https://github.com/elinx/llm-mem-calculator

Interactive KV cache memory calculator for LLMs — supports MLA, GQA, hybrid attention, sliding window, and linear attention architectures. Estimate GPU memory for serving any model at any context length.
https://github.com/elinx/llm-mem-calculator

calculator gpu-memory kv-cache llm llm-serving vllm

Last synced: 3 days ago
JSON representation

Interactive KV cache memory calculator for LLMs — supports MLA, GQA, hybrid attention, sliding window, and linear attention architectures. Estimate GPU memory for serving any model at any context length.

• Host: GitHub
• URL: https://github.com/elinx/llm-mem-calculator
• Owner: elinx
• Created: 2026-06-06T12:49:48.000Z (6 days ago)
• Default Branch: main
• Last Pushed: 2026-06-06T15:01:13.000Z (6 days ago)
• Last Synced: 2026-06-06T15:08:15.138Z (6 days ago)
• Topics: calculator, gpu-memory, kv-cache, llm, llm-serving, vllm
• Language: JavaScript
• Homepage: https://elinx.github.io/llm-mem-calculator/
• Size: 493 KB
• Stars: 0
• Watchers: 0
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# KV Cache Calculator

A web-based tool for estimating LLM KV cache memory requirements. Supports modern architectures including MLA, GQA, hybrid attention, sliding window, and linear attention models.

**Live Demo**: [elinx.github.io/llm-mem-calculator](https://elinx.github.io/llm-mem-calculator/)

## Calculator

Calculate KV cache size for a single model with customizable parameters — context length, batch size, KV precision, and more.

![Calculator](./assets/calculator.png)

## Compare

Compare KV cache memory across multiple models side-by-side with an interactive chart.

![Compare](./assets/compare.png)

## Supported Architectures

| Architecture | Example Models |

|---|---|

| Standard GQA | Qwen3, Llama 3.x, Qwen2.5, MiniMax M2.x |

| MLA (Multi-head Latent Attention) | DeepSeek V3, DeepSeek R1, Kimi K2.5/K2.6 |

| DSA+MLA (DeepSeek V4 Hybrid) | DeepSeek V4 Pro, DeepSeek V4 Flash, DeepSeek V3.2, GLM-5/5.1 |

| Mixed Full + Sliding Window | Gemma 4, Cohere Command, MiMo-V2.5 |

| Linear + Full Hybrid | Qwen3.5, Qwen3.6 |

## Features

- **Precision options**: BF16/FP16, FP8/INT8, FP4/INT4

- **Draft KV cache**: Account for MTP/draft model KV layers

- **Linear attention KV**: Include linear attention layer contributions

- **Context presets**: Quick-select from 1K to 1M tokens

- **Breakdown view**: Detailed per-layer KV cache breakdown

- **Formula display**: Shows the exact formula used for each model

- **Dark mode**: Toggle between light and dark themes

- **Chart export**: Download comparison charts as PNG or copy to clipboard

## Development

No build step required — just open `index.html` in a browser or serve the directory with any static file server.

```bash

# Quick local server

python3 -m http.server 8765

```

## License

MIT