TinyQ Roadmap

[afondiel]

TinyQ Roadmap

Phase 1 - Current Focus

• Basic Linear Layer Quantization
  • W8A32 implementation
  • W8A16 implementation
  • Documentation and examples
  • Unit tests for quantization methods

Phase 2 - Core Features

• W8A8 Quantization Support
  • Activation calibration infrastructure
  • Dynamic range calibration
  • Per-tensor quantization for activations
• Model Support Extensions
  • Custom PyTorch model support
  • TorchVision models integration
  • Additional HuggingFace model architectures

Phase 3 - Advanced Features

• Additional Layer Support
  • Conv2D layers
  • BatchNorm layers
  • Embedding layers
• Lower Bit Precision
  • INT4 quantization (W4A8)
  • Binary quantization exploration
• Performance Optimization
  • CPU instruction set optimization
  • Memory usage optimization
  • Inference latency improvements

Phase 4 - Tools & Infrastructure

• Comprehensive Benchmarking Suite
  • Automated accuracy testing
  • Performance benchmarking (CPU/Memory)
  • Model size comparison tools
• Advanced Calibration Methods
  • MSE-based calibration
  • Entropy-based calibration
  • Per-channel activation calibration
• Developer Tools
  • CLI interface for quantization
  • Model inspection tools
  • Debugging utilities

Future Considerations

• Quantization-Aware Training (QAT) support
• Mixed-precision quantization
• Dynamic quantization options
• ONNX export support
• Integration with other deployment frameworks

If it's your first time contributing/collaborating on this project, please check the Contributing Guidelines to see how to impactful contribute to the project.