pyproject.toml

# LabVLA_v3 pytest + coverage + lint config.
# Plan-A Phase 0 — establishes test infra for all downstream work.
#
# Run tests:   pytest
# Coverage:    pytest --cov=src --cov=data_process
# Marker usage examples:
#   pytest -m "unit"              # fast CPU-only unit tests
#   pytest -m "not requires_data" # skip tests needing real dataset on disk
#   pytest -m "not gpu"           # skip tests needing GPU

# F22 fix: add minimal [project] metadata so the repo declares its name,
# Python requirement, and (where pinned) dependency set. Heavy runtime deps
# (torch, accelerate, deepspeed, transformers, etc.) are intentionally NOT
# pinned here because they are environment-specific — operators install them
# via the project's conda env. The metadata block exists so `pip install -e .`
# and downstream tools like `python -c "import importlib.metadata; ..."` work.
[project]
name = "labvla"
version = "0.3.0"
description = "LabVLA Vision-Language-Action policy (Qwen3-VL + DiT cross-attention head)."
# P2-G fix (round1): removed `readme = "README.md"` because the repo
# intentionally does not ship a root README.md (see the [tool.setuptools]
# NOTE below — this project is not pip-installable). A dangling readme path
# made some PEP 621 build backends warn or fail on metadata read. Add the
# field back if/when a real README lands at the repo root.
requires-python = ">=3.10"
authors = [{name = "BaochangRen"}]
license = {text = "Apache-2.0"}
# Lightweight runtime deps that are stable across the supported environments.
# Heavyweight ML stack (torch/accelerate/deepspeed/transformers) stays out of
# this list to avoid pinning a constellation that conflicts with the conda
# env operators actually use.
dependencies = [
    "numpy",
    "safetensors",
    "msgpack",
    "websockets",
    "pyarrow",
    "pandas",
    "draccus",
]

# NOTE: this project does NOT ship as a pip-installable package today; all
# imports are rooted at the repo root (`from src.X ...`; entry points
# self-insert the repo root into sys.path — see CLAUDE.md). The
# `[project]` block above only exists so distribution-introspection tools
# (importlib.metadata, pip show, etc.) can read the metadata, and so future
# packaging work has a starting point. We explicitly suppress automatic
# package discovery rather than wire it up to broken globs.
[tool.setuptools]
packages = []

[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py"]
python_functions = ["test_*"]
addopts = [
    "--tb=short",
    "--strict-markers",
    "-q",
]
markers = [
    "unit: pure-logic unit test, no I/O, no torch.cuda, no model load",
    "cpu_only: uses torch/numpy on CPU, no GPU required",
    "requires_data: needs a real LeRobot dataset on disk",
    "slow: >10 s on a single CPU core",
    "gpu: needs a real GPU + real VLM weights (skipped by default CI)",
]
filterwarnings = [
    # Don't let a DeprecationWarning from transformers poison the run.
    "default::DeprecationWarning",
]

[tool.coverage.run]
source = ["src", "data_process"]
omit = [
    "*/tests/*",
    "*/__pycache__/*",
    # Deployment + scripts are end-to-end; covered via integration tests
    # (when they land in later phases), not the unit/integration sweep.
    "deployment/*",
    "scripts/*",
]
parallel = false

[tool.coverage.report]
# Coverage baseline evolution:
#   Phase 0 baseline           → ~10%
#   Plan-A Phase 1-4 landed    → 47% (397 tests)
#   Sprint 2 + Sprint 3 landed → 42% (429 tests, +1765 LoC from new helpers)
# The drop from 47 → 42 is the expected denominator lift from Sprint 2/3 adding
# ~1700 lines of hardening/helper code (_reader_base.py, Liger version-gated
# patch, sprint2a/b/3 logic). Gate set to 40 with a small buffer; Sprint 4's
# remaining test-polish work + nightly GPU CI should bring this back to ≥45.
fail_under = 40
show_missing = true
skip_covered = false

[tool.ruff]
line-length = 100
target-version = "py310"

[tool.ruff.lint]
select = ["E", "F", "W", "I"]
ignore = ["E501"]   # line length — many configs have long URLs / block comments

[tool.black]
line-length = 100
target-version = ["py310"]