chore(gb200): update to CUDA 12.9 and improve build process (#8772)

sgl-kernel/build.sh

@@ -39,6 +39,13 @@ docker run --rm \
    # Install CMake (version >= 3.26) - Robust Installation
    export CMAKE_VERSION_MAJOR=3.31
    export CMAKE_VERSION_MINOR=1
+   # Setting these flags to reduce OOM chance only on ARM
+   if [ \"${ARCH}\" = \"aarch64\" ]; then
+      export CUDA_NVCC_FLAGS=\"-Xcudafe --threads=2\"
+      export MAKEFLAGS='-j2'
+      export CMAKE_BUILD_PARALLEL_LEVEL=2
+      export NINJAFLAGS='-j2'
+   fi
    echo \"Downloading CMake from: https://cmake.org/files/v\${CMAKE_VERSION_MAJOR}/cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-${ARCH}.tar.gz\"
    wget https://cmake.org/files/v\${CMAKE_VERSION_MAJOR}/cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-${ARCH}.tar.gz
    tar -xzf cmake-\${CMAKE_VERSION_MAJOR}.\${CMAKE_VERSION_MINOR}-linux-${ARCH}.tar.gz

sgl-kernel/rename_wheels.sh

@@ -7,8 +7,19 @@ wheel_files=($WHEEL_DIR/*.whl)
 for wheel in "${wheel_files[@]}"; do
     intermediate_wheel="${wheel/linux/manylinux2014}"
 
-    if ls /usr/local/ | grep -q "12.8"; then
-        new_wheel="${intermediate_wheel/-cp39/+cu128-cp39}"
+    # Extract the current python version from the wheel name
+    if [[ $intermediate_wheel =~ -cp([0-9]+)- ]]; then
+        cp_version="${BASH_REMATCH[1]}"
+    else
+        echo "Could not extract Python version from wheel name: $intermediate_wheel"
+        continue
+    fi
+
+    # Detect CUDA version and add appropriate suffix
+    if ls /usr/local/ | grep -q "12.9"; then
+        new_wheel="${intermediate_wheel/-cp${cp_version}/+cu129-cp${cp_version}}"
+    elif ls /usr/local/ | grep -q "12.8"; then
+        new_wheel="${intermediate_wheel/-cp${cp_version}/+cu128-cp${cp_version}}"
     else
         new_wheel="$intermediate_wheel"
     fi