ci: fix cu129 wheel tagging + pipefail-abort in install script (follow-up to #23497) (#23587)

scripts/ci/cuda/ci_install_dependency.sh

@@ -315,20 +315,14 @@ if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ -d "sgl-kernel/dist" ]; then
         WHEEL_ARCH="x86_64"
     fi
     # Wheel filenames carry a +cuXYZ local version tag (e.g. sglang_kernel-0.4.0+cu130-...).
-    # When the build matrix produces wheels for multiple CUDA versions (e.g. both cu129 and cu130),
-    # pick the one matching the test runner's $CU_VERSION so we don't trigger the PyPI-fallback reinstall
-    # below (which would replace the PR-built wheel with a public main-branch wheel).
-    KERNEL_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}+${CU_VERSION}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1)
-    if [ -z "$KERNEL_WHL" ]; then
-        # Fallback for older branches that only build a single CUDA variant and whose wheel matches
-        # the current $CU_VERSION. Restrict to the same arch — we intentionally do NOT match a wheel
-        # whose +cuXYZ tag disagrees with $CU_VERSION here, because the post-install reinstall branch
-        # below would otherwise silently replace it with the public main-branch wheel.
-        SINGLE_CUDA_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1)
-        if [ -n "$SINGLE_CUDA_WHL" ]; then
-            KERNEL_WHL="$SINGLE_CUDA_WHL"
-        fi
-    fi
+    # Pick the one matching the test runner's $CU_VERSION; any other tag (or an untagged
+    # wheel of unknown CUDA origin) would at best hit the PyPI-reinstall branch below and
+    # silently replace the PR-built wheel with the public main-branch wheel, and at worst
+    # ABI-crash at cuInit. Better to fail loudly — the error message + dist/ listing below
+    # makes the diagnosis obvious.
+    # `|| true` swallows `ls`'s exit-2-on-no-match so `set -o pipefail` doesn't abort the
+    # script before we reach the explicit error check.
+    KERNEL_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}+${CU_VERSION}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1 || true)
     if [ -z "$KERNEL_WHL" ]; then
         echo "ERROR: No matching sgl-kernel wheel found in sgl-kernel/dist/ for version ${SGL_KERNEL_VERSION_FROM_KERNEL} arch ${WHEEL_ARCH} cuda ${CU_VERSION}"
         ls -alh sgl-kernel/dist/

sgl-kernel/rename_wheels.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
-# Align CUDA wheel filenames (+cu124/+cu128/+cu130) with internal METADATA Version and
+# Align CUDA wheel filenames (+cu124/+cu128/+cu129/+cu130) with internal METADATA Version and
 # WHEEL tags after build (fixes pip "inconsistent version" when only the .whl name changed).
 # Unpack → patch WHEEL/METADATA → wheel pack (RECORD regenerated; no hand-editing).
 set -ex
@@ -11,6 +11,8 @@ detect_cuda_suffix() {
         echo "+cu124"
     elif ls /usr/local/ 2>/dev/null | grep -q "12.8"; then
         echo "+cu128"
+    elif ls /usr/local/ 2>/dev/null | grep -q "12.9"; then
+        echo "+cu129"
     elif ls /usr/local/ 2>/dev/null | grep -q "13.0"; then
         echo "+cu130"
     else