dgx-station

name

dgx-station

description

Inspect and guide NVIDIA DGX Station GB300 development using the local dgx-assist CLI and pinned NVIDIA playbooks. Use for general Station platform questions, Software 1.0 or 2.0 compatibility, GB300 or RTX GPU selection, UUID ordering, mixed ATS/HMM coherency, CDMM, general containers, CDI, CUDA visibility, or vsloshd power-sloshing behavior. Do not use for vLLM or SGLang container selection or tuning, serving a named model, changing MIG, or troubleshooting a reported failure when the dedicated Station skill applies.

DGX Station

Ground every Station-specific answer in current host evidence and a retrieved NVIDIA passage.

Workflow

1. Run scripts/dgx-assist system inspect --json.
2. Read references/software-compatibility.md. Use compatibility.profile_id, support_level, and capabilities; do not branch on a version prefix.
3. Form a narrow playbook query from the user's question and run scripts/dgx-assist playbook search "<query>" --json.
4. Apply the source-precedence rules in references/sources.md. Use only passages applicable to the detected profile. Cite their URL, heading, line span, source commit, and source-file digest.
5. On the Software 1.0 capability-scoped profile, provide the supported inspection, diagnostics, compatibility guidance, and exact recipes qualified for that profile. Do not treat an absent Software 2.0-only service as a fault.
6. Before any platform action, require its named capability to be true. If it is false, explain the profile restriction separately from relevant read-only guidance.
7. If version_specific_guidance is false or retrieval abstains, say no applicable version-specific passage was found and avoid inventing a platform command.
8. Answer from the combined profile and applicable passage. Label unknown evidence as unknown.

Safety requirements

• Before any mutating action, display the exact command to be run and obtain explicit user approval immediately before executing it. Never batch approvals or carry one forward to a later action.
• Never assume an nvidia-smi index is a CUDA ordinal.
• Use GPU UUIDs for every proposed launch. When multiple GPUs are deliberately visible, place the GB300 UUID first.
• Do not apply either legacy or Software 2.0 mixed-device behavior without checking the detected profile and observed ATS/HMM state.
• Do not tell the user to install or operate Fabric Manager as a normal Station requirement.
• When mixed_coherency_service is true, inspect mixed-coherency-gpu-select.service, its generated environment, and container exposure separately.
• When dynamic_power_sloshing is true, inspect vsloshd. Always inspect observed caps and never propose an ad hoc power-cap change.
• Never install or change the driver, kernel, OS, firmware, or packages.
• Never display credential values.

Read references/platform.md when interpreting coherency, container ordering, power, memory placement, or qualification evidence. Read references/software-compatibility.md before applying version-specific guidance. Read references/sources.md when guidance may come from the Development Guide or bring-up guide. Read references/cli.md for command and JSON details.