VRAM Calculator for LLMs

🤖 VRAM Calculator for LLMs - Your AI Hardware Companion

Planning to run Llama, Qwen, Mistral, or other large language models locally? Our VRAM Calculator helps you estimate exactly how much graphics memory you need before downloading or purchasing
hardware.

🎯 Key Features:

✅ Accurate VRAM Estimation
• Calculate memory requirements for weights and KV cache
• Support for dense and MoE (Mixture of Experts) architectures
• Precise formulas based on transformer architecture

✅ 38+ Predefined Models
• Llama 3, 3.1, 3.2 (1B to 70B)
• Llama 2 (7B, 13B, 70B)
• Qwen 2.5, 3, 3.5 (0.5B to 72B)
• Mistral, Mixtral (7B, 8x7B, 8x22B)
• Gemma, Phi, Yi, Command R/R+

✅ HuggingFace Integration
• Search and load model configurations directly
• Auto-detect architecture parameters
• Save time with instant model data

✅ Advanced Customization
• Manual architecture configuration (layers, hidden size, attention heads, KV heads)
• Quantization options (FP16, INT8, INT4, INT2)
• KV cache quantization settings
• Context window and batch size adjustment
• Flash Attention support (60-85% memory savings)

✅ Multi-Language Support
• Automatic language detection (English / Español)
• No manual toggle needed

✅ Dark Mode
• Follows system preference automatically

✅ Offline Capable (PWA)
• Works without internet connection
• Install as a web app on any device

✅ Free & Open Source
• No hidden costs
• Community-driven development

📊 How It Works:

1. Select a predefined model or search HuggingFace
2. Adjust quantization, context window, and batch size
3. Get instant VRAM estimates for:
• Model weights
• KV cache (context memory)
• Total VRAM with GPU recommendations

💡 Perfect For:

• AI enthusiasts planning local LLM deployments
• Developers optimizing inference hardware
• Students learning about transformer architectures
• Anyone comparing GPU options for AI workloads

🔧 Technical Details:

The calculator uses precise formulas based on transformer architecture:
• Weights: Parameters × bytes per parameter
• KV Cache: 2 × layers × KV_heads × head_dim × context × batch × bytes
• Flash Attention: 60-85% KV cache reduction

📱 Available On:
• Android (Google Play Store)
• Web (PWA - installable on any device)

💬 Feedback & Support:
We love hearing from users! Reach out with suggestions, bug reports, or feature requests.

⚠ Disclaimer:
This app provides estimates only. Actual VRAM usage may vary depending on the framework (llama.cpp, vLLM, Transformers), CUDA overhead (1-2GB), and memory fragmentation.

🔒 Privacy:
This app does not collect, store, or transmit any personal data. All calculations are performed locally on your device. No backend servers, no analytics, no tracking.