Add NVIDIA stress loader selection and DCGM 4 support

iso/builder/build-nvidia-module.sh

@@ -10,7 +10,7 @@
 # Output layout:
 #   $CACHE_DIR/modules/   — nvidia*.ko files
 #   $CACHE_DIR/bin/       — nvidia-smi, nvidia-debugdump
-#   $CACHE_DIR/lib/       — libnvidia-ml.so*, libcuda.so* (for nvidia-smi)
+#   $CACHE_DIR/lib/       — libnvidia-ml.so*, libcuda.so*, OpenCL-related libs
 
 set -e
 
@@ -133,7 +133,14 @@ fi
 # Copy ALL userspace library files.
 # libnvidia-ptxjitcompiler is required by libcuda for PTX JIT compilation
 # (cuModuleLoadDataEx with PTX source) — without it CUDA_ERROR_JIT_COMPILER_NOT_FOUND.
-for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
+for lib in \
+    libnvidia-ml \
+    libcuda \
+    libnvidia-ptxjitcompiler \
+    libnvidia-opencl \
+    libnvidia-compiler \
+    libnvidia-nvvm \
+    libnvidia-fatbinaryloader; do
     count=0
     for f in $(find "$EXTRACT_DIR" -maxdepth 1 -name "${lib}.so.*" 2>/dev/null); do
         cp "$f" "$CACHE_DIR/lib/" && count=$((count+1))
@@ -150,7 +157,14 @@ ko_count=$(ls "$CACHE_DIR/modules/"*.ko 2>/dev/null | wc -l)
 [ "$ko_count" -gt 0 ] || { echo "ERROR: no .ko files built in $CACHE_DIR/modules/"; exit 1; }
 
 # Create soname symlinks: use [0-9][0-9]* to avoid circular symlink (.so.1 has single digit)
-for lib in libnvidia-ml libcuda libnvidia-ptxjitcompiler; do
+for lib in \
+    libnvidia-ml \
+    libcuda \
+    libnvidia-ptxjitcompiler \
+    libnvidia-opencl \
+    libnvidia-compiler \
+    libnvidia-nvvm \
+    libnvidia-fatbinaryloader; do
     versioned=$(ls "$CACHE_DIR/lib/${lib}.so."[0-9][0-9]* 2>/dev/null | head -1)
     [ -n "$versioned" ] || continue
     base=$(basename "$versioned")