already
cover the dependencies of the new packages.
Strata follows the newest upstream tag, 0.2.13.1, rather than the 0.2.13
release, because the release list does not show the patch tag; Strata
versions follow the tags (see AGENTS.md). Its ggml pin, 3cf03257, is
unchanged. The redaction patch was re-derived against the new tree: the
same policy text, its hunks at their new line positions, and it applies
with --fuzz=0 at zero offset.
Not verified: the Strata engine does not link on this build host. It was
configured with a test-only sm_75 architecture, because the host GPU is
Pascal and upstream refuses sm_61. The link pulls GCC 14's libstdc++
(CUDA 12.9's host compiler reaches the link through CMake's CUDA implicit
link directories) while the objects are compiled with GCC 16 headers, so
std::__detail::__notify_impl and __wait_impl are undefined. The ebuild's
CUDA compiler setup is unchanged from 0.2.12, and 0.2.12 uses the same
atomic wait/notify code, so this predates this bump.
Closes #36
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
commit 53c42444a316899b153e86b6c68c9e6247e2fc54
Author: user.email <ruizhi.zhong@outlook.com>
Date: Sat Nov 7 17:45:07 2026 +0800
overlay: bump Strata and ChatGPT Desktop from issue #34
sci-ml/strata 0.2.11 -> 0.2.12
app-misc/chatgpt-desktop 26.1103.31037 -> 26.1103.32239
Strata 0.2.12 is the newest tag as well as the newest release, so no tag-only
successor was missed. third_party/ggml/VERSION.txt still pins 3cf0325, so
LLAMA_COMMIT and the llama.cpp distfile are unchanged, and every interface the
ebuild drives (package_engine.py, stage_tuning.py, the CMake options for the
engine and tools/vision, tools/hip) is still present. The redaction patch
applies at fuzz=0 and still guards every writer of the component folder: the
0.2.12 cli only exposes read-only components subcommands, and the writing ones
go through update_cmd.
One file-set change needed a fix. 0.2.12 adds data/dflash2-tuned/ (the
Apache-2.0 text and the "modified by Strata" notice for the tuned drafter),
and strata_control/drafting_fetch.provide_licence() resolves it from the
install root to copy beside the drafter it fetches. The ebuild installs data/
file by file, so the directory was left out and the call would have silently
returned False, dropping the attribution required when redistributing that
Apache-2.0 drafter. It is installed now; upstream's own package_terminal.py
PAYLOAD omits it as well, so this is deliberately wider than their archive.
The new tools/force_check.py is a developer tool no installed module imports,
so TOOLS_SUBSET stays as it is.
ChatGPT Desktop only changes its version: the new deb still ships both Qt
shims, codex-launcher, the apparmor profile naming /usr/lib/chatgpt/ChatGPT,
and Exec=chatgpt %U, which the existing sed still rewrites.
commit 2fd413bdc5768f584aef9424b32fb54f54b54898
Author: user.email <ruizhi.zhong@outlook.com>
Date: Thu Nov 5 21:44:07 2026 +0800
sci-ml/strata: bump to 0.2.11
Engine 0.1.57, and the llama.cpp pin is unchanged
(third_party/ggml/VERSION.txt still 3cf03257), so LLAMA_COMMIT and the ggml
distfile stay as they are.
The portage-managed patch needed re-deriving. Upstream moved run() in
strata_cli/update_cmd.py and added source_store()/default_components_dir()
above it, so both hunks there no longer applied. Regenerated against 0.2.11:
the guard now sits after the fail() closure and reads the component folder
through default_components_dir(), which is what run() itself uses, and
`update --check` still passes through while `update`, `components
update|rollback|install-file|use|remove` refuse. The cli.py hunk only shifted.
The model catalogue grew from 17 to 23 selections (Q8_0 in six shards, the
Gyro TQ quants, more UD sizes); every one is still Qwen3.8-Flash-Next, so the
description and the note in README.gentoo stand.
Checked on amd64 with CUDA 13.4 and an RTX 5080: the patch applies with
--fuzz=0, the patched modules parse, `strata install` and `strata update`
refuse while `update --check` and `components engines` are allowed and
`components use --components-dir <user dir>` reaches its own logic, the engine
and the picture reader compile, package_engine.py accepts the real binaries,
and the component passes the terminal's full SHA-256 verification and runs.
commit ad2921156d1e2649dd9de4c927e44a2e30628913
Author: Ruizhi Zhong <ruizhi.zhong@outlook.com>
Date: Tue Nov 3 21:31:29 2026 +0800
sci-ml/strata: let update --check through the Portage guard
The 0.2.10 re-derivation of redact-portage-managed.patch put the update refusal
in front of everything, including `strata update --check` and `strata
components update --check`, which only read the signed feed and write nothing.
The patch header already said read-only operations were left alone, so the
guard and the documentation disagreed.
`--check` is now exempt like `components engines`; the writes in the component
folder stay refused (`update`, `rollback`, `install-file`, `use`, `remove`)
unless --components-dir names a folder the calling user can write. Checked by
calling run() with STRATA_PACKAGE_MANAGED=1: both update forms exit 5, both
--check forms exit 0 and print the feed, and the component-folder writers
refuse unless given an owned directory.
AGENTS.md now records why the patch is kept (the component folder belongs to
Portage, upstream's STRATA_NO_ACQUIRE does not cover the writers), that it is
gated by a launcher-set variable rather than being a sandbox, and that every
future command writing there must be added to it.
commit 2362000c96a1b8f7ad451e7aadb809c4e8171102
Author: Ruizhi Zhong <ruizhi.zhong@outlook.com>
Date: Tue Nov 3 21:13:21 2026 +0800
overlay: bump quarto-bin, chatgpt-desktop and strata
app-text/quarto-bin 1.11.9: straight version bump of the prebuilt tarball.
The layout is unchanged: bin/quarto and share/man/quarto-man.man are still
where the ebuild installs them.
app-misc/chatgpt-desktop 26.1027.51404: straight version bump. The deb still
carries usr/lib/chatgpt with the Qt5/Qt6 shims, codex-launcher and
usr/bin/chatgpt, the AppArmor profile still names /usr/lib/chatgpt/ChatGPT,
and the desktop entry still starts with Exec=chatgpt, so src_prepare's edits
and the install paths need no change. The prebuilds loop still keeps the
linux-x64 node modules and drops every other platform, including the musl
ones the next find deletes.
sci-ml/strata 0.2.10:
- third_party/ggml/VERSION.txt still pins llama.cpp 3cf03257, so LLAMA_COMMIT
stays put; the regenerated Manifest confirms that distfile is unchanged.
- redact-portage-managed.patch is re-derived against the new tarball: 0.2.10
rewrote the end of _print_check in strata_cli/update_cmd.py, which is where
the helper is inserted. `patch --dry-run --fuzz=0` applies cleanly and the
patched modules compile.
- 0.2.10 adds `strata components use|remove`, which switch the active engine
inside the component folder and delete engine packages from it. The patch
treats them like rollback/install-file: refused under
STRATA_PACKAGE_MANAGED=1 unless --components-dir names a folder the calling
user can write, because the system component folder belongs to Portage.
The read-only `engines` listing and `--check` are unaffected.
- requirements.txt, requirements-monitor.txt, the packaging helpers
(package_engine.py, package_terminal.py, stage_tuning.py) and the terminal
file set package_terminal.py stages are unchanged, as are the data files the
ebuild installs, its tools subset and the docs it names. Every CMake switch
and target the ebuild uses still exists; the engine itself is now 0.1.56,
which the ebuild reads from CMakeLists.txt.
Manifests regenerated with pkgdev manifest. pkgcheck reports no new findings;
xmllint and git diff --check are clean. No build was run locally.
commit ff40a925b41d0d8c868a227c9b81d0fec73160b6
Author: user.email <ruizhi.zhong@outlook.com>
Date: Sat Oct 31 20:56:48 2026 +0800
overlay: bump strata, datagrip, prism-llama-cpp and kvmem-llama-cpp
sci-ml/strata 0.2.9: the ggml pin in third_party/ggml/VERSION.txt is
unchanged, so LLAMA_COMMIT stays put. redact-portage-managed.patch is
re-derived against the new tarball; it applied with a nine-line offset and
now applies cleanly. requirements.txt and requirements-monitor.txt did not
change, and the terminal file set package_terminal.py stages is the same.
dev-util/datagrip 2026.2.6.1: straight version bump, every path the ebuild
sets permissions on still exists in the new tarball.
sci-misc/prism-llama-cpp 12128: upstream tag prism-b12128-d7b2896.
sci-misc/kvmem-llama-cpp 0.19.1:
- Upstream moved to kvmem/kvmem, so HOMEPAGE, the SRC_URI, metadata.xml
and the update checker follow.
- backends/versions.json now pins llama.cpp 620d222 and replaces the single
kvmem patch with an ordered five-patch series (kvmem, pq2, ptq1,
ptq1-performance, checkpoint-policy). scripts/prepare-backends.py needs the
submodule's git metadata to verify the series, which a tarball build does
not have, so the pinned tree is unpacked and the same series, in the same
order, is applied with eapply. All five apply with fuzz=0.
- KVMEM_ENABLE_NVME is gone: the NVMe store was retired upstream in favour
of the shared RAM/SSD tiers. Its build switch and its test target are
dropped; host_kv_storage_test and tiered_kv_storage_test cover the tiers
that replaced it.
- New backport: monitor_hardware.hpp includes <nvml.h>, whose one-argument
DEPRECATED(ver) clobbers llama.cpp's two-argument DEPRECATED(func, hint)
and breaks mtmd.h in llama-kvmem-server.cpp. The patch restores the
llama.cpp spelling after NVML is done with the name.
- README.gentoo: the old example passed --kvmem-gen-reserve 16384, which is
now documented as reusable generation workspace rather than an output cap.
It shows --kvmem-host-mib and the opt-in SSD pair instead.
Validated locally on amd64 with CUDA 13.4 and sm_120a: kvmem 0.19.1 builds
llama-kvmem-cli and llama-kvmem-server and its six model-free tests pass;
strata 0.2.9 builds its engine; prism 12128 configures. Manifest regenerated
with pkgdev manifest; pkgcheck reports no new findings.
commit 0f9b1b8cef4122d833faa92d604c370d8d617fe8
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 22:05:32 2026 +0800
sci-ml/strata: move the published packaging fixes into -r1
0.2.8 has been pushed, so the eleven commits since then are changes to a
published ebuild: they must not ride on the same PV. Create -r1 instead.
The revision matters more here than the wording suggests. With FEATURES
holding buildpkg and getbinpkg, a PV that was already built is served from
$PKGDIR/sci-ml/strata/strata-0.2.8-1.gpkg.tar, and Portage reported the
package as a binary reinstall - the corrected src_prepare, the CUDA
architecture handling and the model note would never have been built or shown.
Under -r1 the same emerge now reports an upgrade and compiles the corrected
ebuild.
PV is unchanged, so SRC_URI, the tarball name, S and the pinned llama.cpp tree
are untouched; only PF, and with it the component's build id, gains the
revision. The Manifest needs no change.
commit 35c1c14ff6fa2da6fb88fd7e0cc8b549628777f5
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 22:04:36 2026 +0800
sci-ml/strata: say which model the engine actually runs
The description claimed "Qwen3.8-Flash-Next and other GGUF models", and the
long description added DeepSeek-V4.1-Flash. Neither is true of this package.
The catalogue in data/gguf_fingerprints.json holds 17 selections under the
qwen, swift, coder, unsloth and gyro keys and all of them are
Qwen3.8-Flash-Next (qwen35moe) in GSQ-RCO, Q2_0 or UD quantizations, split
into two shards with exact upstream names; `strata models import` refuses
anything else, including other Qwen3.8 sizes. DeepSeek needs
STRATA_BUILD_DSV41=ON, which the ebuild does not build.
Verified against an installed terminal with a local GGUF library: a
Qwen3.8-27B GSQ-RCO IQ3_S file (arch qwen35) is refused with "this model type
is not supported by the Strata engine (add it as a Custom model API with
llama-server instead)", and so is the MTP drafter beside it, while the
catalogue's own entries are all Flash-Next.
Correct DESCRIPTION and metadata.xml, and point users at
sci-misc/kvmem-llama-cpp, sci-misc/prism-llama-cpp or guru's
sci-misc/llama-cpp for anything else in README.gentoo and pkg_postinst. Record
the constraint in AGENTS.md so a later bump does not reintroduce the claim.
commit d2382c0150bc798b2d00412679c25340a5354217
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:55:50 2026 +0800
sci-ml/strata: fix a comment naming an array as a function
commit 8ca83a04275a6f9bdebcd5b35f4b90e7d50245e1
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:55:43 2026 +0800
sci-ml/strata: bring the ebuild in line with Gentoo practice
A standards pass over the ebuild, rather than another build fix.
Declare dev-python/rich. The terminal imports rich in twenty places and it
arrived only as textual's dependency; an ebuild must name what its code
imports, or a textual change silently breaks it.
Compute the packaging architecture list in src_configure instead of
src_install. src_install ran nvidia-smi to work out what to write into
BUILD.json, which is a build decision made in the wrong phase and repeated a
third time. It now reuses the list that was already derived, and the fallback
notice fires once instead of per call.
Drop ~arm64 from KEYWORDS. The engine, the reader and the platform key all
support linux-aarch64, but nothing has merged on that hardware here, and a
keyword claims testing that has not happened.
Checked that the four architecture settings still resolve to one list shared by
CMake, BUILD.json's archs and vision_archs, that the fallback warning is
emitted once and goes to stderr, and that emergence still resolves.
commit 688d2bde3447fd8058469ad50fc676d45798d2af
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:53:27 2026 +0800
sci-ml/strata: keep the launcher's shell shebang
python_fix_shebang was run on /usr/bin/strata, which is a POSIX shell script
that finds an interpreter and execs strata.py. The helper only rewrites Python
shebangs and dies on anything else when a file is named explicitly, so the
install phase aborted after the engine had been packaged.
Drop the call; the shell shebang is what the launcher needs, and the modules
under /usr/share/strata are still byte-compiled by python_optimize. Checked
that the installed launcher keeps #!/bin/sh, stays executable, and reports the
engine build it was merged with.
commit 7e10a34ceee74bd558643e1c67e1b7ce0b33a147
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:43:32 2026 +0800
sci-ml/strata: allow the CUDA runtime soname when packaging
The allow-list derived from each binary's DT_NEEDED named the AMD sonames and
libcublas/libcublasLt, but not libcudart, so package_engine.py refused every
CUDA package: strata and strata-vision both need libcudart.so.13. My earlier
checks passed only because they spelled --allow-lib by hand, which is exactly
what the ebuild must not depend on.
Match the rest of the CUDA runtime sonames too (cufft, curand, cusparse,
cusolver, nvrtc, nvJitLink) so a future backend does not reintroduce this.
These stay system libraries: nvidia-cuda-toolkit, already in RDEPEND, puts
them on LDPATH through /etc/env.d.
Checked by deriving the arguments with the ebuild's own loop from the real
binaries and packaging with no hand-written --allow-lib - it succeeds, and the
package's strata and strata-vision both run from the unpacked component
folder. Without the allow-list the packager still refuses, so the check has
not been widened into uselessness.
commit 9a81447c982c3093a1cdb2fc07cef6e526c14b43
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:35:08 2026 +0800
sci-ml/strata: stop the HIP copy from failing a CUDA install
package_engine copied strata-device with
use hip && cp "$/strata-device" "${stage}/" || die
which bash associates left to right, so the failure of `use hip` ran the die.
Every CUDA install aborted in src_install, after the whole engine had compiled.
Group the copy under an if, so `use hip` being false skips the copy instead of
being read as its failure.
Only that line had the shape: the other three `use X && command` lines run a
function that dies on its own, so nothing was hidden.
Checked by reproducing the expression with USE lacking hip (the old form calls
die, the new one returns 0), and by running package_engine.py over a stage
built from this release with a CUDA engine and reader: it packages and the
HIP branch is skipped.
commit dbda4f6a96821433c22f46e851d58488997afe25
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:26:43 2026 +0800
sci-ml/strata: add the missing src_prepare
The ebuild declared PATCHES but never defined src_prepare(), so Portage
aborted the prepare phase with "eapply_user (or default) must be called in
src_prepare()". cmake_src_prepare() is what applies PATCHES and then calls
eapply_user, and without it the portage-managed patch was never applied
either: the build would have installed a terminal whose `strata install` and
`strata update` still try to write into /usr.
Add src_prepare() calling cmake_src_prepare(), and check the phases against a
real source tree instead of reading the ebuild: PATCHES applies with --fuzz=0,
eapply_user runs, the patched modules still parse, `strata update` refuses
with STRATA_PACKAGE_MANAGED=1, and the engine and the picture reader both
build from the patched tree with the paths src_install copies from.
commit 74684f53a47f4bddfe2b55a8b30743db2c8066f0
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:14:22 2026 +0800
sci-ml/strata: make the CUDA default portable and derive the platform
The build was tuned to one machine in two places.
With no architecture configured the ebuild fell back to 120-real, which is
this box's RTX 5080. On any other card that produces an engine that cannot
start, and the fallback was silently chosen. Read the cards from the driver
instead (nvidia-smi compute_cap, giving 86-real, 120-real, ...) so an
unconfigured build is right on whatever host it runs on, and refuse with a
usable message when the driver cannot be asked. The value exposed is always
concrete, never CMake's "native", because BUILD.json has to name the archs.
Naming CMAKE_CUDA_ARCHITECTURES or CUDAARCHS still wins, for building for
another machine or one engine for several cards.
The component's platform key was hardcoded to linux-x86_64 while upstream also
ships linux-aarch64 (the DGX Spark), so derive it from tc-arch and add ~arm64
to KEYWORDS.
README.gentoo described the old behaviour, including an architecture default
the build never had; describe the driver probe instead. Also state the HIP
architecture list accurately: upstream validates gfx1100 and gfx1201, has
community reports for gfx1101 and gfx1200, and refuses anything off its list.
Checked on this host: no configuration yields 120-real and a cubin that runs;
a stubbed nvidia-smi reporting 8.6 yields 86-real; one that fails aborts with
the message; explicit lists pass through unchanged; and 86-real, 120-real,
89-real;120-real and native all resolve consistently into BUILD.json archs.
commit e2e7d6e810328cb72447200f6266465b33b25f96
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:11:26 2026 +0800
sci-ml/strata: describe the architectures that were built, not a guess
src_install computed the arch list in BUILD.json from
$, a hardcoded fallback that ignored
CUDAARCHS. A host with CUDAARCHS=120 and no CMAKE_CUDA_ARCHITECTURES compiles
the engine and the reader for sm_120 alone, while the manifest claimed sm_86,
sm_89 and sm_120 - a superset, so package_engine.py's check passed and nothing
failed until a reader met a card it never had code for.
Take the list from cuda_architectures(), the same call src_configure uses, and
split on ';' as well as space so a multi-card list stays a list. BUILD.json's
archs and vision_archs now state what was compiled.
Checked for 120, 120-real, 89-real;120-real, 86;120 and
sm_86;compute_89;120-virtual: the value handed to CMake, the archs written to
BUILD.json and vision_archs agree in every case, and CUDAARCHS=120 alone
yields [sm_120].
commit f9614d13e642519bf33da9fe2d933348d7a8dc78
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:09:14 2026 +0800
sci-ml/strata: settle the CUDA architecture list in one place
Two problems, both in how the architecture list was chosen.
The ebuild read only CMAKE_CUDA_ARCHITECTURES, so a CUDAARCHS set in make.conf
- what CUDA's own tooling reads, and what this host has - reached CMake by its
own environment lookup rather than through the build. That worked by accident.
Both are read now, CMAKE_CUDA_ARCHITECTURES first as the more specific one, and
one cuda_architectures() supplies the engine and the picture reader alike so a
reader can never be built for a different card than the engine (#3669).
The default was also wrong in effect. Upstream defaults to a bare 120, which
CMake compiles to sm_120 SASS plus compute_120 PTX. PTX is only ever
JIT-compiled for a card newer than the one it was made for, so on the card the
build targets it is dead weight: an extra code path through the driver on first
launch, plus binary size. Builds with nothing configured now ask for
120-real, which is the SASS alone; users who want one binary for several cards
still list them, and a trailing bare entry keeps the PTX for unreleased cards.
Checked on an RTX 5080 (compute 12.0): 120 yields t.1.sm_120.cubin and
t.1.sm_120.ptx, 120-real yields the cubin only, both run, and the two hosts'
settings resolve as intended (CUDAARCHS=120 -> 120, nothing set -> 120-real,
CMAKE_CUDA_ARCHITECTURES wins over CUDAARCHS).
commit 86c561c0de8376bd4181c5ffd0923a9921f9c439
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 21:05:26 2026 +0800
sci-ml/strata: honour CUDAARCHS and correct the default arch note
The ebuild read only CMAKE_CUDA_ARCHITECTURES, so a CUDAARCHS set in
make.conf reached CMake by its own environment lookup rather than through the
build. Read both, and use the same list for the engine and the picture reader
so a reader can never be built for a different card than the engine.
The note also claimed upstream defaults to sm_86, sm_89 and the sm_120 PTX.
It defaults to a bare 120, which CMake compiles to sm_120 SASS plus
compute_120 PTX: that runs on an RTX 50 card and on nothing older. Say that
instead of naming architectures the build does not produce.
Checked on an RTX 5080 (compute 12.0): -DCMAKE_CUDA_ARCHITECTURES=120 emits
t.1.sm_120.cubin and t.1.sm_120.ptx, and the resulting binary runs.
commit 365b2d2e9f052c7704f7850242a75a5f3063a81c
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 20:37:37 2026 +0800
Bump packages reported by the update checker (#28)
Release bumps for the packages issue #28 reported:
app-text/quarto-bin 1.10.19 -> 1.10.20
sci-ml/strata 0.2.7.1 -> 0.2.8
sci-ml/torchaudio 2.11.0-r2 -> 2.11.1
sci-misc/prism-llama-cpp 10826 -> 10831
app-misc/chatgpt-desktop 26.1023.51421 -> 26.1027.30824
app-misc/claude-desktop 4.29.0 -> 4.4212.0
torchaudio 2.11.1 is a packaging and bug-fix release that still supports
PyTorch 2.11 and later, so the existing dependency and PYTHON_COMPAT stand.
Prism moves to prism-b10831-9d9efea. Strata keeps its pinned llama.cpp tree
(third_party/ggml/VERSION.txt is unchanged), and its terminal file set,
package_engine.py/stage_tuning.py interfaces and CMake options are unchanged,
so only the version moves; the Portage redaction patch still applies at fuzz=0.
The remaining two entries needed no ebuild change, but the checker was wrong:
* app-misc/bcompare already tracks 5.2.7.32913. The scooter lookup failed
because the site answers HTTP 500 to curl's default User-Agent; request the
page as a browser does.
* sci-misc/kvmem-llama-cpp stays at 0.18.0. kvmem publishes v0.19.0-rc1 with
the prerelease box unchecked, so get_latest_stable() returned a release
candidate. Filter the version string the way get_latest_tag() already does.
Also correct the launcher comment naming sci-misc/strata instead of sci-ml/strata.
commit d381011ce7e312bc4d83a61d0fb3616dc045f114
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:12:20 2026 +0800
sci-ml/strata: bump to 0.2.7.1
0.2.7.1 is a tag, not a GitHub release: upstream folded its notes into the
v0.2.7 release, and the tag is 127 commits ahead of the v0.2.7 tag. It is the
newest code, so the package follows the tag.
What the bump changes for the package: nothing structural. The llama.cpp pin
in third_party/ggml/VERSION.txt, the engine version (0.1.53), requirements.txt,
the terminal PAYLOAD/TOOLS_SUBSET, the HIP architecture lists and the
hipBLASLt header lookup are all unchanged; package_engine.py and
stage_tuning.py are byte-identical, and package_terminal.py only gains
linux-aarch64 support. files/redact-portage-managed.patch still applies with
--fuzz=0 (two hunks, offset 9).
Verified on amd64 with CUDA 13.4 and an RTX 5080: engine and reader compile,
package_engine.py accepts the real binaries, the component passes the
terminal's full SHA-256 verification, the packaged engine self-reports 0.1.53
and runs, `strata update` still refuses under Portage, and the vision reader
resolves at its installed path.
AGENTS: correct the version note. The update checker reads releases, so it
goes quiet once a release has a newer tag-only successor. Bumps follow tags.
commit 889a103311c4058519b65093bb133b6934b40089
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:05:27 2026 +0800
AGENTS: record the Strata packaging constraints
Collect what governs a Strata bump so it is not rediscovered each time: the
sci-ml placement, the pinned llama.cpp commit and why STRATA_GGML_DIR must be
set (an unset one makes CMake fetch over the network, which the build sandbox
refuses), why no system-ggml flag belongs here (the engine links ggml
statically and installs no headers, so it already coexists with sci-ml/ggml
and sci-misc/llama-cpp), the component manifest and RESTRICT=strip, the vision
reader's architecture rule and its two HIP configure runs, the flat
hipBLASLt version header, the architecture lists that must not be probed from
the build host, the terminal's staged file set, and the launcher's component
and data folders.
Also install docs/TERMINAL_LAUNCH.md: upstream's terminal payload carries it
beside MANAGEMENT_PROTOCOL.md, and the package installed only the latter.
commit 48838b2dd76227b2696be59b49d3b0acc66169a5
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 01:03:15 2026 +0800
sci-ml/strata: build the picture reader, and fix the install layout
Pictures
USE="vision" (on by default) builds tools/vision into strata-vision, so a
prompt can carry images. It is a second CMake project because it links
llama.cpp's mtmd beside the engine's ggml.
Both are built from the same architecture list, and the ebuild records that
list as vision_archs in BUILD.json: package_engine.py refuses a CUDA package
whose reader does not cover the engine's architectures, and a reader carried
over from another build dies with "device kernel image is invalid" on a card
the engine itself handles (#3669).
On HIP the project renames its single target to strata-vision-hip, so the
two readers upstream ships come from two configure runs of the project, the
way packaging/build_hip_linux.sh does it: the CPU reader strata-vision, which
starts even where the HIP runtime does not load, and strata-vision-hip for
the card.
Install layout
The component folder was created under /usr/$(get_libdir)/strata while a
dosym turned the same path into a link to /usr/share/strata, which cannot
both exist. The engine now lives in a real /usr/lib/strata/components and the
terminal's code stays in /usr/share/strata, reached through
/usr/lib/strata/terminal.
Nothing is installed under /usr/include or /usr/$(get_libdir) beside ggml or
llama, and the engine links both statically: a build never calls
cmake --install on the pinned llama.cpp tree, finds no system ggml or llama,
and the packaged binaries need no libggml or libllama. sci-ml/ggml (shared,
include/ggml*.h) and sci-misc/llama-cpp (include/llama.cpp) can be installed
beside this package.
Verified on amd64 with CUDA 13.4 and an RTX 5080: the engine and the reader
compile, the packaging helper accepts the real binaries and refuses a
deliberately mismatched vision_archs, the component passes the terminal's full
SHA-256 verification, and strata_control.vision_reader resolves the reader at
its installed path. The HIP branch is not built here (no ROCm on this host).
commit 0cab938e0f4a884431823a4eb0f517348ff62f5e
Author: user.email <ruizhi.zhong@outlook.com>
Date: Fri Oct 30 00:03:25 2026 +0800
sci-ml/strata: add 0.2.7
Strata is an inference engine for Qwen3.8-Flash-Next and other GGUF models
on consumer NVIDIA and AMD hardware, with an OpenAI/Anthropic-compatible API
and a terminal that picks, verifies and starts the engine.
The package follows the sci-ml category rather than the sci-misc the two
public ebuilds in the wild use: sci-ml is defined for machine learning and
natural language processing, sci-misc is the fallback for packages that fit
no other sci-* category, and the sibling inference engines (sci-ml/ollama,
sci-ml/ggml) already live there.
Engine
Built for exactly one backend, cuda or hip. ggml comes from the llama.cpp
commit pinned in third_party/ggml/VERSION.txt as a distfile, so a build
never downloads; upstream's own ladder of CMake options is used instead of
a patch. The binary is installed as the component package the terminal
resolves and verifies (engine-component.json, SHA-256 per file), which is
why RESTRICT=strip. For HIP the ebuild reads the hipBLASLt version from the
Gentoo-installed header -- hipBLASLt installs it flat as
include/hipblaslt-version.h, while upstream's stage_tuning.py looks under
include/hipblaslt/ -- and passes the number with --version.
Terminal
Installed under /usr/share/strata with the module subset upstream's
package_terminal.py stages, so `strata` and the API server import tools.*
from the install root. A launcher in /usr/bin points the component folder
at /usr/lib/strata/components and the model and run-config folders at the
caller's XDG data directory, which Strata would otherwise put in a
root-owned sibling.
files/redact-portage-managed.patch makes `strata install` and `strata
update` refuse under Portage and name emerge instead, while rollback and
install-file keep working for a user-owned --components-dir.
The picture reader (strata-vision) is not built, so Strata answers text only;
pkg_postinst says so.
Verified on amd64 with CUDA 13.4 and an RTX 5080: the engine compiles, the
packaging helper accepts the real binary, the component passes the terminal's
full SHA-256 verification, the packaged engine reports its version, and
strata/strata update/strata components behave as above from the installed
tree. Patch verified with patch --dry-run --fuzz=0 against the 0.2.7 tarball.

