tinybloat

A library to extend tinygrad functionality with stuff that is probably too bloated for tinygrad.

Installation

Just run pip install -U git+https://github.com/softcookiepp/tinybloat.git

Documentation

Simply open docs/index.html in your browser of choice, or run the following command to open a documentation webserver: pdoc src/tinybloat

Contributing

Contributions to this codebase require only 2 things:

1. New functionality must have docstrings that describe its use in some capacity.
2. Unit tests must be written to validate it against the original implementation (usually pytorch or numpy.)

Roadmap

These are functionality that should be reimplemented in the future:

High priority:

• Function for reliably determining the supported data types of a given hardware device
• Finish implementing ComplexTensor operators such that it reaches feature parity with tinygrad.Tensor

Lower priority:

• All the loss functions from pytorch
• torch.nn.CosineSimilarity
• torch.nn.PairwiseDistance
• any other functionality that torch or other ML libraries have