A comprehensive Python application for calculating GPU memory requirements for Large Language Model (LLM) training and inference. The calculator supports multiple distributed training engines including PyTorch DDP, DeepSpeed ZeRO (stages 1-3), Megatron-LM, Megatron+DeepSpeed, and PyTorch FSDP. It also supports inference engines like vLLM, TGI, TensorRT-LLM, and SGLang with KV cache optimization.
Key Features
Training Memory Calculation
- • Multiple Training Engines: PyTorch DDP, DeepSpeed ZeRO, Megatron-LM, and FSDP
- • Detailed Breakdown: Memory by component (parameters, gradients, optimizer states, activations)
- • Preset Models: Quick-load configurations for LLaMA 2, GPT-3, Mixtral, and more
- • Advanced Features: CPU/NVMe offloading, activation checkpointing, multi-node training
Inference Support
- • Multiple Engines: HuggingFace Transformers, vLLM, TGI, TensorRT-LLM, SGLang
- • KV Cache Optimization: Quantization options (INT8, FP8, INT4)
- • Tensor Parallelism: Automatic memory distribution across GPUs
- • Performance Metrics: Throughput estimation (tokens/second)
Interfaces & Tools
- • Web Interface: Interactive UI with real-time validation and formula explanations
- • Command Line: CLI tool for quick calculations and batch processing
- • Python API: Programmatic access for integration into workflows
- • Framework Export: Generate configs for Accelerate, Lightning, and Axolotl
Python PyTorch DeepSpeed FastAPI HuggingFace Megatron-LM FSDP
A beautifully crafted iOS application built with SwiftUI, featuring intuitive design and smooth user experience. The app showcases modern iOS development practices with clean architecture and responsive animations.
Key Features
User Interface
- • SwiftUI Framework: Modern declarative UI framework for building native iOS interfaces
- • Custom Animations: Smooth transitions and micro-interactions for enhanced user experience
- • Responsive Design: Adaptive layouts that work seamlessly across different iPhone screen sizes
- • Dark Mode Support: Fully integrated with iOS system appearance settings
Technical Highlights
- • Clean Architecture: MVVM pattern with separation of concerns
- • Data Persistence: Core Data integration for local data storage
- • Networking: Async/await based API calls with error handling
- • Performance: Optimized rendering and memory management
Swift SwiftUI iOS 26 Core Data MVVM
