v2: stage CUDA runtime libs into bin/ (fix silent CPU fallback)

build.sh

@@ -40,6 +40,21 @@ if ! cmake -B build "${CMAKE_FLAGS[@]}"; then
     -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=120 -DGGML_NATIVE=OFF
 fi
 cmake --build build -j"$(nproc)" --target llama-server
+
+# Stage CUDA runtime libs next to the binaries: the ZeroGPU image has no CUDA toolkit,
+# and libggml-cuda.so silently fails to dlopen without them (CPU fallback).
+echo "=== ldd before staging ==="
+ldd build/bin/llama-server || true
+ldd build/bin/libggml-cuda.so || true
+for lib in $(ldd build/bin/*.so build/bin/llama-server 2>/dev/null | grep -oP '/usr/local/cuda/lib64/\S+' | sort -u); do
+  base=$(basename "$lib")
+  case "$base" in
+    libcuda.so*) echo "skip driver lib $base (provided by the GPU host)";;
+    *) cp -L "$lib" build/bin/ && echo "staged $base";;
+  esac
+done
+echo "=== ldd after staging (LD_LIBRARY_PATH=build/bin) ==="
+LD_LIBRARY_PATH=build/bin ldd build/bin/libggml-cuda.so | grep -E "not found|cuda" || true
 ls -la build/bin
 
 tar -C build -czf /tmp/llama-ricursive-cuda128-sm120.tar.gz bin