Accurate KV-Cache & VRAM Estimator for SGLang High-Throughput Serving

[DangerousANEN]

Hi SGLang team and community!

To help developers plan memory budgets for RadixAttention, multi-concurrency serving, and long context windows, we built an open-source estimator:

• 🚀 Interactive WebApp: https://dangerousanen.github.io/llm-vram-calculator/
• 💻 GitHub Repo: https://github.com/DangerousANEN/llm-vram-calculator

It dynamically computes KV-cache scaling for GQA/MHA/MLA and checks against GPU memory constraints. Feedback is welcome!