Description

Description

When trying to install nemoclaw using the curl command, it fails during start openclaw gateway stage.
Manually starting a gateway works without issues using
openshell gateway start --gpu

Device:        DGX Spark and DGX Station
OS:            Ubuntu 24.04
Architecture:  x86_64
Node.js:       v22.22.2
npm:           10.9.7
OpenShell CLI: 0.0.36
NemoClaw:      v0.0.35
OpenClaw:      N/A (onboard not completed)

1. curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

Completes install and proceeds to onboard

aerial@spark-6087:~$ curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
   ███╗   ██╗███████╗███╗   ███╗ ██████╗  ██████╗██╗      █████╗ ██╗    ██╗
   ████╗  ██║██╔════╝████╗ ████║██╔═══██╗██╔════╝██║     ██╔══██╗██║    ██║
   ██╔██╗ ██║█████╗  ██╔████╔██║██║   ██║██║     ██║     ███████║██║ █╗ ██║
   ██║╚██╗██║██╔══╝  ██║╚██╔╝██║██║   ██║██║     ██║     ██╔══██║██║███╗██║
   ██║ ╚████║███████╗██║ ╚═╝ ██║╚██████╔╝╚██████╗███████╗██║  ██║╚███╔███╔╝
   ╚═╝  ╚═══╝╚══════╝╚═╝     ╚═╝ ╚═════╝  ╚═════╝╚══════╝╚═╝  ╚═╝ ╚══╝╚══╝
  Launch OpenClaw in an OpenShell sandbox.
[1/3] Node.js
  ──────────────────────────────────────────────────
[INFO]  Node.js found: v22.22.2
[INFO]  Runtime OK: Node.js v22.22.2, npm 10.9.7
[2/3] NemoClaw CLI
  ──────────────────────────────────────────────────
[INFO]  Installer payload is not a persistent source checkout — installing from GitHub…
[INFO]  Installing NemoClaw from GitHub…
[INFO]  Resolved install ref: latest
  ✓  Cloning NemoClaw source
  ✓  Preparing OpenClaw package
  ✓  Installing NemoClaw dependencies
  ✓  Building NemoClaw CLI modules
  ✓  Building NemoClaw plugin
  ✓  Linking NemoClaw CLI
  ✓  Installing OpenShell CLI
[INFO]  Created user-local shim at /home/aerial/.local/bin/nemoclaw
[INFO]  Created user-local shim at /home/aerial/.local/bin/nemoclaw
[INFO]  Verified: nemoclaw is available at /home/aerial/.local/bin/nemoclaw
[3/3] Onboarding
  ──────────────────────────────────────────────────
  Detected container runtime: docker
[INFO]  Installer stdin is piped; attaching the usage notice to /dev/tty…
[INFO]  Running nemoclaw onboard…
[INFO]  Found an interrupted onboarding session — resuming it.
[INFO]  Installer stdin is piped; attaching onboarding to /dev/tty…
  NemoClaw Onboarding
  (resume mode)
  ===================
  [1/8] Preflight checks
  ──────────────────────────────────────────────────
  [resume] Skipping preflight (cached)
  [resume] Continuing GPU passthrough from the saved onboarding session.
  [2/8] Starting OpenShell gateway
  ──────────────────────────────────────────────────
  Using pinned OpenShell gateway image: ghcr.io/nvidia/openshell/cluster:0.0.36
  Starting gateway cluster...
  Still starting gateway cluster... (5s elapsed)
  Still starting gateway cluster... (10s elapsed)
  Gateway start returned before healthy:
    Deploying local gateway nemoclaw...
    Checking Docker
    Downloading gateway
    Initializing environment
    x Gateway failed: nemoclaw
    CDI specs not found on host
    GPU passthrough via CDI was selected (your Docker daemon has CDI spec directories configured) but no CDI device specs were found on the host. Specs must be pre-generated before OpenShell can inject the GPU into the cluster container.
    To fix:
    1. Generate CDI specs on the host (nvidia-ctk creates /etc/cdi/ if it does not exist)
    sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml
    2. Verify the specs were generated and include nvidia.com/gpu entries
    nvidia-ctk cdi list
    3. Then retry: openshell gateway start --gpu
    Error: × failed to start gateway container
    ╰─▶ Docker responded with status code 500: CDI device injection failed:
    unresolvable CDI devices nvidia.com/gpu=all
  Waiting for gateway health...

[NVB#6152045]