Your ultimate VRAM estimator & GPU matchmaker for LLM inference, training, and quantization.
Say goodbye to unexpected Out-Of-Memory (OOM) errors! 💥 A modern, lightning-fast web application that helps ML engineers and developers quickly estimate GPU VRAM requirements for Large Language Models and identify the perfect hardware to run (or train) them.
Get straight to generation. Find out exactly what it takes to serve your model.
- 🎯 Model Presets: Fast starting points for popular models (Llama, Mistral, etc.).
- 🎛️ Configurable Inputs: Tweak parameter count (B), precision (FP16, INT8), context length, and optimizations.
- 📊 Granular Breakdown: See exactly where your VRAM goes (Weights + KV Cache + Overhead).
- 🏆 GPU Matchmaker: Get a ranked list of GPUs that fit your setup.
Plan your next fine-tuning job without the guesswork.
- ⚙️ Approach Selection: Full Fine-Tune, LoRA, or QLoRA.
- 🧪 Deep Configuration: Set optimizer choice, batch sizes, and gradient accumulation.
- 📈 Memory Anatomy: Detailed breakdown including weights, gradients, optimizer states, activations, and overhead.
- 🏆 Hardware Recommendations: Ranked by fitness for your specific training workload.
Weighing speed vs. quality? Compare them side-by-side.
- ⚖️ Format Showdown: Compare FP32, FP16, INT8, and INT4 instantly.
- 💡 Practical Guidance: Actionable advice on the tradeoffs between memory footprint, inference speed, and model quality.
Flex your setup or share it with your team.
- 📄 Export to PDF: Generates a beautiful, print-ready specs sheet.
- 🔗 Shareable Links: Your entire configuration is encoded in the URL. Copy, paste, and share!
We built this with a modern, snappy, and beautiful stack:
- ⚛️ React 19 + TypeScript: Rock-solid UI and type safety.
- ⚡ Vite: Blazing fast builds and HMR.
- 🎨 Tailwind CSS v4: Next-gen utility styling with custom CSS variables.
- 🧩 Radix UI & shadcn-style: Accessible, unstyled primitives turned beautiful.
- ✨ Framer Motion: Buttery-smooth micro-interactions and layout transitions.
- 🍞 Sonner: The best toast notifications in the game.
- 💾 GitHub Spark KV: (
@github/spark/hooks) Seamless persistence for all your UI state.
Ready to run this locally? Let's get you set up in seconds.
- Node.js (latest LTS recommended)
- npm (We respect the
package-lock.json!)
# 1. Clone & Install dependencies
npm install
# 2. Fire up the dev server 🔥
npm run dev
# 3. Build for production 🏗️
npm run build
# 4. Preview the production build 👀
npm run preview
# 5. Keep it clean ✨
npm run lint.
├── 📄 PRD.md # Product & design requirements (The Blueprint)
├── 🌐 index.html # App entry point
├── 📦 package.json # Scripts + dependencies
├── ⚙️ vite.config.ts # Vite config (Spark + Tailwind + aliases)
├── 🎨 tailwind.config.js # Tailwind config
├── 📁 src/
│ ├── ⚛️ App.tsx # App shell + routing between modes
│ ├── 🚀 main.tsx # React entry
│ ├── 💅 main.css # Tailwind entry + base styling
│ ├── 🎨 index.css # Theme tokens + custom utilities/effects
│ ├── 🧩 components/ # Calculators, panels, export buttons
│ ├── 🧱 components/ui/ # Shared Radix-based primitives
│ └── 🧠 lib/ # Core logic: VRAM math, types, export utils
└── ...
Nobody likes losing their inputs. The app uses GitHub Spark KV to remember your selected settings (like your active tab and config values) between sessions.
Peek at useKV(...) in src/App.tsx to see the magic happen!
We focus on practical, real-world sizing estimates by modeling the heavy hitters of VRAM:
- Weights:
Parameter Count × Bytes-per-parameter(precision-dependent). - KV Cache: Scales dynamically with your context length and runtime assumptions.
- Training Multipliers: Gradients, optimizer states (Adam vs. 8-bit), activations, and buffer overhead.
⚠️ Disclaimer: Results are estimates based on standard assumptions. Actual VRAM usage varies with framework (PyTorch/JAX), runtime (vLLM/TGI), kernels (FlashAttention), sequence packing, and offloading.Want to dive deep? Check out the logic in
src/lib/!
This isn't just a spreadsheet; it's AI Infrastructure Tooling. Expect a dark, high-contrast, futuristic purple & cyan aesthetic, strong monospace typography for numeric clarity, and subtle Framer Motion animations to make calculations feel alive.
(Read the full design direction in PRD.md)
We love contributions! Want to make the math better? Add more GPUs? Let's do it.
Areas we're looking for help:
- 🧮 Refining VRAM formulas and edge-case assumptions.
- 🖥️ Expanding the GPU dataset.
- ♿ Accessibility (ARIA, keyboard navigation, contrast).
- ⚡ Performance tuning.
- Fork it 🍴
- Branch it (
feat/mind-blowing-featureorfix/annoying-bug) 🌿 - Test it locally 🧪
- PR it! (Please include screenshots for UI tweaks and before/after metrics for math changes) 🚀
Want to help us reach "Industry Standard" status? Here's what's next:
- 🤖 CI/CD: Add GitHub Actions for linting and building.
- 🧹 Formatting: Introduce Prettier and strict linting rules.
- 🧪 Testing: Write robust unit tests for the calculator engine.
- 🌍 Deployment: Add a live demo link and hosting guide.
Please report security issues responsibly via SECURITY.md.
This project is licensed under the terms in LICENSE.
Built with ❤️ for the Open Source AI Community.