pathfinder: also find ptxas in the nvidia/cu13/bin wheel layout (#2741)

* fix(cuda.pathfinder): discover ptxas from NVIDIA wheels

* test(cuda.pathfinder): add pytest-run-parallel dependency

* docs(pathfinder): add 1.8.1 release notes

Author: rwgk

cuda_pathfinder/cuda/pathfinder/_binaries/supported_nvidia_binaries.py

@@ -13,6 +13,7 @@
 SITE_PACKAGES_BINDIRS = {
     # Core compilation tools
     "nvcc": (_CUDA13_BIN, _CUDA_NVCC_BIN),
+    "ptxas": (_CUDA13_BIN, _CUDA_NVCC_BIN),
     "nvdisasm": (_CUDA13_BIN, _CUDA_NVCC_BIN),
     "cuobjdump": (_CUDA13_BIN, _CUDA_NVCC_BIN),
     "nvprune": (_CUDA_NVCC_BIN,),

cuda_pathfinder/docs/nv-versions.json

@@ -3,6 +3,10 @@
         "version": "latest",
         "url": "https://nvidia.github.io/cuda-python/cuda-pathfinder/latest/"
     },
+    {
+        "version": "1.8.1",
+        "url": "https://nvidia.github.io/cuda-python/cuda-pathfinder/1.8.1/"
+    },
     {
         "version": "1.8.0",
         "url": "https://nvidia.github.io/cuda-python/cuda-pathfinder/1.8.0/"

cuda_pathfinder/docs/source/release/1.8.1-notes.rst

@@ -0,0 +1,33 @@
+.. SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+.. SPDX-License-Identifier: Apache-2.0
+
+.. py:currentmodule:: cuda.pathfinder
+
+``cuda-pathfinder`` 1.8.1 Release notes
+=======================================
+
+Highlights
+----------
+
+* Add ``ptxas`` to ``SUPPORTED_BINARY_UTILITIES``. The
+  :func:`find_nvidia_binary_utility` function can now locate it on Linux and
+  Windows through the existing NVIDIA wheel, Conda, and CUDA Toolkit search
+  pipeline. NVIDIA wheel discovery prefers the CUDA 13 ``nvidia/cu13/bin``
+  layout while retaining the legacy ``nvidia/cuda_nvcc/bin`` fallback.
+  (`PR #2741 <https://github.com/NVIDIA/cuda-python/pull/2741>`_)
+
+Bugfixes
+--------
+
+* Find ``cuobjdump`` in the CUDA 13 ``nvidia/cu13/bin`` NVIDIA wheel layout
+  while retaining discovery from the legacy ``nvidia/cuda_nvcc/bin`` layout.
+  (`PR #2739 <https://github.com/NVIDIA/cuda-python/pull/2739>`_)
+
+Internal maintenance
+--------------------
+
+* Add ``pytest-run-parallel>=0.4.1`` to the normal test dependencies so the
+  existing ``thread_unsafe`` marker and ``thread_unsafe_fixtures``
+  configuration option are registered in standard test environments. This is
+  test-only and does not add a runtime dependency.
+  (`PR #2741 <https://github.com/NVIDIA/cuda-python/pull/2741>`_)

cuda_pathfinder/pixi.lock

@@ -15,6 +15,7 @@ pytest = ">=6.2.4"
 pytest-mock = "*"
 pytest-repeat = "*"
 pytest-randomly = "*"
+pytest-run-parallel = ">=0.4.1"
 
 # Keep this dependency set aligned with cuda_python/docs/environment-docs.yml.
 [feature.docs.dependencies]

cuda_pathfinder/pixi.toml

@@ -16,6 +16,7 @@ test = [
     "pytest-mock==3.15.1",
     "pytest-repeat==0.9.4",
     "pytest-randomly==4.1.0",
+    "pytest-run-parallel>=0.4.1",
 ]
 # Internal organization of test dependencies.
 cu12 = [

cuda_pathfinder/pyproject.toml

@@ -16,7 +16,7 @@
 )
 
 CUDA13_WHEEL_BINARIES = frozenset(
-    ("nvcc", "nvdisasm", "cuobjdump", "fatbinary", "bin2c", "nvlink", "compute-sanitizer")
+    ("nvcc", "ptxas", "nvdisasm", "cuobjdump", "fatbinary", "bin2c", "nvlink", "compute-sanitizer")
 )
 
 
@@ -40,7 +40,7 @@ def test_supported_binaries_consistency():
 
 @pytest.mark.agent_authored(model="claude-opus-5")
 def test_cuda13_wheel_layout_is_declared_and_preferred():
-    """Every utility shipped in nvidia/cu13/bin must declare that layout first.
+    """Every supported utility shipped in nvidia/cu13/bin must declare that layout first.
 
     Omitting it makes the site-packages search step skip an installed CUDA 13
     wheel entirely, which is only visible on hosts without a local CTK to fall