structure

Repository

I've thought for a few weeks on how to structure this whole library, getting inspiration from the pytorch and tinygrad repositories. At a high level, the actual package repository is in pyember/ember, which uses functions pybinded from pyember/aten for fast computations.

I tried to model a lot of the structure from Pytorch and TinyGrad. Very briefly,

1. aten/ contains the header and source files for the C++ low-level tensor library, such as basic operations and an autograd engine.
   1. aten/src contains all the source files and definitions.
   2. aten/bindings contains the pybindings.
   3. aten/test contains all the C++ testing modules for aten.
2. ember/ contains the actual library, supporting high level models, objectives, optimizers, dataloaders, and samplers.
   1. ember/aten contains the stub files.
   2. ember/datasets contains all preprocessing tools, such as datasets/loaders, standardizing, cross validation checks.
   3. ember/models contains all machine learning models.
   4. ember/objectives contain all loss functions and regularizers.
   5. ember/optimizers contain all the optimizers/solvers, such as iterative (e.g. SGD), greedy (e.g. decision tree splitting), and one-shot (e.g. least-squares solution).
   6. ember/samplers contain all samplers (e.g. MCMC, SGLD).
3. examples/ are example python scripts on training models.
4. tests/ are python testing modules for the ember library.
5. docker/ contains docker images of all the operating systems and architectures I tested ember on. General workflows on setting up the environment can be found there for supported machines.
6. setup.py allows you to pip install this as a package.
7. run_tests.sh which is the main test running script.

For a more detailed explanation, look here.

ATen

Aten, short for "a tensor" library (got the name from PyTorch), is a C++ library that provides low level functionality for Tensors. This includes the basic vector and matrix operations like addition, scalar/matrix multiplication, dot products, transpose, etc, which are used everywhere in model training and inference and must be fast.

Compiling and PyBinding

Let's look at aten/CMakeLists.txt and aten/binding/CMakeLists.txt.

• aten/CMakeLists.txt contains the instructions to generate a Makefile for compiling and linking the aten library. It has an optional argument BUILD_PYTHON_BINDINGS when set ON, will generate the .so file through aten/binding/CMakeLists.txt. The executables compiled with aten/main.cpp are compiled to aten/build/main. Same for the test files which are compiled to aten/build/tests.

• aten/binding/CMakeLists.txt contains the instructions to generate the .so file and saves it to pyember/ember/_C.cpython-312-darwin.so. It must be contained within the Python package directory, since embercannot access libraries outside of its base directory.