Description

Description

On DGX Spark with 128GB unified memory, the onboard Ollama model probe rejects nemotron-3-nano:30b (21.2 GiB) because it checks "available" RAM (14.5 GiB) instead of total unified memory. In a normal user scenario (GNOME desktop + browser open per POR playbook path), available RAM drops well below total. The POR default model Qwen3.6 35B is even larger and will fail more easily. This breaks the COMPUTEX express setup path on any Spark with typical desktop usage.

Device:        DGX Spark (spark-dadc, 10.173.104.110)
OS:            DGX Spark FastOS 1.135.29 (2026-04-13, customer build)
Architecture:  aarch64
Node.js:       v22.22.2
npm:           10.9.7
Docker:        Docker 29.2.1
OpenShell CLI: openshell 0.0.36
NemoClaw:      v0.0.37
OpenClaw:      N/A (onboard not completed)
Ollama:        0.23.2
Total RAM:     121Gi (128GB unified memory)
Available RAM: 5.6Gi (with GNOME + Firefox running)

1. DGX Spark with GNOME desktop running (default OOBE state)
2. Open Firefox to build.nvidia.com/spark (POR-designed user path)
3. Run: nemoclaw onboard
4. Select provider: Local Ollama
5. Select model: nemotron-3-nano:30b (wizard default, Choose [2])
6. Wait for 22.6GB model download to complete
7. Observe probe result

Model probe accounts for Spark unified memory architecture (GPU and CPU share 128GB).
Probe should check total memory or GPU-accessible memory, not just free/available RAM.
nemotron-3-nano:30b (21.2 GiB) loads successfully on a 128GB Spark.

Two failure modes observed — both caused by probe not accounting for Spark unified memory:

1. With desktop running (5.6GB available, 128GB total):
   Selected Ollama model 'nemotron-3-nano:30b' failed the local probe: model requires
   more system memory (21.2 GiB) than is available (14.5 GiB)

2. After freeing memory (116GB available):
   Loading Ollama model: nemotron-3-nano:30b
   Selected Ollama model 'nemotron-3-nano:30b' did not answer the local probe in time.
   It may still be loading, too large for the host, or otherwise unhealthy.

Additionally, the wizard recommends nemotron-3-nano:30b as the default (Choose [2]).
User downloads 22.6GB, then probe rejects the model — terrible first-run UX.
Wizard should not recommend a model it cannot successfully probe.

$ free -h
               total        used        free      shared  buff/cache   available
Mem:           121Gi       116Gi       3.3Gi       118Mi       3.3Gi       5.6Gi
Swap:           15Gi       6.9Gi       9.1Gi

Bug Details

[NVB#6157916]