xlnstorch 1.0.0

A PyTorch backend for LNS experimentation

xlnstorch

XLNSTorch: a PyTorch addon python package for simulating Logarithmic Number System (LNS) arithmetic. You can find the docs at https://xlnsresearch.github.io/xlns/.

Getting Started

To install xlnstorch, you can run pip3 install xlnstorch. If you have a C++ compiler, this will allow you to use more efficient operations and layers. Otherwise, xlnstorch will fallback to pure python, slower implementations.

To learn more about LNS, see the LNS_Intro.md and src/xlnsconf/README.md files. For examples, look at the documentation and the examples directory.

Dependencies

xlnstorch has several dependencies that are automatically installed when you install it:

• torch
• xlns
• numpy

xlnstorch also has several optional dependencies for additional features:

• matplotlib: For many of the viz submodule's graphs
• graphviz: For visualizing the computational graph
• torchvision: For tensor transforms and datasets

To Do list

• Implement more transformer layers.
• Improve support for saving and loading LNSTensor weights, and copying weights between torch and xlnstorch.
• Rework float64 storage to bitcast rather than reinterpret types.
• Implement more layers in the C++ backend.
• Implement positive and negative infinity sentinel values.
• Improve type and shape checks/error messages.
• Support LNSTensors and operations performed on the GPU.
• Add more implementations from xlnsconf.