[dhruvnathawani]

📋 Summary

Allow Slurm recipes to use hosted model endpoints alongside managed vLLM models, including hosted-only recipes in CPU allocations. Existing DD async execution remains unchanged.

Targets feat/slurm-execution (integration PR #1040). Keep draft until mixed-provider live qualification is complete.

🔗 Related Issue

Closes #1079

🔄 Changes

• Add explicit external providers with runtime secret references, request parameters, and headers. Require complete, disjoint model coverage and preserve each external model's identity and inference settings.
• Allow zero local deployments: request one CPU node with --gres=none and skip GPU discovery, serving images, containers, and proxies.
• Reuse existing shard, retry, resume, and collection contracts. Preserve serialized local-only plans when new defaults are omitted.
• Add per-node submission.memory for generation and collection; an explicit value overrides GPU memory requests for generation.
• Fix native fresh-output handoff: let DD exclusively create the directory, then restrict its permissions before winner publication. Preserve existing nonempty and linked paths.
• Add mixed and hosted-only examples and public Fern documentation.

Async execution does not guarantee continuous GPU activity. Representative utilization and cluster policy remain prerequisites for qualifying the SQL migration.

🧪 Testing

• Baseline make test across all four packages: 6,531 passed, 3 skipped.
• Final make test-slurm including live-run regressions: 2,019 passed.
• make check-all, Slurm Ruff checks, and git diff --check passed.
• Native DD HTTP tests cover mixed and hosted-only routing, seeded and unseeded outputs, credentials, and validation through the real candidate-artifact verifier.
• make check-fern-links passed across 101 pages; full Fern validation and all visible PR checks passed on the final head.
• Real hosted-only Slurm smoke on the final code: 2/2 authored SQLite fixture rows, both judgments pass, successful attempt and winner publication, then successful winner-driven collection of the same two rows. Generation and collection each allocated four CPUs, 16 GB, and zero GPUs. Persisted text artifacts passed the hosted-credential scan.
• Mixed live qualification is blocked by the independent serving-runtime gap in Slurm: support Qwen3.8-capable vLLM runtimes #1085: the preserved model-specific image reports a development version outside the vLLM 0.21/0.22 matrix. Native import and inspection of the official vLLM 0.22.0 image succeeded, but a CPU-only check against the unchanged Qwen3.8-Flash-Next checkpoint failed because Transformers does not recognize qwen4_exp and the image lacks Qwen4ExpForConditionalGeneration. No GPU allocation or utilization measurement has occurred.

✅ Checklist

• Conventional commits and signed-off DCO.
• Internal contribution; external-contributor triage does not apply.
• Public documentation updated; architecture boundaries preserved.

[github-actions]

All contributors have signed the DCO ✍️ ✅
_{Posted by the DCO Assistant Lite bot.}

[github-actions]

Fern preview: https://nvidia-preview-pr-1080.docs.buildwithfern.com/nemo/datadesigner

Fern previews include the docs-website version archive with PR changes synced into latest. Notebook tutorials are rendered without execution outputs in previews.