LocalAI: pin the CPU backends when the container has no GPU runtime

LocalAI resolves meta backends from the GPU vendor it reads out of sysfs,
which is the host's card even with no driver installed in the container,
so llama-cpp became cuda12-llama-cpp and never started (#17908). Force
the default capability unless a CUDA, ROCm or Level Zero runtime is
present; update adds the pin to existing installs.

ct/localai.sh

@@ -37,6 +37,15 @@ function update_script() {
     exit
   fi
 
+  # Older installs have no CPU pin; on a host with a GPU but no driver in the
+  # container every model then ends in "backend not found".
+  if ! grep -q '^LOCALAI_FORCE_META_BACKEND_CAPABILITY=' /opt/localai.env 2>/dev/null &&
+    ! ldconfig -p 2>/dev/null | grep -qE 'libcuda[.]so|libamdhip64[.]so|libze_intel_gpu[.]so|libze_loader[.]so'; then
+    echo "LOCALAI_FORCE_META_BACKEND_CAPABILITY=default" >>/opt/localai.env
+    systemctl restart localai
+    msg_ok "Pinned LocalAI to the CPU backends (no GPU runtime in the container)"
+  fi
+
   if check_for_gh_release "localai" "mudler/LocalAI"; then
     msg_info "Stopping Service"
     systemctl stop localai

install/localai-install.sh

@@ -39,6 +39,12 @@ LOCALAI_MODELS_PATH=/opt/localai_data/models
 LOCALAI_BACKENDS_PATH=/opt/localai_data/backends
 LOCALAI_LOG_LEVEL=info
 EOF
+# LocalAI picks the backend variant from the GPU it sees in sysfs, which is the
+# host's card even when no driver was installed in here, and a CUDA/ROCm/SYCL
+# backend then never starts. Pin the CPU variants unless a runtime is present.
+if ! ldconfig -p 2>/dev/null | grep -qE 'libcuda[.]so|libamdhip64[.]so|libze_intel_gpu[.]so|libze_loader[.]so'; then
+  echo "LOCALAI_FORCE_META_BACKEND_CAPABILITY=default" >>/opt/localai.env
+fi
 # LocalAI refuses to start on a wildcard bind with nothing to identify callers,
 # and binding the wildcard is what makes the container reachable at all.
 if [[ "$var_auth" == "yes" ]]; then