Script 'mail_helper' called by obssrc
Hello community,
here is the log from the commit of package python-torchvision for
openSUSE:Factory checked in at 2026-09-18 22:05:11
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-torchvision (Old)
and /work/SRC/openSUSE:Factory/.python-torchvision.new.383539 (New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-torchvision"
Fri Sep 18 22:05:11 2026 rev:2 rq:1378659 version:0.29.0
Changes:
--------
--- /work/SRC/openSUSE:Factory/python-torchvision/python-torchvision.changes
2026-08-21 17:02:21.927269426 +0200
+++
/work/SRC/openSUSE:Factory/.python-torchvision.new.383539/python-torchvision.changes
2026-09-18 22:05:40.899438734 +0200
@@ -1,0 +2,35 @@
+Wed Sep 9 16:49:56 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 0.29.0 (needs python-torch 2.14, not yet in the tree):
+ * ABI-stable against torch 2.14: upstream states torchvision
+ 0.29 stays compatible with future torch releases
+ * Deprecate the torchvision.io image decoders and encoders,
+ moved to TorchCodec >= 0.16 (removal in a future release)
+ * Fix deformable convolution kernels always running on the
+ default stream; gradient overaccumulation in ROI ops on MPS;
+ JPEG transform on non-contiguous batches
+- Require python-torch >= 2.14 and < 2.15 for build and runtime
+ (upstream PyPI metadata pins torch == 2.14.0)
+
+-------------------------------------------------------------------
+Sun Aug 30 10:34:09 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 0.28.0 (built against python-torch 2.13.0):
+ * Quantized nms/roi_align moved off torch's deprecated
+ quantized dtypes onto new torchvision::qnms /
+ torchvision::qroi_align ops dispatched from Python:
+ torch.ops.torchvision.nms() no longer takes quantized
+ tensors, and the quantized paths no longer script
+ * v2 geometry and auto-augment transforms accept interpolation
+ as a string and now default to the equivalent strings rather
+ than InterpolationMode members
+ * Fix F.resize on tv_tensors.Mask ignoring its interpolation
+ argument, silently downgrading NEAREST_EXACT to NEAREST
+ * wrap() keeps metadata for custom TVTensor subclasses
+- Drop fix-gif-decoder-oob.patch: merged upstream in 0.28.0,
+ already shipped here as a backport in 0.27.0
+- Require python-torch >= 2.13.0 and < 2.14 for build and runtime:
+ rpm sees only unversioned libtorch_cpu.so()/libc10.so(), and
+ libtorch's C++ ABI moves at torch minor boundaries in practice
+
+-------------------------------------------------------------------
Old:
----
fix-gif-decoder-oob.patch
torchvision-0.27.0.tar.gz
New:
----
torchvision-0.29.0.tar.gz
----------(Old B)----------
Old: * wrap() keeps metadata for custom TVTensor subclasses
- Drop fix-gif-decoder-oob.patch: merged upstream in 0.28.0,
already shipped here as a backport in 0.27.0
----------(Old E)----------
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Other differences:
------------------
++++++ python-torchvision.spec ++++++
--- /var/tmp/diff_new_pack.YdMnzd/_old 2026-09-18 22:05:43.737557680 +0200
+++ /var/tmp/diff_new_pack.YdMnzd/_new 2026-09-18 22:05:43.739557763 +0200
@@ -16,9 +16,19 @@
#
+# rpm sees only unversioned libtorch_cpu.so()/libc10.so(), so nothing catches a
+# mismatched torch. libtorch's C++ ABI moves at torch minor boundaries in
+# practice (upstream guarantees none) - the granularity upstream's README
matrix
+# pairs at - hence a range, not "=": a torch patch release stays installable,
+# 2.14 does not. Upstream ships PYTORCH_VERSION_GE/_LT for exactly this
+# third-party case. Not expressible in Requires-Dist: our python-torch's wheel
+# version is a pre-release string (e.g. 2.13.0a0+gitunknown), so a PEP 440
+# pin there is unsatisfiable.
+%define torch_ver_ge 2.14.0
+%define torch_ver_lt 2.15
%{?sle15_python_module_pythons}
Name: python-torchvision
-Version: 0.27.0
+Version: 0.29.0
Release: 0
Summary: Image datasets, models and transforms for PyTorch
# Legal-Review-Notice: ships a bundled giflib 5.2.2 (MIT) under
@@ -26,17 +36,16 @@
License: BSD-3-Clause AND MIT
URL: https://github.com/pytorch/vision
Source0:
https://github.com/pytorch/vision/archive/refs/tags/v%{version}.tar.gz#/torchvision-%{version}.tar.gz
-# PATCH-FIX-UPSTREAM fix-gif-decoder-oob.patch gh#pytorch/vision#9520
-# Backport of 4e05dc22f5 (0.28): GIF decoder heap OOB read/write.
GHSA-vp9x-48wq-4wc3.
-Patch0: fix-gif-decoder-oob.patch
BuildRequires: %{python_module Pillow >= 5.3.0}
BuildRequires: %{python_module devel}
BuildRequires: %{python_module numpy}
BuildRequires: %{python_module packaging}
BuildRequires: %{python_module pip}
BuildRequires: %{python_module setuptools}
-BuildRequires: %{python_module torch-devel}
-BuildRequires: %{python_module torch}
+BuildRequires: %{python_module torch < %{torch_ver_lt}}
+BuildRequires: %{python_module torch >= %{torch_ver_ge}}
+BuildRequires: %{python_module torch-devel < %{torch_ver_lt}}
+BuildRequires: %{python_module torch-devel >= %{torch_ver_ge}}
BuildRequires: %{python_module wheel}
BuildRequires: fdupes
BuildRequires: gcc
@@ -51,7 +60,8 @@
# setuptools wheel; declare them by hand.
Requires: python-Pillow
Requires: python-numpy
-Requires: python-torch
+Requires: python-torch < %{torch_ver_lt}
+Requires: python-torch >= %{torch_ver_ge}
# Match Factory python-torch (ExcludeArch: %%ix86 %%{arm})
ExcludeArch: %{ix86} %{arm}
%python_subpackages
++++++ torchvision-0.27.0.tar.gz -> torchvision-0.29.0.tar.gz ++++++
/work/SRC/openSUSE:Factory/python-torchvision/torchvision-0.27.0.tar.gz
/work/SRC/openSUSE:Factory/.python-torchvision.new.383539/torchvision-0.29.0.tar.gz
differ: char 13, line 1
