Script 'mail_helper' called by obssrc
Hello community,

here is the log from the commit of package python-timm for openSUSE:Factory 
checked in at 2026-08-25 13:20:01
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-timm (Old)
 and      /work/SRC/openSUSE:Factory/.python-timm.new.1258 (New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Package is "python-timm"

Tue Aug 25 13:20:01 2026 rev:2 rq:1373474 version:1.0.28

Changes:
--------
--- /work/SRC/openSUSE:Factory/python-timm/python-timm.changes  2026-08-24 
15:43:37.766732772 +0200
+++ /work/SRC/openSUSE:Factory/.python-timm.new.1258/python-timm.changes        
2026-08-25 13:20:15.156302300 +0200
@@ -1,0 +2,31 @@
+Mon Aug 24 18:07:15 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 1.0.28:
+  * New models: DINOv3 ConvNeXt/ViT, Gemma4 ViT, TIPSv2, EUPE,
+    CSATv2, MobileCLIP-2 and MetaCLIP-2 encoders, SigLIP-2
+    NaFlex weights
+  * Add the Muon optimizer plus AdaMuon/NAdaMuon variants, and
+    a cautious mode for AdamP and SGDP
+  * Add a lightweight task abstraction with logits, feature and
+    token distillation support (new timm.task subpackage)
+  * Add ROPE support to NaFlexViT and set_input_size() to EVA
+    models; extend NaFlexViT patch layout and forward_intermediates
+  * Add differential attention, coordinate attention and the
+    LsePlus/SimPool/PRR pooling modules
+  * Harden checkpoint loading: default weights_only=True and
+    address several pickle security concerns
+  * Support device/dtype factory kwargs and meta-device model
+    initialisation
+  * Fix LayerScale ignoring init_values, reg_token weight decay
+    grouping, Hiera SDPA dispatch and a ZeroDivisionError in
+    AverageMeter
+  * Remove APEX AMP support and reduce torch.jit usage ahead of
+    its deprecation
+  * Compat break: correct QKV vs MLP bias in ParallelScalingBlock
+    and DiffParallelScalingBlock; no shipped weights are affected
+  * Compat break: MobileNet-V5 gains a stem bias and uses the tanh
+    GELU approximation, breaking pre-1.0.17 weights
+- Extend the %check smoke test with a forward pass and a guard for
+  the SigLIP ViT that python-sglang builds via timm.create_model()
+
+-------------------------------------------------------------------

Old:
----
  timm-1.0.16.tar.gz

New:
----
  timm-1.0.28.tar.gz

++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

Other differences:
------------------
++++++ python-timm.spec ++++++
--- /var/tmp/diff_new_pack.gVWESU/_old  2026-08-25 13:20:16.362344884 +0200
+++ /var/tmp/diff_new_pack.gVWESU/_new  2026-08-25 13:20:16.371345202 +0200
@@ -18,7 +18,7 @@
 
 %{?sle15_python_module_pythons}
 Name:           python-timm
-Version:        1.0.16
+Version:        1.0.28
 Release:        0
 Summary:        PyTorch Image Models
 License:        Apache-2.0
@@ -59,8 +59,18 @@
 
 %check
 # Upstream pytest downloads pretrained weights and needs network.
-# Smoke-test the import and model registry instead.
-%python_expand PYTHONPATH=%{buildroot}%{$python_sitelib} $python -B -c "import 
timm; assert timm.__version__ == '%{version}', timm.__version__; assert 
timm.list_models()"
+# Smoke-test import, the model registry and a real forward pass instead.
+cat > smoke.py <<'EOF'
+import timm, torch
+assert timm.__version__ == "%{version}", timm.__version__
+assert timm.list_models()
+# guard the SigLIP ViT that python-sglang builds via timm.create_model()
+assert "vit_so400m_patch14_siglip_384" in timm.list_models()
+m = timm.create_model("resnet18", pretrained=False, num_classes=10).eval()
+with torch.no_grad():
+    assert m(torch.randn(1, 3, 64, 64)).shape == (1, 10)
+EOF
+%python_expand PYTHONPATH=%{buildroot}%{$python_sitelib} $python -B smoke.py
 
 %files %{python_files}
 %license LICENSE

++++++ timm-1.0.16.tar.gz -> timm-1.0.28.tar.gz ++++++
++++ 55072 lines of diff (skipped)

Reply via email to