[nabinchha]

📋 Summary

Default pipeline parallelism to one stage, independently of the number of nodes per replica. A TP16 deployment across two eight-GPU nodes now works with PP omitted; additional pipeline stages must be requested explicitly. Expert parallelism remains disabled by default and opt-in for compatible MoE models.

Earlier persisted runs are not migrated; compatibility with their prior topology and serialization semantics is intentionally outside this change.

🔗 Related Issue

Related to #850; follows #1112.

🔄 Changes

• Use pipeline_parallel: 1 in authored configuration and topology derivation, removing the node-count fallback.
• Cover omitted, disabled, and enabled EP across single-node and multi-node TP/PP layouts, including TP16 on eight-GPU nodes with PP and EP omitted.
• Verify benchmark child plans preserve default PP1 or explicit PP2 and replica counts when node placement changes.
• Provide a dedicated TP8 benchmark base run and compile every child plan from both the guide and shipped examples in regression tests. Document the two-node/eight-GPU requirement and per-child validation.
• Make pipeline fixtures request PP2 explicitly and refresh their derived checksums and benchmark identities.
• Clarify TP, PP, independent replicas, GPU arithmetic, and optional EP in the model-agnostic documentation. Include a two-node/eight-GPU example and update profile and benchmark guidance.

🧪 Testing

• Full Slurm suite: 2,127 passed, 1 platform-specific skip.
• Slurm-wide Ruff lint and formatting checks.
• Fern validation: zero errors; one authentication warning for the redirects check.
• Internal links across 101 Fern pages and all eight documented topology-table calculations.
• Live cluster rerun: N/A; this changes configuration defaults, with generated launch commands covered by integration tests.

✅ Checklist

• Follows commit message conventions
• Commits are signed off (DCO)
• User-facing Slurm documentation updated
• Architecture docs: N/A; the serving architecture is unchanged.

[github-actions]

Fern preview: https://nvidia-preview-pr-1116.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.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge within its stated saved-run compatibility limits.

Summary

This PR defaults pipeline_parallel to one, independently of node count.

• Slurm runs keep one pipeline stage unless you ask for more.
• Benchmark examples use a TP8 base run for node-placement comparisons.

nabinchha explicitly accepted saved-run incompatibility because preserving earlier topology and serialization behavior is outside this change. The prior unnumbered finding is not reposted.

_{Reviews (3) · Last reviewed commit: "fix(slurm): use a compatible benchmark b..." · Reviewed by Greptile}

[andreatnvidia]

Please address the benchmark example issue before merging.