Prepare the 13.4.2 release (#2900)

* Prepare the 13.4.2 release

* Simplify release note

* Fix typo

* test(pathfinder): handle cuTENSORMg removal in cuTENSOR 2.8 (#2809)

* Add pathfinder support for cupti64_2026.3.1.dll (CTK 13.4.1) (#2896)

The nvidia-cu13 wheel now ships cupti64_2026.3.1.dll, which was not
listed in the cupti descriptor's windows_dlls, causing
test_load_nvidia_dynamic_lib[cupti] to fail under all_must_work.

---------

Co-authored-by: Ralf W. Grosse-Kunstleve <rwgkio@gmail.com>
Co-authored-by: Ralf Juengling <rjuengling@nvidia.com>

Author: 3 people

CONTRIBUTING.md

@@ -48,7 +48,7 @@ correct.** Each package matches its own tag prefix:
 
 | Package | Tag pattern |
 | --- | --- |
-| `cuda-bindings`, `cuda-python` | `v*` (e.g. `v13.4.1`) |
+| `cuda-bindings`, `cuda-python` | `v*` (e.g. `v13.4.2`) |
 | `cuda-core` | `cuda-core-v*` (e.g. `cuda-core-v1.1.0`) |
 | `cuda-pathfinder` | `cuda-pathfinder-v*` (e.g. `cuda-pathfinder-v1.6.0`) |
 
@@ -109,7 +109,7 @@ version-check failure:
    `cuda-bindings` to that same bogus version.
 2. **Stale tags** (a fork that has not fetched upstream in a while): you get a
    plausible-looking but wrong version, e.g. `13.0.4.dev650+g0d22cb44` when the
-   real latest tag is `v13.4.1`. Nothing warns you. Note there is no leading
+   real latest tag is `v13.4.2`. Nothing warns you. Note there is no leading
    `v` — the tag prefix is stripped by `tag_regex`.
 3. **No git metadata** (source zip): the build fails with
    `LookupError: setuptools-scm was unable to detect version`.

ci/test-matrix.yml

@@ -5,6 +5,6 @@ backport_branch: "12.9.x"  # keep in sync with target-branch in .github/dependab
 
 cuda:
   build:
-    version: "13.4.1"
+    version: "13.4.2"
   prev_build:
     version: "12.9.1"

ci/versions.yml

@@ -3,6 +3,10 @@
         "version": "latest",
         "url": "https://nvidia.github.io/cuda-python/cuda-bindings/latest/"
     },
+    {
+        "version": "13.4.2",
+        "url": "https://nvidia.github.io/cuda-python/cuda-bindings/13.4.2/"
+    },
     {
         "version": "13.4.1",
         "url": "https://nvidia.github.io/cuda-python/cuda-bindings/13.4.1/"

cuda_bindings/docs/nv-versions.json

@@ -0,0 +1,18 @@
+.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+.. SPDX-License-Identifier: Apache-2.0
+
+.. module:: cuda.bindings
+
+``cuda-bindings`` 13.4.2 Release notes
+======================================
+
+Bugfixes
+--------
+
+- ``cuda-bindings`` is rebuilt against CUDA toolkit 13.4.2, and thus gets all of its bugfixes.
+
+Known issues
+------------
+
+* Updating from older versions (v12.6.2.post1 and below) via ``pip install -U cuda-python`` might not work. Please do a clean re-installation by uninstalling ``pip uninstall -y cuda-python`` followed by installing ``pip install cuda-python``.
+* ``nvml.system_get_process_name`` on WSL can return incorrect values.  To work around this, set the locale to "C" before calling ``nvml.device_get_compute_running_processes_v3`` (which sets the process names) and before calling ``nvml.system_get_process_name``. ``cuda_core`` does this automatically, but users of the raw NVML API will need to do this manually.

cuda_bindings/docs/source/release/13.4.2-notes.rst

@@ -362,6 +362,7 @@ class DescriptorSpec:
         packaged_with="ctk",
         linux_sonames=("libcupti.so.13", "libcupti.so.12"),
         windows_dlls=(
+            "cupti64_2026.3.1.dll",
             "cupti64_2026.3.0.dll",
             "cupti64_2026.2.1.dll",
             "cupti64_2026.2.0.dll",

cuda_pathfinder/cuda/pathfinder/_dynamic_libs/descriptor_catalog.py

@@ -1,16 +1,20 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 import functools
 import importlib.metadata
 import re
 
+from packaging.version import Version
+
 
 @functools.cache
-def have_distribution(name_pattern: str) -> bool:
+def have_distribution(name_pattern: str, *, minimum_version: str | None = None) -> bool:
     re_name_pattern = re.compile(name_pattern)
+    parsed_minimum_version = Version(minimum_version) if minimum_version is not None else None
     return any(
         re_name_pattern.match(dist.metadata["Name"])
+        and (parsed_minimum_version is None or Version(dist.version) >= parsed_minimum_version)
         for dist in importlib.metadata.distributions()
         if "Name" in dist.metadata
     )

cuda_pathfinder/tests/local_helpers.py

@@ -4,6 +4,7 @@
 import os
 import platform
 from pathlib import Path
+from types import SimpleNamespace
 
 import pytest
 from child_load_nvidia_dynamic_lib_helper import (
@@ -123,12 +124,39 @@ def test_known_but_platform_unavailable_libname_raises_dynamic_lib_not_available
 def _is_expected_load_nvidia_dynamic_lib_failure(libname):
     if libname == "nvpl_fftw" and platform.machine().lower() != "aarch64":
         return True
+    if libname == "cutensorMg":
+        # cuTENSOR 2.8 removed cuTENSORMg in favor of cuTENSORMp.
+        return have_distribution(r"^cutensor-cu(?:12|13)$", minimum_version="2.8")
     dist_name_pattern = IMPORTLIB_METADATA_DISTRIBUTIONS_NAMES.get(libname)
     if dist_name_pattern is not None:
         return not have_distribution(dist_name_pattern)
     return False
 
 
+@pytest.mark.parametrize(
+    ("installed_distributions", "expected"),
+    [
+        ([], False),
+        ([SimpleNamespace(metadata={"Name": "cutensor-cu13"}, version="2.7.0")], False),
+        ([SimpleNamespace(metadata={"Name": "cutensor-cu12"}, version="2.8.0")], True),
+        ([SimpleNamespace(metadata={"Name": "cutensor-cu13"}, version="2.9.0")], True),
+        ([SimpleNamespace(metadata={"Name": "unrelated-package"}, version="2.8.0")], False),
+    ],
+)
+@pytest.mark.agent_authored(model="gpt-5.6-sol")
+def test_cutensor_mg_expected_failure_follows_installed_cutensor_version(
+    mocker,
+    installed_distributions,
+    expected,
+):
+    mocker.patch("local_helpers.importlib.metadata.distributions", return_value=installed_distributions)
+    have_distribution.cache_clear()
+    try:
+        assert _is_expected_load_nvidia_dynamic_lib_failure("cutensorMg") is expected
+    finally:
+        have_distribution.cache_clear()
+
+
 @pytest.mark.parametrize(
     "libname",
     supported_nvidia_libs.SUPPORTED_WINDOWS_DLLS if IS_WINDOWS else supported_nvidia_libs.SUPPORTED_LINUX_SONAMES,

cuda_pathfinder/tests/test_load_nvidia_dynamic_lib.py

@@ -3,6 +3,10 @@
         "version": "latest",
         "url": "https://nvidia.github.io/cuda-python/latest/"
     },
+    {
+        "version": "13.4.2",
+        "url": "https://nvidia.github.io/cuda-python/13.4.2/"
+    },
     {
         "version": "13.4.1",
         "url": "https://nvidia.github.io/cuda-python/13.4.1/"