git clone https://github.com/vignesh2027/TemporalMesh-Transformer.git
cd TemporalMesh-Transformer
python3 -m venv .venv && source .venv/bin/activate
# Minimal test install (CPU-only torch, fast)
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install einops pytest pytest-cov ruff
pip install -e . --no-deps# All 15 tests must pass
pytest tests/ -v
# No lint errors
ruff check tmt/ tests/- Shape test for a new module →
tests/test_shapes.py - End-to-end / invariant test →
tests/test_forward.py - Use a small config (
d_model=64, n_heads=4) so tests stay fast on CPU
tmt/model/ ← Architecture (config, embedding, mesh, attention, ffn, exit_gate, memory, layers, model)
tmt/training/ ← Trainer, loss, LR scheduler
tmt/data/ ← Dataset loader, tokenizer wrapper
tests/ ← pytest test suite
docs/ ← GitHub Pages site
paper/ ← Figures + PDF
Every PR triggers GitHub Actions on Python 3.10, 3.11, and 3.12 (Ubuntu, CPU torch).
The lint job runs ruff check separately. Both must be green to merge.