fix: onboard fails on GPUs with insufficient VRAM for local NIM

[CalebDeLeeuwMisfits]

Problem

nemoclaw onboard fails at sandbox creation with:

GPU sandbox requested, but the active gateway has no allocatable GPUs.

This happens because detectGpu() sets nimCapable: true for any NVIDIA GPU, regardless of VRAM. The --gpu flag is then passed to both openshell gateway start and openshell sandbox create — even when no NIM model fits in the available VRAM and the user selects cloud inference.

The sandbox silently fails to create (the error is piped through awk), but the CLI reports success. Then step 7 (policy presets) fails with "sandbox not found", leaving the user stuck.

Affects all consumer NVIDIA GPUs with <40GB VRAM (RTX 3060, 4060, 4070, etc.) and WSL2 environments where GPU passthrough to nested k3s containers is unavailable.

Fix

bin/lib/nim.js — nimCapable now checks whether at least one NIM model in nim-images.json has minGpuMemoryMB <= totalMemoryMB:

const canRunNim = nimImages.models.some((m) => m.minGpuMemoryMB <= totalMemoryMB);

bin/lib/onboard.js — Added an informational message during preflight when GPU is detected but too small for local NIM:

✓ NVIDIA GPU detected: 1 GPU(s), 8188 MB VRAM
ⓘ GPU VRAM too small for local NIM — will use cloud inference

Testing

Tested on WSL2 (Ubuntu) with an 8GB VRAM RTX GPU where the smallest NIM model requires 8192 MB.

Before: Onboard fails every time at step 3 — sandbox never created, policies can't apply.
After: Onboard completes successfully — gateway starts without --gpu, sandbox creates as CPU-only, cloud API inference works, agent responds.

Summary by CodeRabbit

• Bug Fixes

  • GPU capability detection now correctly validates VRAM requirements against model specifications, rather than assuming any detected GPU is suitable for local execution.

• New Features

  • Informational message now displays when a GPU lacks sufficient VRAM, notifying users that cloud inference will be used instead.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 7803fc0a-8b62-4471-996c-d22962b6cf7e

📥 Commits

Reviewing files that changed from the base of the PR and between bc509b9 and 2b5603d.

📒 Files selected for processing (2)

• bin/lib/nim.js
• bin/lib/onboard.js

---

📝 Walkthrough

Walkthrough

Updated GPU capability detection in NVIDIA path to conditionally set nimCapable based on whether available VRAM meets model requirements. Added informational logging in preflight flow when local NIM is unsuitable due to insufficient VRAM.

Changes

Cohort / File(s)	Summary
GPU Capability Detection bin/lib/nim.js	Modified detectGpu() NVIDIA path to compute nimCapable by checking if at least one model in nimImages.models has sufficient VRAM (minGpuMemoryMB <= totalMemoryMB), replacing unconditional true assignment.
Preflight Logging bin/lib/onboard.js	Added informational log emission in GPU detection flow when gpu.nimCapable is false, notifying that local NIM is unsuitable due to insufficient VRAM and cloud inference will be used.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Poem

🐰 GPU memory checked with care,
VRAM requirements laid bare,
When models won't fit locally tight,
Cloud inference shines so bright!
A rabbit's code, precise and fair ✨

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}