Add local cuDSS spack recipe with CUDA 13 redistributables

The upstream spack cudss recipe lists only the CUDA 12 redistributables while
declaring depends_on("cuda@12:"), so on a CUDA 13 toolchain spack selects the
CUDA 12 binary, which fails to load libcudart.so.12 at runtime. Override cudss
in the local repo with both the CUDA 12 and CUDA 13 redistributables, each bound
to its matching CUDA major (cuda@12.0:12 / cuda@13.0:13), so the concretizer
picks the build matching the CUDA in the DAG. Remove once upstream ships the
CUDA 13 archives.

Author: hughcars

spack_repo/local/packages/cudss/package.py

@@ -0,0 +1,60 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+# SPDX-License-Identifier: Apache-2.0
+
+# Local override of the builtin spack `cudss` package. Upstream only lists the
+# CUDA 12 redistributables together with `depends_on("cuda@12:")`, so on a
+# CUDA 13 toolchain spack hands out the CUDA 12 binary, which then fails to load
+# `libcudart.so.12` at runtime. Here each cuDSS build is bound to its matching
+# CUDA major version, so the concretizer selects the redistributable that
+# matches the CUDA in the DAG (CUDA 12 -> cuda12 build, CUDA 13 -> cuda13 build).
+# Remove once the upstream spack recipe ships the CUDA 13 redistributables.
+
+import platform
+
+from spack_repo.builtin.build_systems.generic import Package
+
+from spack.package import *
+
+_url = (
+    "https://developer.download.nvidia.com/compute/cudss/redist/libcudss/"
+    "linux-{arch}/libcudss-linux-{arch}-{ver}_cuda{cuda}-archive.tar.xz"
+)
+
+
+class Cudss(Package):
+    """NVIDIA cuDSS is a GPU-accelerated Direct Sparse Solver library for
+    solving linear systems with very sparse matrices. This local override adds
+    the CUDA 13 redistributables alongside the upstream CUDA 12 ones."""
+
+    homepage = "https://developer.nvidia.com/cudss"
+
+    if platform.machine() == "x86_64":
+        # CUDA 12 build (default; matches the upstream recipe).
+        version(
+            "0.7.1",
+            sha256="946571d9ea164f948e402dd97a14541cb90fbec800336cfa7ae644af5937632f",
+            url=_url.format(arch="x86_64", ver="0.7.1.4", cuda="12"),
+            preferred=True,
+        )
+        # CUDA 13 build.
+        version(
+            "0.7.1-cuda13",
+            sha256="84b34ebe7fad40ec10f2aab2957a63b6070bd8ce16e3ada3e6bcac7317256347",
+            url=_url.format(arch="x86_64", ver="0.7.1.4", cuda="13"),
+        )
+    elif platform.machine() == "aarch64":
+        # CUDA 12 build (upstream only ships aarch64 for CUDA 12 so far).
+        version(
+            "0.7.1",
+            sha256="f283c31b6badf4a5277014295705eac8e6e28f27de30d746719ea1cef7a750b8",
+            url=_url.format(arch="aarch64", ver="0.7.1.4", cuda="12"),
+            preferred=True,
+        )
+
+    # Bind each redistributable to its matching CUDA major so the concretizer
+    # never pairs a cuda12 binary with a CUDA 13 toolchain (or vice versa).
+    depends_on("cuda@12.0:12", when="@0.7.1")
+    depends_on("cuda@13.0:13", when="@0.7.1-cuda13")
+
+    def install(self, spec, prefix):
+        install_tree(".", prefix)