Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
58 lines (40 loc) · 1.57 KB

CONTRIBUTING.md

File metadata and controls

58 lines (40 loc) · 1.57 KB
Raw

Contributing to meTile

Setup

pip install -e ".[dev]"

Development Workflow

make format    # auto-format + fix lint
make check     # lint + format check (no changes)
make test      # run tests
make code-qual # dead code detection
make ci        # all of the above
make bench     # run all benchmarks

Debug Output

Set METILE_DEBUG to inspect the compilation pipeline:

METILE_DEBUG=all python your_script.py

Flags: tile_ir, metal_ir, metal_ir_opt, msl, all (comma-separated). Output is written to debug_output/ by default.

Running Tests

python -m pytest tests/ -x -q           # all tests
python -m pytest tests/test_gemm.py -v  # specific file

Performance Regression Dashboard

Kernel performance is tracked across commits:

https://andreslavescu.github.io/meTile/dev/bench/

Every push to main runs benchmarks on an M1 runner and publishes results. PRs automatically compare against the baseline and fail if any kernel regresses by more than 15%.

To run regression benchmarks locally:

python benchmarks/regression.py                          # print results
python benchmarks/regression.py --output baseline.json   # save baseline
python benchmarks/regression.py --compare baseline.json  # compare + fail on regression

PR Guidelines

  • Run make ci before submitting
  • If modifying compiler passes, verify both GEMM and elementwise tests pass
  • If modifying codegen, check generated MSL with METILE_DEBUG=msl
  • Performance-sensitive changes should include benchmark results in the PR description