Add Tinker SDK LLM router tutorial for Nemotron Nano 30B

[vikalluru]

Summary

• Adds a Jupyter notebook tutorial: DPO + RLVR fine-tuning of Nemotron Nano 30B as a production LLM router using Tinker SDK
• Based on Glean's real production deployment: LoRA rank 32, RouterSearchEnv, reward = 0.7·R_search + 0.3·R_termination, vLLM on Vertex AI
• Uses open-source training data: Salesforce/xlam-function-calling-60k
• Covers full workflow: architecture overview → DPO fine-tuning → RLVR fine-tuning → vLLM serving config → offline eval
• Production latency achieved: P50=250ms, P95=2.5s on H100 single-replica

Location

usage-cookbook/Nemotron-3-Nano/tinker_llm_router_tutorial.ipynb

Test plan

• Notebook runs end-to-end on H100 with Tinker SDK installed
• DPO training cell completes with LoRA rank 32 config
• RLVR training cell completes with RouterSearchEnv reward function
• vLLM serving config verified against production params
• Offline eval cell produces router accuracy metrics

🤖 Generated with Claude Code