Description

Description

On DGX Station GB300, NemoClaw v0.0.43 express install (curl|bash → license=yes →
express=Y) auto-selects model Qwen/Qwen3.6-27B-FP8 and proceeds directly to vLLM
image pull. The 4-option vLLM-preflight model picker described in test case
T6015293 (1) Qwen2.5 72B, 2) DeepSeek-R1 Distill 70B, 3) MiniMax M2.7,
4) Nemotron-3 Super 120B NVFP4 default) is never displayed. Users wanting any
non-default model on the express path cannot select it; the gated DeepSeek-R1
Distill 70B HuggingFace-token discovery flow is also never exercised.

Either v0.0.43 regressed and removed the picker, or the picker moved elsewhere
and the case spec is now outdated. Dev needs to clarify expected behavior so
QA can mark T6015293 / T5987924 / T5987925 (DeepSeek-on-Station coverage).

Device:        DGX Station GB300 (galaxy-sku2-018, host 10.176.192.158)
OS:            Ubuntu 24.04.4 LTS
Architecture:  aarch64
Kernel:        6.17.0-1014-nvidia-64k
GPU:           NVIDIA GB300 (256703 MB) + NVIDIA RTX PRO 6000 Blackwell Max-Q (97887 MB)
NVIDIA driver: 610.39
Node.js:       v22.22.3 (auto-installed by curl|bash)
npm:           10.9.8
Docker:        29.5.0 (build 98f1464)
nvidia-ctk:    1.19.0
OpenShell CLI: 0.0.39
NemoClaw:      v0.0.43
OpenClaw:      2026.4.24 (cbcfdf6) — confirmed when sandbox is reached

1. Start from a clean Station (no nemoclaw / openshell binaries, no ~/.nemoclaw)
2. export HF_TOKEN=
3. Run: curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
4. Type 'yes' to accept the third-party software notice
5. At "Detected DGX Station. Run express install? [Y/n]:" → type 'y'
6. Observe the [3/8] Configuring inference (NIM) section

Per T6015293 spec, the wizard should show a 4-option model picker:
   1) Qwen2.5 72B
   2) DeepSeek-R1 Distill 70B
   3) MiniMax M2.7
   4) Nemotron-3 Super 120B NVFP4   (default — pre-selected on bare Enter)
User selects 2 to install DeepSeek-R1 Distill 70B; HuggingFace token discovery
runs, whoami-v2 validates, gated license check passes, vLLM serves DeepSeek-R1.

No picker is displayed. The wizard prints:

    [3/8] Configuring inference (NIM)
      ──────────────────────────────────────────────────
      [non-interactive] Provider: install-vllm
      vLLM (DGX Station):
        Image: nvcr.io/nvidia/vllm:26.03.post1-py3
        Model: Qwen/Qwen3.6-27B-FP8
        Image download on first run, cached after
        Model download on first run, cached after
      Installing vLLM. This can take 10–30 minutes; progress markers (==>) will print below.
      ==> Pulling vLLM image: nvcr.io/nvidia/vllm:26.03.post1-py3

Wizard is committed to Qwen3.6-27B with no user choice exposed.

Reproduced 2026-05-15 ~05:18-05:25 UTC. Full transcript saved at
  /home/lab/day0-automation/20260511/station1-t6015293-deepseek.txt
on local-lab. Same auto-pick behavior also observed in T6015057 run
(express full E2E, same Qwen3.6-27B auto-selection).

Blocks DevTest task 580929 (v0.0.43 T6015293) from being marked Pass per spec.
DeepSeek-R1 Distill 70B path on Station is unreachable via the documented
express flow; HF-token discovery and gated-model license check paths are not
exercised in QA on v0.0.43.

Bug Details

[NVB#6179554]