Component / Page

• Page: docs/deployment/deploy-to-remote-gpu.mdx ("Deploy NemoClaw to a Remote GPU Instance")
• Validation machine: 2u2g-gen-0642 (Ubuntu 24.04.4, NVIDIA RTX PRO 6000 Blackwell)
• NemoClaw CLI: v0.0.61

Description

This page describes how to run NemoClaw on a remote GPU instance and explicitly states the preferred path is to provision the VM, run the standard installer on that host, and then run nemoclaw onboard. However, all code blocks on the page use the deprecated nemoclaw deploy Brev wrapper, and there is no concrete example of the recommended installer + onboard workflow. This makes the doc internally inconsistent and encourages the deprecated path despite the warning.

Steps to Reproduce

1. Open docs/deployment/deploy-to-remote-gpu.mdx in the rendered docs.

2. Read the introduction and Prerequisites section — it states:

"The preferred path is to provision the VM, run the standard NemoClaw installer on that host, and then run nemoclaw onboard."

3. Scroll through all code blocks on the page and note they all use nemoclaw deploy:

• nemoclaw deploy <instance-name> (deploy instance)
• nemoclaw deploy <instance-name> again to reconnect
• ssh <instance-name> 'cd ~/nemoclaw && ... && openshell term'
• openclaw agent --agent main -m "Hello from the remote sandbox" ...
• export CHAT_UI_URL=...; nemoclaw deploy <instance-name>
• export NEMOCLAW_SANDBOX_READY_TIMEOUT=600; nemoclaw onboard
• export NEMOCLAW_PROXY_HOST=...; export NEMOCLAW_PROXY_PORT=...; nemoclaw onboard
• export NEMOCLAW_GPU="..."; nemoclaw deploy <instance-name>

Note there is no example showing: provisioning the VM, running the curl installer on the remote host, or running nemoclaw onboard as the primary flow.

4. Observe runtime output from nemoclaw deploy on the validation machine:

⚠ nemoclaw deploy is deprecated and will be removed in a future release.
  Prefer provisioning the remote host separately, then run the standard
  installer and nemoclaw onboard on that host.

Expected Result

The page should lead with and show actual commands for the preferred path:

1. Provision VM via Brev or cloud console.
2. SSH to the VM (brev ssh <instance-name> or ssh <host>).
3. Run the standard installer on the VM: curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
4. Run nemoclaw onboard on the remote VM, with env vars (CHAT_UI_URL, proxy, NEMOCLAW_SANDBOX_READY_TIMEOUT, etc.) set there.

The deprecated nemoclaw deploy flow should be clearly scoped as legacy compatibility only — either moved to a subsection or labelled as such.

Actual Result

• The intro and <Warning> banner state nemoclaw deploy is deprecated and the preferred path is manual provisioning + installer + nemoclaw onboard.
• Despite that, every concrete deployment example uses nemoclaw deploy. There is no copy-pasteable example of the preferred workflow.
• nemoclaw deploy itself prints a deprecation warning when executed, reinforcing that the doc steers users toward a path the CLI discourages.

Additional flow gaps:

• No expected-output examples for any deploy/onboard commands.
• No error-handling guidance for missing brev CLI, even though "Brev CLI installed and authenticated" is listed as a prerequisite.
• ssh <instance-name> assumes Brev has configured an SSH alias; the doc never explains where <instance-name> comes from.
• openclaw agent ... "Verify Inference" example does not clarify whether it runs on the VM host vs. inside the sandbox shell.
• NEMOCLAW_GPU examples use a GCP-specific value without noting the format differs per cloud provider.

Suggested Fix

1. Add a concrete "preferred" installer + onboard flow section:

# On your local machine
brev create <instance-name> --gpu <type>   # or console-based provisioning

# Then on the remote VM
ssh <instance-name>
curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
export CHAT_UI_URL=...
export NEMOCLAW_SANDBOX_READY_TIMEOUT=600  # if needed
nemoclaw onboard

Label this section "Preferred deployment path" and make it the primary flow.

2. Re-scope nemoclaw deploy content — move it into a "Legacy Brev compatibility" subsection with a clear label that this flow is deprecated and will be removed.

3. Clarify ssh <instance-name> and command context — add a note that <instance-name> is the Brev SSH alias, and clarify whether openclaw agent ... runs on the VM host or inside the sandbox shell.

4. Augment error-handling and expected output — add a note for missing/unauthenticated brev CLI, and add a short "you should see…" snippet after a successful onboard.

5. Clarify NEMOCLAW_GPU format — note that a2-highgpu-1g:nvidia-tesla-a100:1 is GCP-specific and other providers use different GPU type strings.

Severity / Priority

• Severity: Medium (doc recommends one flow but only shows commands for the deprecated one; can mislead users)
• Priority: Medium (affects users deploying NemoClaw to remote GPUs; the CLI itself warns about the deprecation)