feat: Add Gemma-4 Multi-token-prediction Guide for TPUs

[syeda-anjum]

Overview:
This PR introduces a complete, end-to-end reference implementation for running Multi-Token Prediction (MTP) speculative decoding using Gemma-4 on Trillium TPUs. It provides the necessary vLLM manifests, inference-perf benchmarking configurations, and comprehensive architectural guidance.

Key Changes:

• Inference Overlays: Added the v6e-gemma-4-31b-it-mtp Kustomize overlay to support MTP on TPUs, including updating the base vllm-tpu image to v0.25.0 and ensuring the deployment is distinctly separated from the base Gemma-4 workload.
• Benchmarking Suite: Created dedicated inference-perf benchmark manifests tailored for sd-mtp to accurately measure MTP latency improvements.
• Comprehensive Documentation: Added the vllm-gemma-4-mtp.md guide, which includes:
  • Step-by-step instructions for tfvars configuration and downloading the assistant/drafter model.
  • A comparative analysis of MTP versus Standard Speculative Decoding, Draft-Head architectures (EAGLE), dSpark (SGLang), and N-gram prompt lookup.
  • Guidance on selecting optimal benchmarking datasets (e.g., CNN Daily Mail, HumanEval) to maximize speculative decoding acceptance rates.

[fernandorubbo]

How the result compares with standard inference (without MTP)?

Can you make an analysis here and share this with customers? Is it 30% better for this use case, 50% better? how customer analyse the results?