Fix CUDA 13.4 arch list: nvcc fatal 'Unsupported gpu architecture compute_50'

get_cuda_arch_list() matched "13.0" and "13.2" exactly, so 13.4 fell through to
the legacy fallback list "5.0;6.0;7.0;7.5;8.0;8.6;9.0". CUDA 13 removed sm_50
and sm_60, so every CUDA translation unit failed:

    /usr/local/cuda-13.4/bin/nvcc ...
    nvcc fatal : Unsupported gpu architecture 'compute_50'
    subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1
    RuntimeError: Error compiling objects for extension

seen on pytorch/vision build-wheel-py3_10-cuda-aarch6413_4-aarch64.

Match the whole 13.x series with startswith("13.") rather than adding 13.4 to
the exact-match chain, so the next CUDA minor does not silently fall through to
a list containing architectures the toolkit no longer supports.

Add direct coverage for the arch list on 13.0/13.2/13.4 x linux/linux-aarch64:
the existing tests compare get_cuda_variables() against get_cuda_arch_list()
output, so they cannot catch a wrong list.

Author: atalman

tools/pkg-helpers/pytorch_pkg_helpers/cuda.py

@@ -16,7 +16,10 @@ def get_cuda_arch_list(
     # however we would like to keep sm_70 architecture see: https://github.com/pytorch/pytorch/issues/157517
     if sanitized_version == "12.8":
         return "7.0;7.5;8.0;8.6;9.0;10.0;12.0"
-    elif sanitized_version == "13.0" or sanitized_version == "13.2":
+    elif sanitized_version.startswith("13."):
+        # CUDA 13 removed sm_50/sm_60 outright, so the fallback list above makes
+        # nvcc fail with "Unsupported gpu architecture 'compute_50'". Match the
+        # whole 13.x series so a new minor does not silently fall through.
         arch_list = "7.5;8.0;8.6;9.0;10.0;12.0+PTX"
         # Add sm_110 for aarch64
         if "aarch64" in platform:

tools/pkg-helpers/tests/test_cuda.py

@@ -72,3 +72,25 @@ def test_cuda_variables_cuda_windows_wheels(gpu_arch_version):
         f'export PATH="{cuda_home}/bin:${{PATH}}"',
         "export FORCE_CUDA=1",
     ]
+
+
+@pytest.mark.parametrize("sanitized_version", ["13.0", "13.2", "13.4"])
+@pytest.mark.parametrize("platform", ["linux", "linux-aarch64"])
+def test_cuda_arch_list_cuda13_excludes_removed_arches(sanitized_version, platform):
+    # CUDA 13 removed sm_50/sm_60; requesting them makes nvcc fail with
+    # "Unsupported gpu architecture 'compute_50'".
+    arch_list = get_cuda_arch_list(sanitized_version, platform=platform)
+    assert "5.0" not in arch_list
+    assert "6.0" not in arch_list
+
+
+@pytest.mark.parametrize("sanitized_version", ["13.0", "13.2", "13.4"])
+def test_cuda_arch_list_cuda13(sanitized_version):
+    assert (
+        get_cuda_arch_list(sanitized_version, platform="linux")
+        == "7.5;8.0;8.6;9.0;10.0;12.0+PTX"
+    )
+    assert (
+        get_cuda_arch_list(sanitized_version, platform="linux-aarch64")
+        == "8.0;9.0;10.0;11.0;12.0+PTX"
+    )