development.md

TessellateIPU Development Guidelines

There are some guidelines for doing development on the TessellateIPU library.

Local install & packaging

For local development, we recommend using a pip editable install:

pip install cmake ninja nanobind scikit-build-core[pyproject]
pip install --no-build-isolation -ve .

The first line will install the required build dependencies, and the second will create an editable install of TessellateIPU. Note that the argument --no-build-isolation is optional, but speed up the compilation by avoiding re-creating a compilation virtual environnment at every call.

Source (sdist) and build (wheels) packages can easily be generated using build Python package:

pip install build
python -m build --no-isolation

Source tarball and compiled wheels will be created in the dist directory.

Development

We rely on pre-commit to perform basic checks on the Python code. Set up pre-commit with:

pip install pre-commit
pre-commit install
pre-commit run --all-files

Run unit tests using pytest:

pip install --no-build-isolation -ve .
pip install -r test-requirements.txt
JAX_IPU_USE_MODEL=true JAX_IPU_MODEL_NUM_TILES=16 pytest -v --tb=short ./tests/

Run a terminal with an IPU model (useful for local debugging):

JAX_IPU_USE_MODEL=true JAX_IPU_MODEL_NUM_TILES=8 ipython
TF_CPP_MIN_LOG_LEVEL=0 TF_CPP_VMODULE='poplar_compiler=1,poplar_executor=1' JAX_IPU_USE_MODEL=true JAX_IPU_MODEL_NUM_TILES=8 ipython

Useful dev. flags/env variables

• XLA_FLAGS='--xla_dump_to=./xla_dumps --xla_dump_hlo_pass_re=.': XLA passes dump;
• "debug.branchRecordFlush":"false": Deactivate Poplar loop profiling flush => smaller overhead;