Description

Description

On DGX Spark (spark-8158), NemoClaw v0.0.41 preflight reports "no GPU
detected" and "Sandbox GPU: disabled" (due to Bug #6175939: JMJWOA-
Generic-GPU detection failure). Despite this, the install-ollama provider
path proceeds without any guard: it installs Ollama, pulls the default
Spark model qwen3.6:35b (22.29 GB download), and completes onboard.

The install succeeds because the GPU physically exists (Ollama uses it
directly via the host), but the preflight's "no GPU" determination is
never consulted by the install-ollama flow. If the GPU were truly absent,
a 35B model on CPU-only inference would be unusable (minutes per response).

The installer should either:
(a) Block install-ollama when hostGpuDetected=false and surface a clear
    "GPU required for local inference" message, or
(b) Automatically downgrade to a smaller model (e.g. qwen2.5:7b) when
    no GPU is detected, with a warning.

Device:        DGX Spark (spark-8158, NVIDIA JMJWOA-Generic-GPU)
OS:            Ubuntu 24.04 LTS (kernel 6.17.0-1014-nvidia)
Architecture:  aarch64
Node.js:       v22.22.3
npm:           10.9.8
Docker:        29.2.1
NVIDIA Driver: 580.142 / CUDA 13.0
OpenShell CLI: 0.0.39
NemoClaw:      v0.0.41
OpenClaw:      2026.4.24

1. On DGX Spark where detectGpu() returns null (JMJWOA-Generic-GPU issue):
   export NEMOCLAW_NON_INTERACTIVE=1
   export NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1
   export NEMOCLAW_PROVIDER=install-ollama
2. Run: curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
3. Observe preflight output and subsequent Ollama install + model pull

When preflight determines hostGpuDetected=false, the install-ollama path
should either:
- Warn the user that local inference requires a GPU and suggest installing
  NVIDIA drivers first, or
- Automatically select a smaller model suitable for CPU inference, or
- At minimum, display a warning that the 35B model will be extremely slow
  without GPU acceleration

Preflight output:
  ⓘ Local NIM unavailable — no GPU detected
  ⓘ Sandbox GPU: disabled (no NVIDIA GPU detected)
  NVIDIA GPU hardware detected but nvidia-smi is not available.

install-ollama proceeds without any check:
  [non-interactive] Provider: install-ollama
  Pulling Ollama model 'qwen3.6:35b' (22.29 GB)
  ...
  success

No warning about GPU absence. No model downgrade. 22GB downloaded.

Preflight contradicts itself:
  ⓘ Sandbox GPU: disabled (no NVIDIA GPU detected)   ← no GPU
  NVIDIA GPU hardware detected but nvidia-smi...     ← but also yes GPU?

Then install-ollama ignores the GPU status entirely:
  Pulling Ollama model: qwen3.6:35b
  pulling f5ee307a2982: 100%  23 GB/23 GB  45 MB/s
  success

Note: On this specific machine the GPU works (Ollama uses it via host),
so inference actually succeeds. But the install path has no guard for
the case where GPU is truly absent.

Bug Details

[NVB#6175944]