> ## Documentation Index > Fetch the complete documentation index at: https://docs.poolside.ai/llms.txt > Use this file to discover all available pages before exploring further. # On-premises installation prerequisites > Requirements to complete before you install Poolside model inference on an on-premises GPU host. Before you [install Poolside on-premises](/deployment/on-prem/install), complete the following steps in order. Review the STIG guidance in Step 1 only if it applies to your host. ## Step 1: Verify host requirements Ensure that the host meets the following requirements: * A supported operating system: * Ubuntu 22.04 LTS or 24.04 LTS * Red Hat Enterprise Linux (RHEL) 9.6 * SUSE Linux Enterprise Server (SLES) 16 * A supported host CPU architecture: * `amd64` (`x86_64`) * `arm64` (`aarch64`) * Storage capacity and partitions planned before you install the operating system, because partitions are difficult to change afterward. See [Storage requirements](/deployment/on-prem/storage) for current guidance. * `sudo` access on the host Some prerequisite tools provide architecture-specific Linux binaries. Use `amd64` downloads for `x86_64` hosts and `arm64` downloads for `aarch64` hosts. To check the host architecture, run: ```bash theme={null} uname -m ``` ### STIG-hardened hosts If you deploy on a STIG-hardened host, two decisions happen before or during operating system installation: FIPS-validated cryptography and the filesystem partition scheme. For both, see [STIG hardening considerations](/deployment/on-prem/stig). ## Step 2: Gather installation materials Ensure that the following materials are available on the host: * The Poolside installation bundle for your host CPU architecture * Poolside model checkpoint files for your deployment Model checkpoint files are provided separately based on your deployment. You upload them during the model upload step of the installation. ## Step 3: Choose a network mode The network mode changes how the host obtains NVIDIA GPU drivers and Terraform providers. | Network mode | Description | | ------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | Internet-connected | The host can reach external package repositories and the Terraform registry. | | Air-gapped | The host has no external network access. The bundle supplies Terraform providers from a local cache. Complete the air-gapped GPU driver preparation for your operating system. | ## Step 4: Choose certificates and hostnames Decide the following values before you install: * The ingress hostnames you plan to expose for model inference * Whether you use installer-generated self-signed certificates or your own TLS certificates If you use custom TLS certificates, prepare CA and server certificate files whose Subject Alternative Names (SANs) cover every hostname you expose, including each model ingress hostname, `seaweedfs.poolside.local`, and `seaweedfs-s3.poolside.local`. ## Step 5: Choose a GPU driver path The NVIDIA GPU Operator checks whether a driver is already installed on the host and prefers a preinstalled driver. Find your operating system, then start at the top of its table and use the first row that applies. ### Ubuntu | Host conditions | Driver installation path | | ---------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | The host requires Secure Boot | Preinstall the validated NVIDIA driver because the driver compiled by the GPU Operator is not compatible with Secure Boot. See [Preinstall the NVIDIA driver on Ubuntu](/deployment/on-prem/prerequisites/ubuntu#preinstall-the-nvidia-driver). | | The host can reach external Ubuntu package repositories and Secure Boot is disabled | Start with no host NVIDIA driver. The GPU Operator installs the driver by using the Ubuntu package repositories. See [Use external Ubuntu package repositories](/deployment/on-prem/prerequisites/ubuntu#use-external-ubuntu-package-repositories). | | The host cannot reach external repositories, can reach an internal APT mirror, and Secure Boot is disabled | Configure the GPU Operator driver container to use the mirror. See [Use an internal APT mirror](/deployment/on-prem/prerequisites/ubuntu#use-an-internal-apt-mirror). | | The host is air-gapped without an accessible mirror, Secure Boot is disabled, and the APT cache is complete for the running kernel | The GPU Operator can install the driver by using the cached packages. See [Use the host APT cache](/deployment/on-prem/prerequisites/ubuntu#use-the-host-apt-cache). | | None of the preceding conditions apply | Preinstall the validated NVIDIA driver. This is the simplest path for a disconnected installation. See [Preinstall the NVIDIA driver on Ubuntu](/deployment/on-prem/prerequisites/ubuntu#preinstall-the-nvidia-driver). | ### RHEL | Host conditions | Driver installation path | | ---------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | The host requires Secure Boot | Preinstall the validated NVIDIA driver because the driver compiled by the GPU Operator is not compatible with Secure Boot. See [Preinstall the NVIDIA driver on RHEL](/deployment/on-prem/prerequisites/rhel#preinstall-the-nvidia-driver). | | The host is air-gapped and Secure Boot is disabled | Preinstall the validated NVIDIA driver. This is the simplest path for a disconnected installation. See [Preinstall the NVIDIA driver on RHEL](/deployment/on-prem/prerequisites/rhel#preinstall-the-nvidia-driver). | | The host is internet-connected and Secure Boot is disabled | Start with no host NVIDIA driver. The GPU Operator installs the driver by using the RHEL package repositories. See [Let the GPU Operator install the driver](/deployment/on-prem/prerequisites/rhel#let-the-gpu-operator-install-the-driver). | ### SLES Preinstall the NVIDIA driver in either network mode. The GPU Operator does not install the driver on SLES. See [Install the NVIDIA driver on SLES](/deployment/on-prem/prerequisites/sles#step-4-install-the-nvidia-driver). ## Step 6: Prepare your operating system Complete the prerequisites page for your operating system: * [Prepare Ubuntu](/deployment/on-prem/prerequisites/ubuntu) * [Prepare RHEL](/deployment/on-prem/prerequisites/rhel) * [Prepare SLES](/deployment/on-prem/prerequisites/sles) ## Next steps * [Install on-premises](/deployment/on-prem/install) ## Related resources * [Install on-premises](/deployment/on-prem/install) * [On-premises deployment](/deployment/on-prem/overview) * [Storage requirements](/deployment/on-prem/storage) * [Certified stacks](/deployment/on-prem/certified-stacks/overview) * [Supported configurations](/deployment/supported-configurations)