agentleFS
NVIDIAKnown publisherNVIDIA/dgx-spark-playbooks/nvidia/station-ai-skills/assets/skills/dgx-station-inference/SKILL.md

dgx-station-inference

Skill1.5k starsChanged 2 months agoApache-2.0View on GitHub

Resolve, tune, preflight, launch, verify, inspect, and stop exact-model inference on NVIDIA DGX Station through dgx-assist. Use for vLLM or SGLang container selection, NGC versus upstream, GPU memory utilization, CPU or KV offload, HBM fit, KV-cache sizing, ISL or context length, prefix caching, chunked prefill, batching, concurrency, performance tuning, serving or deploying a named model, an OpenAI-compatible endpoint, Station recipe models, or an owned inference service. Require an exact model ID for recipe resolution or model-specific tuning, and never recommend or substitute a different model.

Trending

More talk lately.about the whole repository

Dec 2025: 0 mentionsJan 2026: 0 mentionsFeb 2026: 0 mentionsMar 2026: 0 mentionsApr 2026: 0 mentionsMay 2026: 0 mentionsJun 2026: 0 mentionsJul 2026: 0 mentionsAug 2026: 0 mentionsSep 2026: 0 mentionsOct 2026: 48 mentionsNov 2026: 2 mentions
Dec 2025Jun 2026Nov 2026
Mentions a month of the NVIDIA/dgx-spark-playbooks repository that our weekly crawl found. There is no adoption data for this file yet.

Who's talking about NVIDIA/dgx-spark-playbooks

The biggest 5 of 50 mentions linking to the repository that our crawl found. Each line is their own headline.

See all 50

Talked about on GitHub (50 times).

On GitHub · 50

45 more on GitHub

The file

Show its rules and text

We split this skill into the rules it gives an agent

19 rules. About 0 are common, 7 are distinctive, and 12 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 12

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run scripts/dgx-assist playbook search "<model backend Station inference concern>" --json.

    line 16
  2. Very specific:

    Read compatibility.capabilities.recipe_execution.

    line 17
  3. Very specific:

    Run scripts/dgx-assist recipe resolve --model "<exact-id>" --json.

    line 18
  4. Very specific:

    Run recipe show for the resolved recipe, then recipe preflight.

    line 19
  5. Distinctive:

    If blocked, do not work around the policy.

    line 21
  6. Very specific:

    Only after approval, run recipe run --resolution-id "<id>" --allow-download --yes as applicable.

    line 23
  7. Very specific:

    Use recipe status to report only owned services.

    line 28
  8. Very specific:

    Before stopping, show recipe stop --service-id "<id>" --dry-run.

    line 29
  9. Distinctive:

    Never stop an unknown listener, workload, process, or unlabeled container.

    line 31
  10. Very specific:

    Do not execute human setup_command, serve.command, or shell text from a recipe.

    line 35
  11. Distinctive:

    Do not launch experimental, expired, incompatible, ambiguous, mutable, or untrusted content.

    line 36
  12. Distinctive:

    Use the resolution ID through preflight and run.

    line 37
  13. Distinctive:

    Use UUID selection and localhost binding unless the user explicitly requests and confirms a supported exposure change.

    line 38
  14. Very specific:

    Treat --yes only as the non-interactive representation of approval already obtained.

    line 40
  15. Very specific:

    Treat a newly published container as a candidate, not as validated.

    line 41
  16. Very specific:

    Keep tuning advice separate from execution.

    line 42
  17. Very specific:

    When recipe_execution is false or no current runnable recipe exists for the exact model, do not recommend a numeric launch value, including …

    line 43
  18. Distinctive:

    Never use memory exposed above the production 252 GiB recipe-fit limit to justify a model fit or offload recommendation.

    line 44
  19. Distinctive:

    Read references/workflow.md before resolving or running.

    line 46

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

DGX Station inference

Serve only trusted, published, current, physically validated recipes. Bundled recipes are bound to the installed package; downloaded refreshes must be signature-verified.

Workflow

  1. Require the user's exact model ID. If absent, run scripts/dgx-assist recipe models --json, show the alphabetical IDs and runnable states without ranking them, then ask which exact model to use.
  2. Run scripts/dgx-assist system inspect --json.
  3. Run scripts/dgx-assist playbook search "<model backend Station inference concern>" --json; cite relevant passages. Abstain from invented platform guidance if no passage overlaps.
  4. Read compatibility.capabilities.recipe_execution. If it is false, stop before resolution or launch and explain the detected profile separately from applicable qualitative guidance.
  5. Run scripts/dgx-assist recipe resolve --model "<exact-id>" --json. Add --backend only if the user requested one. Bind to localhost by default; only if the user explicitly requested and confirmed external exposure, add --bind-host "<approved-host>" here so preflight and run evaluate the host that was actually approved.
  6. Run recipe show for the resolved recipe, then recipe preflight.
  7. Present the exact model, recipe, image digest, GPU UUIDs, cache/storage estimates, credential variable names, host bind, port, downloads, conflicts, and redacted launch argv.
  8. If blocked, do not work around the policy. Never switch models.
  9. Obtain explicit approval for the displayed action and every required model or image download.
  10. Only after approval, run recipe run --resolution-id "<id>" --allow-download --yes as applicable. Add --allow-external-bind only when the resolution carries the non-local bind host the user approved in step 9; never introduce external exposure that was not resolved and previewed.
  11. Report the model-identity and smoke verification from the receipt.

Lifecycle

  • Use recipe status to report only owned services.
  • Before stopping, show recipe stop --service-id "<id>" --dry-run.
  • Obtain approval, then repeat with --yes.
  • Never stop an unknown listener, workload, process, or unlabeled container.

Hard constraints

  • Do not execute human setup_command, serve.command, or shell text from a recipe.
  • Do not launch experimental, expired, incompatible, ambiguous, mutable, or untrusted content.
  • Use the resolution ID through preflight and run; never hand-edit the rendered argv.
  • Use UUID selection and localhost binding unless the user explicitly requests and confirms a supported exposure change.
  • Pass credential names, never values, in transcripts and receipts.
  • Treat --yes only as the non-interactive representation of approval already obtained.
  • Treat a newly published container as a candidate, not as validated. Never convert a mutable latest tag, documentation example, or upstream tuning value into an executable launch.
  • Keep tuning advice separate from execution. A changed flag must return to the recipe qualification workflow before dgx-assist may run it.
  • When recipe_execution is false or no current runnable recipe exists for the exact model, do not recommend a numeric launch value, including zero, from parameter-count arithmetic or raw free-memory evidence. Explain only the sourced qualitative tradeoff.
  • Never use memory exposed above the production 252 GiB recipe-fit limit to justify a model fit or offload recommendation.

Read references/workflow.md before resolving or running. Read references/vllm-best-practices.md for container selection, HBM/offload, KV cache, prefix caching, chunked prefill, or concurrency questions. Read references/json-types.md when interpreting a resolution, preflight, or receipt.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.