Linux ARM64v1.3.2Installation

Install Vault on DGX Spark or ASUS Ascent GX10

Install Wave2 Vault on NVIDIA DGX Spark or ASUS Ascent GX10 (Linux ARM64) — from USB preparation through licence activation and first start.

Use this guide for a new Vault installation on:

NVIDIA DGX Spark or ASUS Ascent GX10
NVIDIA DGX OS based on Ubuntu 24.04
An ARM64 processor with an NVIDIA GB10 GPU
128 GB of unified system memory

Complete every step using the same Ubuntu user. Do not run the vault launcher with sudo.

If Vault is already installed on this machine, use the upgrade guide instead.

Step 0 — Prepare the installation USB

Wave2 will provide vault-1.3.2-linux-arm64.zip. The installer preparing the USB must also download these models from Hugging Face:

Chat/Wiki LLM Model: Unsloth Qwen3.6 35B A3B GGUF (opens in a new tab)
Embedding Model: Qwen3 Embedding 0.6B GGUF (opens in a new tab)
Embedding tokenizer: Qwen3 Embedding 0.6B (opens in a new tab)
Reranker: BGE Reranker v2 M3 (opens in a new tab)

Do the following on an internet-connected preparation computer. Install the official Hugging Face hf command using one of these methods:

bash

# macOS with Homebrew
brew install hf

# Or macOS/Linux with Python and pip
python3 -m pip install --upgrade huggingface_hub

Create the installation folder:

bash

INSTALL_ROOT="$HOME/Vault-Install"

mkdir -p "$INSTALL_ROOT/models/gguf/qwen3-embedding-tokenizer"
mkdir -p "$INSTALL_ROOT/models/reranker/bge-reranker-v2-m3"

Download the exact model files used by Vault:

bash

INSTALL_ROOT="$HOME/Vault-Install"

hf download \
  unsloth/Qwen3.6-35B-A3B-GGUF \
  Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  --local-dir "$INSTALL_ROOT/models/gguf"

hf download \
  Qwen/Qwen3-Embedding-0.6B-GGUF \
  Qwen3-Embedding-0.6B-Q8_0.gguf \
  --local-dir "$INSTALL_ROOT/models/gguf"

hf download \
  Qwen/Qwen3-Embedding-0.6B \
  tokenizer.json tokenizer_config.json \
  --local-dir "$INSTALL_ROOT/models/gguf/qwen3-embedding-tokenizer"

hf download \
  BAAI/bge-reranker-v2-m3 \
  config.json model.safetensors sentencepiece.bpe.model \
  special_tokens_map.json tokenizer.json tokenizer_config.json \
  --local-dir "$INSTALL_ROOT/models/reranker/bge-reranker-v2-m3"

For Chat and Wiki, download only Qwen3.6-35B-A3B-UD-Q4_K_M.gguf. The other quantizations, imatrix, and vision-projector files are not required by this guide.

Extract vault-1.3.2-linux-arm64.zip and move the extracted vault-1.3.2-linux-arm64 folder into $HOME/Vault-Install.

Copy the complete Vault-Install folder to a USB drive named VAULTDRIVE. Use an exFAT-formatted USB drive with at least 40 GB of free space. The finished USB must have this structure:

text

VAULTDRIVE/
└── Vault-Install/
    ├── vault-1.3.2-linux-arm64/
    │   ├── vault
    │   ├── docker-images.tar
    │   ├── docker-compose.prod.yml
    │   ├── release-manifest.json
    │   └── runtimes/
    └── models/
        ├── gguf/
        │   ├── Qwen3.6-35B-A3B-UD-Q4_K_M.gguf
        │   ├── Qwen3-Embedding-0.6B-Q8_0.gguf
        │   └── qwen3-embedding-tokenizer/
        │       ├── tokenizer.json
        │       └── tokenizer_config.json
        └── reranker/
            └── bge-reranker-v2-m3/
                ├── config.json
                ├── model.safetensors
                ├── sentencepiece.bpe.model
                ├── special_tokens_map.json
                ├── tokenizer.json
                └── tokenizer_config.json

Keep all filenames and folder names unchanged.

Step 1 — Prepare DGX OS (internet access required)

Complete the machine's first-time setup before installing Vault. Use the NVIDIA DGX Dashboard to install all available system updates:

Open DGX Dashboard from the Ubuntu application menu.
Open Settings.
Select Update, then Update Now.
Wait for the update to finish and restart the machine if prompted.

This is the recommended update method for both NVIDIA DGX Spark (opens in a new tab) and ASUS Ascent GX10 (opens in a new tab).

DGX OS already includes the NVIDIA driver, Docker, CUDA, and the NVIDIA Container Runtime. Do not replace them with generic Ubuntu packages.

Open Terminal and run:

bash

sudo systemctl enable --now docker
sudo usermod -aG docker "$USER"

Log out of Ubuntu and log back in. Then run:

bash

uname -m
. /etc/os-release && printf '%s %s\n' "$ID" "$VERSION_ID"
nvidia-smi
docker info
docker compose version

uname -m must show aarch64. The operating system line must show ubuntu 24.04. The remaining commands must finish without an error.

On DGX Spark and GX10, nvidia-smi may show Memory-Usage: Not Supported. This is normal for the integrated GPU and unified memory architecture.

You may now disconnect the machine from the internet.

Step 2 — Copy Vault from the USB

Insert the installation USB. Ubuntu normally mounts it under /media/$USER/VAULTDRIVE.

Change VERSION below only if Wave2 supplied a different version:

bash

VERSION=1.3.2
USB_ROOT="/media/$USER/VAULTDRIVE/Vault-Install"

mkdir -p "$HOME/Vault"
cp -a "$USB_ROOT/vault-$VERSION-linux-arm64/." "$HOME/Vault/"

cd "$HOME/Vault"
chmod 755 ./vault

If your USB has a different name, replace VAULTDRIVE in USB_ROOT with the name shown by Ubuntu.

Step 3 — Install Vault

bash

cd "$HOME/Vault"
./vault install --bundle .

The installer creates Vault's data folders, loads the application, and creates a licence request for this machine. Vault remains read-only until its licence is activated.

Step 4 — Send the licence request to Wave2

Send this file to Wave2:

text

/home/<ubuntu-user>/.vault/license/vault-license-request.vlr

Replace <ubuntu-user> with your Ubuntu username. For example, user alex would find the file at:

text

/home/alex/.vault/license/vault-license-request.vlr

Wave2 will return a signed file named customer.vlic for this machine.

Step 5 — Activate Vault

Copy customer.vlic to the USB under Vault-Install, insert the USB into the Ubuntu machine, and run:

bash

USB_ROOT="/media/$USER/VAULTDRIVE/Vault-Install"

cd "$HOME/Vault"
./vault license activate --license "$USB_ROOT/customer.vlic"
./vault license status

The status must show Mode: active. If you saved customer.vlic somewhere else, replace the path after --license with its actual path.

Step 6 — Copy and configure the models

bash

USB_MODELS="/media/$USER/VAULTDRIVE/Vault-Install/models"

cp -a "$USB_MODELS/gguf/." "$HOME/.vault/models/gguf/"
cp -a "$USB_MODELS/reranker/." "$HOME/.vault/models/reranker/"

cd "$HOME/Vault"
./vault models configure \
  --chat Qwen3.6-35B-A3B-UD-Q4_K_M.gguf \
  --embedding Qwen3-Embedding-0.6B-Q8_0.gguf \
  --tokenizer qwen3-embedding-tokenizer \
  --wiki chat \
  --reranker bge-reranker-v2-m3

Wait for the configuration checks to finish. Do not close Terminal while the checks are running.

Step 7 — Start Vault

bash

cd "$HOME/Vault"
./vault start --open
./vault status

Vault opens at http://127.0.0.1:3000 (opens in a new tab). The status must show that Vault is active and that llama.cpp is running with the cuda backend.

Installation is complete.

If any command fails, stop and send the complete Terminal output to Wave2. Do not repeat the command with sudo.