Dual Station Deployment Guide

[agudanv]

Description

Adds a documented two-DGX Station deployment path for NVIDIA Nemotron 3 Ultra. The guide complements the existing single-Station recipe with a Ray-based PP=2 deployment across two GB300 systems connected by dual-rail CX8/RoCE networking.

Related Issue

N/A

Type of Change

• Documentation update

Changes Made

• Added a standalone DualStationDeploymentGuide covering fabric preflight, model-cache distribution, worker-first startup, Ray/vLLM launch, API validation, troubleshooting, and cleanup.
• Documented the PP=2, TP=1, 256K-context configuration, including NeMoClaw tool-calling and reasoning parser flags.
• Added an AIPerf benchmark procedure and measured-performance table for the synthetic shared-prefix workload.
• Linked the new guide from the Nemotron 3 Ultra cookbook index and cross-linked it from the single-Station guide.

Testing

Documentation-only change.

git diff --check origin/main...HEAD

The command completed successfully. No runtime deployment or model benchmark was run as part of this documentation-only update.

Checklist

• My code follows the existing style conventions (N/A: Markdown-only change)
• I have added type hints and docstrings where applicable (N/A: Markdown-only change)
• I have added tests for new functionality (N/A: Markdown-only change)
• All existing and new tests pass (N/A: Markdown-only change)
• I have updated documentation where needed
• My commits are signed off (git commit -s) per the DCO

Additional Notes

This is intentionally opened as a draft to collect technical review of the two-Station deployment guidance before it is considered ready for merge.