Script 'mail_helper' called by obssrc
Hello community,
here is the log from the commit of package python-xgrammar for openSUSE:Factory
checked in at 2026-09-09 16:21:05
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Comparing /work/SRC/openSUSE:Factory/python-xgrammar (Old)
and /work/SRC/openSUSE:Factory/.python-xgrammar.new.1265 (New)
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Package is "python-xgrammar"
Wed Sep 9 16:21:05 2026 rev:5 rq:1376540 version:0.2.6
Changes:
--------
--- /work/SRC/openSUSE:Factory/python-xgrammar/python-xgrammar.changes
2026-09-08 16:55:17.338903897 +0200
+++
/work/SRC/openSUSE:Factory/.python-xgrammar.new.1265/python-xgrammar.changes
2026-09-09 16:23:08.153993221 +0200
@@ -1,0 +2,26 @@
+Wed Sep 9 06:21:19 UTC 2026 - Martin Pluskal <[email protected]>
+
+- Update to 0.2.6:
+ * Harden C++ boundary so invalid args, corrupted serialized
+ data and worker-thread errors raise exceptions instead of
+ crashing the process
+ * Reject NUL in regex instead of segfaulting; preserve
+ tokenizer JSON round-trips including binary vocabulary
+ * EBNF/Lark: capture, lazy matching, suffix/stop conditions,
+ token/character budgets, sampling temperature, regex flags,
+ and cached GrammarCompiler.compile_lark
+ * JSON Schema: positional prefixItems, CR whitespace, and
+ fixes for repeat-edge metadata and serialization keys
+ * Structural tags for Cohere XML, Kimi K3 and EXAONE 4.0;
+ max_tokens/max_chars for AnyTextFormat/AnyTokensFormat
+ * NPU token-bitmask backend; wheel CMake build type is
+ Release, so the native extension is no longer full of
+ RelWithDebInfo debug sections
+- Switch the source back to the PyPI sdist (now ships the
+ 3rdparty tree the C++ build needs) and drop _service
+- Drop SKBUILD_WHEEL_EXCLUDE: upstream skips installing the
+ static library into the wheel
+- Raise apache-tvm-ffi floor to >= 0.1.10 and scikit-build-core
+ to >= 0.10.0
+
+-------------------------------------------------------------------
Old:
----
_service
xgrammar-0.2.5.post1.tar.gz
New:
----
xgrammar-0.2.6.tar.gz
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Other differences:
------------------
++++++ python-xgrammar.spec ++++++
--- /var/tmp/diff_new_pack.gAp1Tz/_old 2026-09-09 16:23:08.934025822 +0200
+++ /var/tmp/diff_new_pack.gAp1Tz/_new 2026-09-09 16:23:08.937025947 +0200
@@ -17,22 +17,20 @@
%{?sle15_python_module_pythons}
Name: python-xgrammar
-Version: 0.2.5.post1
+Version: 0.2.6
Release: 0
Summary: Efficient, Flexible and Portable Structured Generation
License: Apache-2.0
URL: https://xgrammar.mlc.ai/
-# Upstream's PyPI sdist is unreliable (absent for 0.2.4 and 0.2.5), so the
-# tarball is generated from the git tag via _service, with the 3rdparty
-# submodules the C++ build needs.
# https://github.com/mlc-ai/xgrammar
-Source: xgrammar-%{version}.tar.gz
-BuildRequires: %{python_module apache-tvm-ffi >= 0.1.9}
+# PyPI sdist ships the 3rdparty sources the C++ build needs.
+Source:
https://files.pythonhosted.org/packages/source/x/xgrammar/xgrammar-%{version}.tar.gz
+BuildRequires: %{python_module apache-tvm-ffi >= 0.1.10}
BuildRequires: %{python_module devel}
BuildRequires: %{python_module numpy}
BuildRequires: %{python_module pip}
BuildRequires: %{python_module pydantic}
-BuildRequires: %{python_module scikit-build-core}
+BuildRequires: %{python_module scikit-build-core >= 0.10.0}
BuildRequires: %{python_module setuptools}
BuildRequires: %{python_module torch >= 1.10.0}
BuildRequires: %{python_module transformers >= 4.38.0}
@@ -45,7 +43,9 @@
BuildRequires: libopenblas_pthreads0
BuildRequires: ninja
BuildRequires: python-rpm-macros
-Requires: python-apache-tvm-ffi >= 0.1.9
+# triton is a Linux x86_64 PyPI dep for the CUDA bitmask kernel only; CPU
+# auto-selects the bundled CPU backend. python-triton is not in Factory.
+Requires: python-apache-tvm-ffi >= 0.1.10
Requires: python-numpy
Requires: python-pydantic
Requires: python-torch >= 1.10.0
@@ -64,12 +64,6 @@
%build
export CMAKE_GENERATOR=Ninja
-# Exclude the internal C++ build artifacts (headers plus the static
-# libxgrammar.a, unused at runtime) from the wheel instead of deleting them in
-# %%install: the fat-LTO .a is not reproducible (random .gnu.lto_* section
-# names, ar timestamps) and deleting it after the fact left its varying
-# checksum behind in dist-info/RECORD.
-export SKBUILD_WHEEL_EXCLUDE='**.a;**.h'
%pyproject_wheel
%install
++++++ xgrammar-0.2.5.post1.tar.gz -> xgrammar-0.2.6.tar.gz ++++++
++++ 35846 lines of diff (skipped)
