docs(minimax-h3): add RTX PRO 5000 profile

[lishunyang12]

Summary

• Add a validated 4× RTX PRO 5000 Blackwell PCIe hardware profile for MiniMax-H3.
• Use BF16, TP2 × Ulysses2, Ring1, text-encoder TP4, VAE patch parallel 4, and CUDNN attention.
• Document the five-step topology screen, peak memory, Nsight kernel breakdown, and topology-aware GPU ordering.
• Note the combined root service added by vLLM-Omni #5720 while retaining the qualified partition-specific commands.
• Mark the unreleased 1440p/2K mode as coming soon.

Validation

• YAML parse for taxonomy.yaml and models/MiniMaxAI/MiniMax-H3.yaml
• node scripts/build-recipes-api.mjs
• git diff --check

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 6, 2026 1:20pm

[gemini-code-assist]

Code Review

This pull request introduces a new docker_source configuration option for recipes, allowing Docker containers to run against a newer local source checkout without rebuilding the image. This is achieved by bind-mounting the host source path read-only and exporting it as PYTHONPATH inside the container. The MiniMax-H3 recipe is updated to utilize this feature. Feedback on the changes points out a potential issue in buildDockerRun where duplicate PYTHONPATH environment flags could be generated if the recipe's environment already defines one. Merging the paths instead of appending duplicate flags is recommended to prevent overwriting existing configurations.