gpo.zugaina.org

Search Portage & Overlays:

sci-libs/onnxruntime

Cross-platform, high performance ML inferencing and training accelerator

Screenshots

  • onnxruntime-1.32.0
    ~amd64 ~arm64
    cuda python test python_targets_python3_12 python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: Apache-2.0 BSD MIT   
    Overlay: stuff
  • onnxruntime-1.32.0
    ~amd64
    python test python_targets_python3_12 python_targets_python3_13 python_targets_python3_14 python_targets_python3_15

    View      Download      Browse     License: MIT   
    Overlay: bentoo
  • onnxruntime-1.31.1
    ~amd64 ~arm64
    cuda python test python_targets_python3_12 python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: Apache-2.0 BSD MIT   
    Overlay: stuff
  • onnxruntime-1.31.0
    ~amd64 ~arm64
    cuda python test python_targets_python3_12 python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: Apache-2.0 BSD MIT   
    Overlay: stuff
  • onnxruntime-1.29.0
    ~amd64
    python test python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: MIT   
    Overlay: guru
  • onnxruntime-1.29.0
    ~amd64
    python test python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: MIT   
    Overlay: local-ai
  • onnxruntime-1.28.0
    ~amd64
    python test python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: MIT   
    Overlay: local-ai
  • onnxruntime-1.28.0
    ~amd64
    python test python_targets_python3_13 python_targets_python3_14

    View      Download      Browse     License: MIT   
    Overlay: guru

ChangeLog

commit c226dfd8c98c9ae5bd9562ce722fac6e162a9702
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Nov 14 22:53:48 2026 +0100

sci-libs/onnxruntime: drop 1.30.0-r2

Superseded; no in-tree atom needs this version. Its three patches
(1.29.0-fix-cuda-test-linking, 1.29.0-onnx-1.23-float6 and
1.30.0-use-system-libraries) have no other user and go with it.

commit 27153ea5b056741d4f26e6ed3f63dbfbd9960759
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Nov 14 22:29:45 2026 +0100

sci-libs/onnxruntime: add 1.32.0

Upstream now pins the Eigen 5.0.1 release, which replaces the 3.4
snapshot, and ships the float6 element-type support the ONNX 1.23
patch added, so that patch is dropped. Prestaging Eigen skips
upstream's patch step, so the SSE2 int64 compare fix it carries for
Eigen is applied in src_prepare.

commit 12441e115e7db7b6a04e18e61e8f6c28836635df
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 28 21:02:34 2026 +0100

sci-libs/onnxruntime: add an ABI subslot to 1.31.1

Upstream exports every symbol under one VERS_$ node and keeps no
older ones, while the SONAME stays libonnxruntime.so.1. Binaries linked
against 1.31.0 therefore fail to load against 1.31.1 ("version
'VERS_1.31.0' not found"), and with a plain SLOT="0" the := dependency
in sci-ml/sherpa-onnx never fires. A 0/$ subslot makes portage
rebuild := consumers on every release. 1.30.0-r2 and 1.31.0 keep their
slot: moving from either to 1.31.1 changes the subslot anyway, and a
revision would mean a long CUDA rebuild for no code change.

commit 2610d4bd99da242d42dc31948e27d9f5ee1cd5a6
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 28 17:48:44 2026 +0100

sci-libs/onnxruntime: drop 1.29.0-r4

Keep the last three releases.

commit 5418a36cac5349a84d8b346aec9f38a02d561cc2
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Oct 28 17:48:22 2026 +0100

sci-libs/onnxruntime: add 1.31.1

A version bump and one cherry-picked fix; the dependency pins and
build files are unchanged, so the 1.31.0 patches apply as they are.

commit a854e1dc17882dd2215ab83e7335ff2ae1f0a4bc
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Oct 9 14:32:57 2026 +0200

sci-libs/onnxruntime: drop 1.28.0-r3

commit 44af92197d758e84cf665ee20c5bf171c213ffb7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Oct 9 14:32:33 2026 +0200

sci-libs/onnxruntime: let USE=cuda configure without CUDAARCHS, revbump

With CUDAARCHS empty the ebuilds passed
-DCMAKE_CUDA_ARCHITECTURES=all-major. Upstream resolves that keyword by
unsetting the variable and then testing whether it is defined. A value
given with -D lives in the cache and stays defined, so the literal
reached the architecture parser and configure aborted. Leave the
variable alone instead: upstream then selects the set the keyword
stood for.

That set includes SM90, for which 1.30.0 fetches DeepGEMM during
configure; prestage it. Prestaging CUTLASS had also bypassed the patch
upstream applies to it in the same step, so src_prepare applies it.

Revision bump: a CUDA build with CUDAARCHS set worked before and now
compiles against the patched CUTLASS. Both versions were taken through
the configure phase only.

commit a328e1edb2a09ba4e5b73aae546d2313dfee275f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Oct 9 14:32:33 2026 +0200

sci-libs/onnxruntime: add 1.31.0

Three patches are rebased. The float6 one also raises the bound in the
new onnxruntime_type_conversion.h and extends two test tables sized
from the ONNX enum. A CUDA header new in this release lacks <cstdint>.

The CUDA setup differs from 1.30.0-r1: an empty CUDAARCHS leaves the
architecture list to upstream, DeepGEMM is prestaged for SM90 and newer
targets, and the prestaged CUTLASS gets upstream's own patch. The CUDA
provider was compiled only in part; a complete CUDA build is still to
be run.

commit f199223938469abf38148a17f356cd560b693031
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Oct 8 16:38:50 2026 +0200

sci-libs/onnxruntime: stop requiring gcc 15 for cuda

The cuda build takes its compiler from cuda_gccdir, the newest gcc the
installed toolkit accepts. These ebuilds already require toolkit 13.4,
which accepts gcc up to 16, so gcc:15 pulled in a second compiler where
the current one was usable.

A gcc dependency is still needed: the toolkit's own can be met by
clang alone, and cuda_gccdir dies without a usable gcc. Use the bound
the toolkit ebuild uses.

Build-time only; no revision bump.

commit 3cf550a4ba530b3dc9678549613a5c8f613d13b3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 26 20:11:05 2026 +0200

sci-libs/onnxruntime: revbump 1.28.0 for the current Abseil

The CUDA build of 1.28.0 still shadowed the system Abseil headers with a
patched copy of the 20250814.1 sources and pinned the dependency to that
exact release. Default-USE installs never touch it, but asking for CUDA
downgraded Abseil and rebuilt everything linked to it, exactly as the
two newer versions did before the same fix.

Apply that fix here: CUDA 13.4's nvcc accepts the installed headers, so
the bundled sources, the include-path override and the pin go, and the
toolkit dependency gains the matching floor. That removes the last user
of the nvcc header patch and of the Abseil tarball, so both go too.
Unlike the newer versions, 1.28.0 predates the consteval enum checks
ONNX 1.23 trips, so it needs no float6 patch.

Bring the rest of the ebuild level with its successors while it is open.
CUDA builds lacked the four-job cap the newer versions carry, so a host
with a high MAKEOPTS could exhaust memory at more than 3 GiB per nvcc
job. The ONNX floor still admitted the unvalidated 1.20.1 fallback its
own comment said to retire; it now matches the 1.22.0 upstream pins in
cmake/deps.txt. And the system-library patches for the long-dropped
1.26.0 and 1.27.0 had lost their last reference.

commit 9757c3684b531ecf96124bc0a797f71b81a0f9c4
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Sep 26 13:22:38 2026 +0200

sci-libs/onnxruntime: build against the current Abseil and ONNX

Neither retained version could be installed against these dependencies
as they stand.

CUDA builds shadowed the system Abseil headers with a patched copy of
the 20250814.1 sources, because the nvcc shipped with CUDA 13.3 rejected
the lifetime-bound insert_or_assign declarations in raw_hash_map.h.
Objects compiled that way still link the system library, so header and
library had to be the same release. That exact pin is what downgraded
Abseil and drew protobuf, re2, OpenCV, ONNX, sentencepiece and caffe2
into rebuilds, the last of which collided with the installed PyTorch.
CUDA 13.4's nvcc takes those declarations unchanged, so the bundled
sources, the include-path override and the pin all go, and the toolkit
dependency gains a floor at the release that allows it.

Separately, ONNX 1.23 appends FLOAT6E2M3 and FLOAT6E3M2 to
TensorProto::DataType. Two consteval self-checks tie onnxruntime's
tensor element-type tables to the size of that enum and both stop
resolving once it grows: the element-size table is declared with the
enum's ARRAYSIZE but initialised only through INT2, and the
ONNX-to-runtime mapping asserts a bijection against a public enum that
still ends at FLOAT8E8M0. Append both formats to the public enum,
leaving every existing value where it is, move the bijection's bound to
the new last entry, and record both as byte-granular in the element-size
table the way the packed sub-byte types already there are. Upstream
still pins ONNX 1.22, where none of this arises, so the floor moves to
the release that added the types.

One patch serves both versions and is named for the older of the two, as
the rest of this FILESDIR is.

Both versions get a revision, because both changes alter what lands on
disk: the patch adds two enumerators to the installed public header, and
CUDA objects now compile against the system Abseil headers rather than
the bundled copy. The Abseil subslot operator happens to force the
rebuild for CUDA installs, since the pin recorded the old subslot, but
nothing forces it for -cuda ones -- ONNX carries no subslot to track.

commit 9f02bfb5883bffaaef299623d70c9de4b35ac044
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 11 02:42:45 2026 +0200

sci-libs/onnxruntime: add 1.30.0

The release still needs the overlay's system-library integration,
while its generated patch must remain version-specific.

commit bb9ce83c51cba8733531c0f4898b2d901edfa261
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 11 02:37:25 2026 +0200

sci-libs/onnxruntime: fix 1.29.0-r2 Python tests

The Python suite must resolve the staged native library instead of
an incidental source-tree artifact.

commit 8d9d5b6d6d78383ba2958a0cf79abc7e01c06ab9
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 11 02:37:22 2026 +0200

sci-libs/onnxruntime: fix 1.28.0-r2 Python tests

The Python suite must resolve the staged native library instead of
an incidental source-tree artifact.

commit 36171520bf45be64d5905b781c30d9032b494084
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:57:46 2026 +0200

sci-libs/onnxruntime: trim 1.28.0-r2 comment prose

commit 7a60fff0a44e893cfd735fa2b63eafaaf0afc610
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:56:45 2026 +0200

sci-libs/onnxruntime: trim 1.29.0-r2 comment prose

commit fa0c85628af961e27659099d5afd9c4013a758ff
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 17 01:30:09 2026 +0200

sci-libs/onnxruntime: pin C/C++ compiler to gcc:15 for 1.28.0-r2 cuda

The cuda block pinned only CUDAHOSTCXX, so on a host whose default gcc is
newer than the CUDA toolkit supports (gcc 16 vs CUDA 13's gcc 15
ceiling), ordinary C++ translation units compiled with the system gcc
and the CUDA host translation units compiled with gcc 15 link two
different libstdc++ ABIs into one shared library. Resolve cuda_gccdir
once and pin CC, CXX, and CUDAHOSTCXX to that slot together, matching
1.29.0-r2, so all objects share one ABI.

commit ab5b9a60ae04326175c870384c06da63f2558655
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 16 11:16:44 2026 +0200

sci-libs/onnxruntime: derive the CUDA host compiler from cuda_gccdir

Both -r2 ebuilds hardcoded /usr/bin/gcc-15 and /usr/bin/g++-15 to keep nvcc off gcc
16 (which CUDA 13 rejects), baking in the exact moving target the cuda.eclass helper
tracks. Replace the hardcodes with cuda_gccdir, which selects the newest gcc slot the
active toolkit supports: 1.29.0-r2 sets CC/CXX/CUDAHOSTCXX from its bindir so
ordinary C++, CUDA host compilation, and linking share one libstdc++ ABI, and
1.28.0-r2 sets CUDAHOSTCXX from it, matching the ollama pattern. The cuda? sys-
devel/gcc:15 BDEPEND stays as the single slot knob and the guarantee that a
compatible gcc is installed; a future CUDA bump now moves one line instead of four.

commit a51e5ba9237312c089026e887eff489ab8a293a1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sun Aug 16 11:01:47 2026 +0200

sci-libs/onnxruntime: restore sympy runtime dep for 1.29.0-r2

1.29.0-r2's python? block dropped dev-python/sympy, which the shipped
symbolic_shape_infer tool imports (onnxruntime/python/tools/
symbolic_shape_infer.py). Upstream exposes sympy as the optional
[symbolic] extra, but with no USE flag to gate the tool it ImportErrors
without the dep, so pull it unconditionally as 1.28.0-r2 already does.

Also drop the inert throttle wrapper around cmake_src_install (the
install phase copies files and spawns no nvcc) and remove a stray
leading tab on 1.28.0-r2's SRC_URI assignment.

commit eb639b57264abd960b8402b2452b82b8e170661c
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 15 02:50:33 2026 +0200

sci-libs/onnxruntime: fix CUDA build and dependencies

CUDA 13 requires GCC 15, so use it consistently for C, C++, and nvcc host compilation
instead of mixing libstdc++ ABIs. Add the missing test-only framework and provider-
option link edges exposed by shared CUDA builds. Align the Python dependency
declarations and ONNX floor with the metadata bundled in the 1.29.0 source.

commit f725e0b23a02819fe09ddf28193137a962d18060
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 20:46:13 2026 +0200

sci-libs/onnxruntime: drop unused coloredlogs dep

The 1.29.0 Python requirements no longer declare coloredlogs and the shipped Python
modules do not import it.

commit 180cd3606965d17b3ca7f39cbb2ecbca58af4e4f
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 10:03:41 2026 +0200

sci-libs/onnxruntime: enforce Python protobuf floor

Upstream 1.29.0 installs requirements.txt verbatim, which requires protobuf 4.25.8 or
newer. Express that floor so older Python protobuf versions cannot satisfy the
runtime dependency.

commit d09a92cc75cd000955a9cfcd3cf0857bc3a2f838
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 05:11:06 2026 +0200

sci-libs/onnxruntime: limit CUDA build concurrency

Flash-attention translation units can each require several GiB of memory, making
unrestricted parallel CUDA compilation prone to compiler OOM termination. Apply the
same bounded concurrency to compile and install while preserving lower user limits
and leaving non-CUDA builds unchanged.

commit 478ac46e77dfffe3cc2b9f6425a00319d5122ae3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Aug 13 20:24:01 2026 +0200

sci-libs/onnxruntime: add CUDA execution provider

CUDA builds require the toolkit, cuDNN, upstream-pinned CUTLASS and cuDNN Frontend
sources, and an Abseil header compatibility correction for nvcc. Install the provider
library with the Python bindings and retain the matching system Abseil ABI.

commit 1a8423f40555da94e216276d092ff3bafc456f0a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Aug 12 09:28:05 2026 +0200

sci-libs/onnxruntime: drop 1.27.1-r1

Retention: keep the last two (1.28.0-r1, 1.29.0). The only in-tree
consumer, sci-ml/sherpa-onnx, depends on sci-libs/onnxruntime:= with no
version floor, so nothing anchors 1.27.1.

commit 138e6566de041d79b232e33dfeabd7918a170a3a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Aug 12 10:36:00 2026 +0200

sci-libs/onnxruntime: add 1.29.0

Copy of 1.28.0-r1. eigen pin (1d8b82b) and the bundled onnx pin (v1.22.0)
are unchanged from 1.28.0, so the onnx floor stays at >=sci-ml/onnx-1.20.1
(::gentoo still tops out at 1.20.1). use-system-libraries is reused from
1.28.0 unchanged -- the 1.29.0 onnxruntime_external_deps.cmake drift falls
outside the hunks it touches -- matching how 1.27.1 reuses the 1.27.0 patch.

1.29.0 adds a non-Windows telemetry path (1DS SDK plus a bundled static
curl + mbedTLS) as three new deps.txt FetchContent entries, all gated
behind onnxruntime_USE_TELEMETRY (default OFF, not enabled here), so they
never fetch under the network sandbox.

commit 7685bad266f589fc2c500245b687e0c4154463eb
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 10:38:17 2026 +0200

sci-libs/onnxruntime: floor onnx and record the version gap, revbump

The onnx dependency was unversioned while upstream pins an exact
release: cmake/deps.txt names onnx v1.21.0 for onnxruntime 1.27.1 and
v1.22.0 for 1.28.0. Our use-system-libraries patch rewrites that
FetchContent entry to FIND_PACKAGE_ARGS ... REQUIRED, so the system
copy is what actually gets linked, with no version check of its own.

The correct floor cannot be written today. sci-ml/onnx lives only in
::gentoo and tops out at 1.20.1, so >=sci-ml/onnx-1.22.0 would make the
package uninstallable rather than correct.

Floored at 1.20.1 instead, which is strictly better than bare: it keeps
out 1.18.0-r1, the other version in the tree and further still from
what upstream builds against. Whether 1.20.1 is genuinely sufficient is
not something this change establishes, and the inline note says so
along with the version each ebuild really wants, so a later audit can
raise it rather than re-derive the situation.

commit d8d65410da1776e7c674353ce9eb7f2305bb9bc8
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 25 10:37:24 2026 +0200

sci-libs/onnxruntime: drop 1.27.0

commit 78d5f235d1e75aaf518240c0cf58c01e3eedbc76
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 25 10:37:11 2026 +0200

sci-libs/onnxruntime: add 1.28.0

Regenerate the use-system-libraries patch for 1.28.0: upstream split the cpuinfo
FetchContent block into three branches (Windows-ARM / Linux / else) in the 1.28
cmake, adding a Linux elseif that the 1.27 patch did not cover. Rename to
onnxruntime-1.28.0-use-system-libraries.patch and make all three cpuinfo
FIND_PACKAGE_ARGS REQUIRED so the Linux build cannot silently vendor cpuinfo. Eigen
pin unchanged (deps.txt still 1d8b82b0).

commit 46b8b8af7cfa59db009ce9f3ef072eadd81326a7
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 11 09:20:24 2026 +0200

sci-libs/onnxruntime: drop 1.26.0

commit 6df016beddb458007d0674cc710fd7838efb17fb
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 11 09:20:23 2026 +0200

sci-libs/onnxruntime: add 1.27.1

commit 5856daa5d1f9c35c27d7f4da1b54b7db9fe7af7b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jun 20 10:01:12 2026 +0200

sci-libs/onnxruntime: add 1.27.0

Rebase the use-system-libraries patch for 1.27.0: upstream collapsed the
Microsoft.GSL FetchContent declaration from a USE_CUDA/else split into a
single block (and moved its sub-patch 1064->1213), so the two old GSL
hunks become one. The other system-lib REQUIRED hunks apply unchanged;
the eigen pin is unchanged (cmake/deps.txt still references the same
commit). Build-verified with USE=python: configure resolves all system
libraries (no FetchContent download under network-sandbox) and
libonnxruntime.so.1.27.0 links.

commit 9e48c0af7bcdcfd491539d61d6341a9c4e8def9a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 18:01:58 2026 +0200

sci-libs/onnxruntime: Keyword 1.26.0 for ~arm64

ONNX Runtime supports aarch64 as a first-class target. Builds the
default C++ configuration (python USE is off by default and pulls
::gentoo deps that still lack arm64, e.g. coloredlogs, so it stays
opt-in). aarch64-portable.

commit 6cefd2e68968302cbaf5f7f888d77248130b2159
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 15:15:40 2026 +0200

sci-libs/onnxruntime: new package, add 1.26.0

Microsoft's cross-platform ML inference runtime. Source build with
system libs (abseil, protobuf, re2, onnx, boost, etc.), USE=python
for Python bindings.

Forked from ::guru @ same version; ebuild + 3 patches copied as-is
(relax-flatbuffers-dep, no-werror, use-system-libraries). Build
verified end-to-end with PYTHON_SINGLE_TARGET=python3_13
+ USE=python; Python `import onnxruntime` resolves to 1.26.0.

Pulled in to support sci-ml/sherpa-onnx (source build).