.gitignore

# ── Python bytecode / build artifacts ────────────────────────────────
__pycache__/
*.py[cod]
*$py.class
*.so

# Distribution / packaging
build/
dist/
wheels/
*.egg
*.egg-info/
.eggs/
pip-wheel-metadata/

# ── Test / type / lint caches ────────────────────────────────────────
.pytest_cache/
.mypy_cache/
.ruff_cache/
.tox/
.nox/
.coverage
.coverage.*
htmlcov/
coverage.xml
*.cover
.hypothesis/

# ── Triton / CUDA / cuTeDSL JIT caches ───────────────────────────────
.triton/
# Per-primitive TRITON_CACHE_DIR populated by
# benchmarks/vs_cuml/heavy/run_all_parallel.py — its _run_one() sets
# TRITON_CACHE_DIR=<repo>/.triton_cache/<prim> per subprocess so each
# primitive's autotune state is isolated.
.triton_cache/
.cute_dsl/
.nv_cache/
.cuda_cache/

# ── Notebooks ────────────────────────────────────────────────────────
.ipynb_checkpoints/

# ── Editors / OS ─────────────────────────────────────────────────────
.vscode/
.idea/
*.swp
*.swo
.DS_Store
Thumbs.db

# ── Tuner artifacts (machine-local) ──────────────────────────────────
# Per-device JSONL grids written by `python -m benchmarks.tune.<op>`.
# The hand-written rules in `flashlib/.../<op>/route.py` are committed;
# the raw measurements are not.
benchmarks/tune/results/*
!benchmarks/tune/results/.gitkeep

# ── Benchmark results (regenerated locally by the user) ──────────────
# Every script under benchmarks/ writes to benchmarks/results/. The
# results are machine + version specific and meant to be reproduced
# locally — they are not committed. Re-run e.g.:
#   python -m benchmarks.vs_cuml.heavy.run_all_parallel --gpus 8
#   python -m benchmarks.vs_cuml.broad.run_all --gpus 8
#   python -m benchmarks.micro.bench_<name>
benchmarks/results/

# ── Local-only ───────────────────────────────────────────────────────
*.log
.env
.env.local