Add llm-vram-calculator to Local / On-device Inference

[159753a52]

Project

• Name: llm-vram-calculator
• URL: https://github.com/159753a52/llm-vram-calculator (web version: https://modelvram.com/llm-vram-calculator/)
• Category: 3. Inference Engines & Serving → Local / On-device Inference (next to llmfit)

Why it belongs

It helps people running models locally decide what fits before downloading: it estimates inference memory (weights, KV cache, runtime overhead) for any Hugging Face model from its config.json and safetensors metadata, rather than from a fixed model table. It handles cases generic calculators get wrong: all MoE experts counted toward weights, MLA's compressed latent cache, sliding-window layers, linear-attention / state-space layers, and MXFP4 / 4-bit expert checkpoints.

Quality signals

• License: MIT
• Maintenance status: actively maintained (last push 2026-09-29)
• Documentation/examples: README documents each architecture case it models; tests pin results to real checkpoint sizes
• Distinction from similar projects: llmfit scores models against your detected hardware from a curated list; this computes memory per model from the Hub's own files and runs in the browser without install

python3 tools/validate_awesome.py --skip-remote passes with 0 errors (the 10 warnings are pre-existing and unrelated).

Disclosure: I built and maintain this project. GitHub stars are low since it's new; happy to adjust or drop it if it doesn't meet the bar.