sci-ml/pytorch
Tensors and Dynamic neural networks in Python
Screenshots
-
pytorch-2.13.0-r91~amd64 ~arm64python_single_target_python3_12 python_single_target_python3_13 python_single_target_python3_14 debug
View 
Download 
Browse License: BSD Overlay: stuff -
pytorch-2.13.0-r3~amd64 ~arm64cuda cusparselt distributed fbgemm flash gloo kineto memefficient mimalloc mkl mpi nccl nnpack +numpy onednn openblas opencl openmp qnnpack rocm xnnpack python_single_target_python3_12 python_single_target_python3_13 python_single_target_python3_14 debug +amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +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: BSD Overlay: gentoo
ChangeLog
commit 6932a03e3f2f57ff08f00a0f6b4110bded397373
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 8 12:01:30 2026 +0200
sci-ml/pytorch: add 2.11.0-r90
Restores the thin torch-2.11 frontend that vllm (0.24/0.25/0.26) pins via
~sci-ml/pytorch-2.11.0. ::gentoo dropped torch 2.11 and our fork jumped
straight to 2.13.0, leaving vllm unresolvable. Adapted from 2.13.0-r91
(same caffe2/pytorch split: dontbuildagain reuse of the caffe2 build)
against the existing sci-ml/caffe2-2.11.0-r90 backend. One deviation from
torch 2.11.0's Requires-Dist: its setuptools<82 cap is unsatisfiable on a
current system (it forces setuptools-79, colliding with consumers that
need >=80 like dev-python/ipython, and nothing exists in [80,82)), so
relax to the same >=77.0.3 floor our 2.13.0-r91 frontend already uses.
Both frontend patches apply to 2.11.0.
commit ca8261c4ffaadde41193177e0267b04df6593353
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 3 14:49:49 2026 +0200
*/*: normalize metadata.xml whitespace
Use the overlay-wide two-space XML indentation convention and remove the remaining trailing whitespace so metadata stays consistent and diff-safe.
commit 71285ad8b369f6bdd6b64718b23b4574dda4273d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 08:54:04 2026 +0200
sci-ml/pytorch: declare torch's unconditional runtime deps, revbump
RDEPEND listed sympy and typing-extensions, both without floors, and
nothing else. torch declares seven unconditional Requires-Dist. The
five that were missing entirely are filelock, fsspec, jinja2, networkx
and setuptools; sympy and typing-extensions were present but unbounded
where upstream floors them at 1.13.3 and 4.10.0.
Only typing-extensions is pulled in by `import torch`, which is why the
gap went unnoticed: the rest are reached lazily. sympy and networkx
come in through torch.fx, jinja2 and filelock through the inductor
codegen and its compile cache, fsspec through torch.load and save on
remote paths, setuptools through torch.utils.cpp_extension. Upstream
marks none of them as an extra -- the ones that are optional are listed
separately under extra == "optree", "opt-einsum" and "pyyaml", and
those are correctly absent here. Anyone reaching torch.compile without
them installed gets an ImportError rather than a dependency error.
caffe2 does not cover them either; its RDEPEND is C++ and system
libraries plus a USE-gated numpy.
The block is written to mirror Requires-Dist verbatim, floors included,
so it can be diffed against the built torch-2.13.0.dist-info/METADATA
instead of re-derived.
Revbump because the package is installable at -r90 and portage does not
re-resolve dependencies for an installed package without --changed-deps;
without a new revision an existing install would keep the old set. -r91
stays inside the band this fork uses, below the -r99 that marks a
straight fork of the gentoo ebuild.
gentoo's own sci-ml/pytorch-2.12.0 carries the same incomplete RDEPEND,
so this is worth reporting upstream rather than only fixing here.
commit 4cb9a73ef8fe789add28e419383e788e8e5db826
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 18 16:40:59 2026 +0200
sci-ml/pytorch: add 2.13.0-r90
Python-bindings metapackage for sci-ml/caffe2-2.13.0 (RDEPEND ~caffe2-$).
Fork of ::gentoo's pytorch-2.12.0, which pairs with caffe2 at 2.12.0; carries
its two patches (dontbuildagain, cpp-extension-multilib) and the build-dir /
libdir seds unchanged.
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Aug 8 12:01:30 2026 +0200
sci-ml/pytorch: add 2.11.0-r90
Restores the thin torch-2.11 frontend that vllm (0.24/0.25/0.26) pins via
~sci-ml/pytorch-2.11.0. ::gentoo dropped torch 2.11 and our fork jumped
straight to 2.13.0, leaving vllm unresolvable. Adapted from 2.13.0-r91
(same caffe2/pytorch split: dontbuildagain reuse of the caffe2 build)
against the existing sci-ml/caffe2-2.11.0-r90 backend. One deviation from
torch 2.11.0's Requires-Dist: its setuptools<82 cap is unsatisfiable on a
current system (it forces setuptools-79, colliding with consumers that
need >=80 like dev-python/ipython, and nothing exists in [80,82)), so
relax to the same >=77.0.3 floor our 2.13.0-r91 frontend already uses.
Both frontend patches apply to 2.11.0.
commit ca8261c4ffaadde41193177e0267b04df6593353
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Aug 3 14:49:49 2026 +0200
*/*: normalize metadata.xml whitespace
Use the overlay-wide two-space XML indentation convention and remove the remaining trailing whitespace so metadata stays consistent and diff-safe.
commit 71285ad8b369f6bdd6b64718b23b4574dda4273d
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jul 27 08:54:04 2026 +0200
sci-ml/pytorch: declare torch's unconditional runtime deps, revbump
RDEPEND listed sympy and typing-extensions, both without floors, and
nothing else. torch declares seven unconditional Requires-Dist. The
five that were missing entirely are filelock, fsspec, jinja2, networkx
and setuptools; sympy and typing-extensions were present but unbounded
where upstream floors them at 1.13.3 and 4.10.0.
Only typing-extensions is pulled in by `import torch`, which is why the
gap went unnoticed: the rest are reached lazily. sympy and networkx
come in through torch.fx, jinja2 and filelock through the inductor
codegen and its compile cache, fsspec through torch.load and save on
remote paths, setuptools through torch.utils.cpp_extension. Upstream
marks none of them as an extra -- the ones that are optional are listed
separately under extra == "optree", "opt-einsum" and "pyyaml", and
those are correctly absent here. Anyone reaching torch.compile without
them installed gets an ImportError rather than a dependency error.
caffe2 does not cover them either; its RDEPEND is C++ and system
libraries plus a USE-gated numpy.
The block is written to mirror Requires-Dist verbatim, floors included,
so it can be diffed against the built torch-2.13.0.dist-info/METADATA
instead of re-derived.
Revbump because the package is installable at -r90 and portage does not
re-resolve dependencies for an installed package without --changed-deps;
without a new revision an existing install would keep the old set. -r91
stays inside the band this fork uses, below the -r99 that marks a
straight fork of the gentoo ebuild.
gentoo's own sci-ml/pytorch-2.12.0 carries the same incomplete RDEPEND,
so this is worth reporting upstream rather than only fixing here.
commit 4cb9a73ef8fe789add28e419383e788e8e5db826
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 18 16:40:59 2026 +0200
sci-ml/pytorch: add 2.13.0-r90
Python-bindings metapackage for sci-ml/caffe2-2.13.0 (RDEPEND ~caffe2-$).
Fork of ::gentoo's pytorch-2.12.0, which pairs with caffe2 at 2.12.0; carries
its two patches (dontbuildagain, cpp-extension-multilib) and the build-dir /
libdir seds unchanged.