TinyInfer-WASM

TinyInfer-WASM is a production-ready machine learning inference engine that runs entirely in the browser. Built with Rust and compiled to WebAssembly, it delivers near-native performance for neural network inference.

Architecture Overview

┌─────────────────────────────────────────────────────────┐
│                    Browser Environment                   │
├─────────────────────────────────────────────────────────┤
│  React UI  ←→  Web Workers  ←→  TinyInfer WASM Core    │
├─────────────────────────────────────────────────────────┤
│  Model Loader │ Tensor Engine │ Op Registry │ Memory   │
│  (ONNX/JSON)  │ (SIMD Accel)  │ (17+ Ops)   │ Pool     │
└─────────────────────────────────────────────────────────┘

Key Features

• 17+ Neural Network Operators: Conv2D, MatMul, Attention, LayerNorm, GELU, etc.
• SIMD Acceleration: 4x speedup using WebAssembly SIMD instructions
• Transformer Support: Full multi-head attention implementation
• INT8 Quantization: Model compression for faster inference
• Web Workers: Non-blocking inference with progress callbacks
• Memory Efficient: Smart tensor pooling and reuse

Performance Metrics

Supported Models

• MLP classifiers
• Convolutional networks (ResNet-style)
• Transformer encoders
• Custom architectures via JSON config