Add vLLM CT script and GPU passthrough support

Introduce full vLLM container integration with new `ct/vllm.sh` and `install/vllm-install.sh` scripts, including uv/Python setup, hardware acceleration setup, systemd service creation, and update flow. Add `json/vllm.json` metadata with defaults, access details, and operational notes, and register `vllm` in `misc/passthrough.func` so GPU passthrough selection supports the app.

Author: MickLesk

ct/vllm.sh

@@ -0,0 +1,62 @@
+#!/usr/bin/env bash
+source "$(dirname "${BASH_SOURCE[0]}")/../misc/build.func" 2>/dev/null || source <(curl -fsSL "${COMMUNITY_SCRIPTS_URL:-https://raw.githubusercontent.com/community-scripts/ProxmoxVED/main}/misc/build.func")
+# Copyright (c) 2021-2026 community-scripts ORG
+# Author: community-scripts ORG
+# License: MIT | https://github.com/community-scripts/ProxmoxVED/raw/main/LICENSE
+# Source: https://github.com/vllm-project/vllm
+
+APP="vLLM"
+var_tags="${var_tags:-ai;llm;inference}"
+var_cpu="${var_cpu:-4}"
+var_ram="${var_ram:-8192}"
+var_disk="${var_disk:-20}"
+var_os="${var_os:-debian}"
+var_version="${var_version:-13}"
+var_arm64="${var_arm64:-no}"
+var_unprivileged="${var_unprivileged:-1}"
+
+header_info "$APP"
+variables
+color
+catch_errors
+
+function update_script() {
+  header_info
+  check_container_storage
+  check_container_resources
+
+  if [[ ! -d /opt/vllm ]]; then
+    msg_error "No ${APP} Installation Found!"
+    exit 1
+  fi
+
+  if check_for_gh_release "vllm" "vllm-project/vllm"; then
+    msg_info "Stopping Service"
+    systemctl stop vllm
+    msg_ok "Stopped Service"
+
+    create_backup /opt/vllm/vllm.env
+
+    msg_info "Updating vLLM (Patience)"
+    $STD uv pip install --python /opt/vllm/bin/python --upgrade vllm
+    msg_ok "Updated vLLM"
+
+    restore_backup
+
+    msg_info "Starting Service"
+    systemctl start vllm
+    msg_ok "Started Service"
+    msg_ok "Updated successfully!"
+  fi
+  cleanup_lxc
+  exit
+}
+
+start
+build_container
+description
+
+msg_ok "Completed Successfully!\n"
+echo -e "${CREATING}${GN}${APP} setup has been successfully initialized!${CL}"
+echo -e "${INFO}${YW}Access it using the following URL:${CL}"
+echo -e "${TAB}${GATEWAY}${BGN}http://${IP}:8000/v1${CL}"

install/vllm-install.sh

@@ -0,0 +1,68 @@
+#!/usr/bin/env bash
+
+# Copyright (c) 2021-2026 community-scripts ORG
+# Author: community-scripts ORG
+# License: MIT | https://github.com/community-scripts/ProxmoxVED/raw/main/LICENSE
+# Source: https://github.com/vllm-project/vllm
+
+source /dev/stdin <<<"$FUNCTIONS_FILE_PATH"
+color
+verb_ip6
+catch_errors
+setting_up_container
+network_check
+update_os
+
+msg_info "Installing Dependencies"
+$STD apt install -y \
+  build-essential \
+  python3-dev
+msg_ok "Installed Dependencies"
+
+PYTHON_VERSION="3.12" setup_uv
+setup_hwaccel
+
+msg_info "Installing vLLM (Patience)"
+$STD uv venv --python 3.12 /opt/vllm
+$STD uv pip install --python /opt/vllm/bin/python vllm
+cat <<EOF >~/.vllm
+$(get_latest_github_release "vllm-project/vllm")
+EOF
+msg_ok "Installed vLLM"
+
+msg_info "Configuring vLLM"
+mkdir -p /opt/vllm/models
+cat <<EOF >/opt/vllm/vllm.env
+VLLM_MODEL=Qwen/Qwen2.5-0.5B-Instruct
+VLLM_HOST=0.0.0.0
+VLLM_PORT=8000
+HF_HOME=/opt/vllm/models
+HF_TOKEN=
+EOF
+msg_ok "Configured vLLM"
+
+msg_info "Creating Service"
+cat <<EOF >/etc/systemd/system/vllm.service
+[Unit]
+Description=vLLM OpenAI-Compatible Inference Server
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+WorkingDirectory=/opt/vllm
+EnvironmentFile=/opt/vllm/vllm.env
+ExecStart=/opt/vllm/bin/vllm serve \${VLLM_MODEL} --host \${VLLM_HOST} --port \${VLLM_PORT}
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=0
+
+[Install]
+WantedBy=multi-user.target
+EOF
+systemctl enable -q --now vllm
+msg_ok "Created Service"
+
+motd_ssh
+customize
+cleanup_lxc

json/vllm.json

@@ -0,0 +1,50 @@
+{
+  "name": "vLLM",
+  "slug": "vllm",
+  "categories": [
+    20
+  ],
+  "date_created": "2026-07-24",
+  "type": "ct",
+  "updateable": true,
+  "privileged": false,
+  "has_arm": false,
+  "interface_port": 8000,
+  "documentation": "https://docs.vllm.ai/",
+  "website": "https://vllm.ai/",
+  "logo": "https://cdn.jsdelivr.net/gh/selfhst/icons@main/webp/vllm.webp",
+  "config_path": "/opt/vllm/vllm.env",
+  "description": "vLLM is a high-throughput, memory-efficient inference and serving engine for LLMs. It exposes an OpenAI-compatible HTTP API (/v1) so existing OpenAI clients can serve local models.",
+  "install_methods": [
+    {
+      "type": "default",
+      "script": "ct/vllm.sh",
+      "config_path": "/opt/vllm/vllm.env",
+      "resources": {
+        "cpu": 4,
+        "ram": 8192,
+        "hdd": 20,
+        "os": "Debian",
+        "version": "13"
+      }
+    }
+  ],
+  "default_credentials": {
+    "username": null,
+    "password": null
+  },
+  "notes": [
+    {
+      "text": "Select NVIDIA GPU passthrough at container creation. vLLM installs the CUDA build by default and requires an NVIDIA GPU for usable performance; CPU-only inference is impractically slow.",
+      "type": "warning"
+    },
+    {
+      "text": "The model to serve is set via VLLM_MODEL in /opt/vllm/vllm.env (default Qwen/Qwen2.5-0.5B-Instruct). Restart the service after changing it: systemctl restart vllm",
+      "type": "info"
+    },
+    {
+      "text": "The configured model is downloaded from HuggingFace on first service start, which needs internet access and disk space. For gated/private models set HF_TOKEN in /opt/vllm/vllm.env. Resource needs scale heavily with the chosen model.",
+      "type": "info"
+    }
+  ]
+}

misc/passthrough.func

@@ -77,6 +77,7 @@ _GPU_APPS=(
   fileflows
   "open webui"
   tunarr
+  vllm
   debian
 )