Changelog

Changelog#

This is the initial release of Panther. As the library evolves, this document will track changes between versions.

Current Version (0.1.2)#

Core Features

  • SKLinear: Sketched linear layers with GPU acceleration

  • SKConv2d: Sketched 2D convolution layers

  • RandMultiHeadAttention: Randomized multi-head attention mechanism

  • CQRRPT: Randomized QR decomposition with column pivoting

  • Randomized SVD: Fast approximate singular value decomposition

  • SKAutoTuner: Hyperparameter tuning with multiple search algorithms (Grid Search, Random Search, Bayesian Optimization, etc.)

  • CUDA Backend: GPU-accelerated operations via pawX C++/CUDA extension

  • PyTorch Integration: Drop-in replacements for standard PyTorch layers

Platform Support

  • Windows and Linux support (CPU and GPU)

  • CPU-only operation fully supported (requires building from source)

  • CUDA 12.4+ for optional GPU acceleration

  • Python 3.12+

Current Limitations#

CUDA Support

  • CUDA 12.4+ required for GPU operations (optional)

  • All core features available in CPU-only mode

  • Tensor Core optimizations work best with dimensions that are multiples of 16

Performance

  • Small layers (< 512 parameters) may be slower than standard layers

  • First run may be slower due to CUDA kernel compilation

  • Memory benefits are most noticeable for large layers

Contributing#

Contributions are welcome! To contribute:

  1. Fork the repository on GitHub

  2. Create a feature branch

  3. Make your changes with tests

  4. Submit a pull request

See the repository for more details on development setup and guidelines.

Acknowledgments#

Panther builds upon excellent open source projects:

  • PyTorch: Deep learning framework

  • OpenBLAS: Optimized linear algebra routines

  • CUDA: GPU computing platform

  • Optuna: Bayesian optimization

  • Poetry: Python dependency management

License#

Panther is released under the MIT License. See the LICENSE file in the repository for details.