pyproject.toml

[build-system]
requires = ["uv_build==0.7.6"]
build-backend = "uv_build"

[tool.uv]
environments = [
    "sys_platform == 'linux' and platform_machine == 'x86_64'",
    "sys_platform == 'linux' and platform_machine == 'aarch64'",
    "sys_platform == 'darwin' and platform_machine == 'x86_64'",
    "sys_platform == 'darwin' and platform_machine == 'arm64'",
]
# bfcl-eval hard-pins an old dependency set (numpy==1.26.4, filelock, etc.).
# Mark bfcl as conflicting with the tooling extras so uv resolves it in its own
# fork; otherwise those old pins drag shared dev/test/performance deps (filelock,
# virtualenv) down to versions with known CVEs. CI installs the tooling extras
# without bfcl; bfcl is installed standalone (pip install -e ".[bfcl]").
conflicts = [
    [{ extra = "bfcl" }, { extra = "dev" }],
    [{ extra = "bfcl" }, { extra = "test" }],
    [{ extra = "bfcl" }, { extra = "performance" }],
]
# CVE floors for transitive deps we don't declare directly. click is pulled in
# via transformers -> typer; 8.3.2 carries PYSEC-2026-2132 (fixed in 8.3.3).
# setuptools is pulled in via pbr/torch; 82.0.1 carries PYSEC-2026-3447 (fixed
# in 83.0.0).
constraint-dependencies = ["click>=8.3.3", "setuptools>=83.0.0"]

[tool.uv.build-backend]
module-root = "src"
source-exclude = [
    "inference_endpoint/evaluation/livecodebench/_server.py",
    # Isolated uv subproject (own pyproject.toml/uv.lock; pinned transformers /
    # numpy<2 / prm800k that conflict with the parent env). Invoked only via
    # `uv run --project`, never imported - keep it out of the parent wheel.
    "inference_endpoint/evaluation/legacy_mlperf_deepseek_r1/**",
    # Isolated native uv service that owns SWE-bench Docker execution.
    "inference_endpoint/evaluation/swebench_service/**",
]

[project]
name = "inference-endpoint"
version = "0.1.0"
description = "High-performance LLM endpoint benchmarking system with 50k QPS capability"
readme = "README.md"
license = {text = "Apache-2.0"}
authors = [
    {name = "MLPerf Inference Endpoint Team"}
]
keywords = ["mlperf", "benchmarking", "llm", "endpoint", "performance"]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: Apache Software License",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Software Development :: Testing",
    "Topic :: System :: Benchmark",
]
requires-python = ">=3.12"
dependencies = [
    # Core Python dependencies
    "typing-extensions==4.15.0",
    # System utilities
    "psutil==7.2.2",
    # Async and networking
    "pyzmq==27.1.0",
    "uvloop==0.22.1",
    "httptools==0.7.1",
    "websocket-client==1.9.0",
    # Data handling
    "duckdb==1.5.1",
    "msgspec==0.20.0",
    "pydantic==2.12.5",
    "pydantic_core==2.41.5",
    # YAML/CLI from Pydantic Schema
    "cyclopts==4.10.0",
    "rich==14.3.3",
    # Needed for tokenization and OSL reporting
    "transformers==5.16.1",
    "tiktoken==0.13.0",
    # Required by transformers' apply_chat_template
    "jinja2==3.1.6",
    "numpy>=1.26.4",
    "datasets==5.0.1",
    "Pillow==12.3.0",
    "sentencepiece==0.2.1",
    "protobuf==7.34.1",
    "openai_harmony==0.0.8",
    # HDR Histogram for live percentile/histogram approximations in the
    # metrics aggregator (PyPI: hdrhistogram, importable as hdrh.histogram).
    "hdrhistogram==0.10.3",
    # Color support for cross-platform terminals
    "colorama==0.4.6",
    # Fix pytz-2024 import warning
    "pytz==2026.1.post1",
    "urllib3==2.8.0",
    "pyyaml==6.0.3",
    # anyio is pulled in transitively (openai -> httpx -> anyio); pinned here to
    # force past CVE-2026-63374 / CVE-2026-64847 (both fixed in 4.14.2), which
    # otherwise fail `uv run pip-audit`.
    "anyio==4.14.2",
]

[project.optional-dependencies]
sql = [
    # SQL event logger (swappable backends, default sqlite)
    "sqlalchemy==2.0.48",
]
dev = [
    # Code quality
    "ruff==0.15.8",
    # Development tools
    "pre-commit==4.5.1",
    # Profiling tools
    "line-profiler==5.0.2",
    "Pympler==1.1",
    # Documentation
    "sphinx==9.1.0",
    "sphinx-rtd-theme==3.1.0",
    "sphinx-autodoc-typehints==3.9.11",
    "myst-parser==5.0.0",
    # Security auditing
    "pip-audit==2.10.0",
    # bfcl-eval hard-pins filelock==3.20.0, which uv would otherwise share across
    # every resolution fork. Because bfcl conflicts with this extra (see
    # [tool.uv].conflicts), these floors force uv to fork filelock/virtualenv:
    # patched here, pinned only inside the bfcl fork. Closes CVE-2025-68146 /
    # CVE-2026-22701 (filelock) and CVE-2026-22702 (virtualenv).
    "filelock>=3.20.3",
    "virtualenv==21.7.13",
    # pip is pulled in transitively (pip-audit -> pip-api -> pip); force past
    # PYSEC-2026-3721 (fixed in 26.2), which otherwise fails `uv run pip-audit`.
    "pip==26.2",
]
test = [
    # Includes optional dependencies for full test coverage
    "inference-endpoint[sql]",
    "inference-endpoint[rouge]",
    # Testing framework
    "pytest==9.0.3",
    "pytest-asyncio==1.3.0",
    "pytest-cov==7.1.0",
    "pytest-benchmark==5.2.3",
    "pytest-timeout==2.4.0",
    "pytest-xdist==3.8.0",
    # Test utilities
    "coverage==7.13.4",
    "line-profiler==5.0.2",
    "Pympler==1.1",
    "scipy==1.17.1",
    # HTTP server and client for mock server fixture
    "aiohttp==3.14.3",
    # Plotting for benchmark sweep mode
    "matplotlib==3.10.8",
    # Property-based testing (CLI fuzz)
    "hypothesis==6.151.10",
]
performance = [
    "pytest-benchmark==5.2.3",
    "memory-profiler==0.61.0",
]
rouge = [
    # ROUGE text-generation scoring (eval_method: "rouge"), used by the Llama
    # CNN-DailyMail / OpenOrca accuracy paths. Imported lazily in
    # evaluation/scoring.py, so the base install stays lean. Resolves cleanly
    # against the pinned datasets==5.0.1 - no conflicts fork needed, unlike bfcl.
    # NOTE: not self-contained at runtime - nltk needs the punkt/punkt_tab
    # corpora and evaluate.load("rouge") fetches its metric script from the HF
    # Hub, so an offline image must prefetch both.
    # nltk advisory PYSEC-2026-3740 has no fixed release and is ignored in CI's
    # pip-audit (affected APIs unused); tracked in
    # https://github.com/mlcommons/endpoints/issues/520
    "nltk==3.10.3",
    "evaluate==0.4.6",
    "rouge-score==0.1.2",
]
bfcl = [
    # BFCL v4 function-calling evaluation. Pins numpy==1.26.4, which is why
    # the top-level numpy requirement is a lower bound (>=1.26.4).
    "bfcl-eval==2026.3.23",
    # bfcl-eval's qwen model handler transitively imports qwen_agent → soundfile;
    # soundfile is not used by our scorer but must be present for the import to succeed.
    "soundfile==0.13.1",
    # Used directly by the multi-turn runner (bfcl_v4_multi_turn_runner.py).
    # Present transitively via openai, but declared explicitly so the import
    # does not depend on another package keeping it.
    "httpx==0.28.1",
]

[project.scripts]
inference-endpoint = "inference_endpoint.main:run"

[project.urls]
Homepage = "https://github.com/mlperf/inference-endpoint"
Documentation = "https://github.com/mlperf/inference-endpoint#readme"
Repository = "https://github.com/mlperf/inference-endpoint.git"
Issues = "https://github.com/mlperf/inference-endpoint/issues"

[tool.ruff]
target-version = "py312"
line-length = 88
exclude = [
    "tests/assets/datasets/*",
    "src/inference_endpoint/openai/openai_types_gen.py",
    "src/inference_endpoint/openai/openapi.yaml",
    # Vendored third-party MLCommons evaluator - keep verbatim.
    "src/inference_endpoint/evaluation/legacy_mlperf_deepseek_r1/mlperf_eval/",
    "datasets/*",
    "datasets/",
]


[tool.ruff.lint]
select = [
    "E",  # pycodestyle errors
    "W",  # pycodestyle warnings
    "F",  # pyflakes
    "I",  # isort
    "B",  # flake8-bugbear
    "C4", # flake8-comprehensions
    "UP", # pyupgrade
]
ignore = [
    "E501",  # line too long, handled by ruff-format
    "B008",  # do not perform function calls in argument defaults
    "C901",  # too complex
]

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
skip-magic-trailing-comma = false
line-ending = "auto"

[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py", "*_test.py"]
python_classes = ["Test*"]
python_functions = ["test_*"]
asyncio_mode = "strict"
timeout = 300
timeout_func_only = true
addopts = [
    "--strict-markers",
    "--strict-config",
    "--cov=src",
    "--cov-report=term-missing",
    "--cov-report=html",
    "--cov-report=xml",
    "-m not run_explicitly",
]
markers = [
    "slow: marks tests as slow (skip in CI)",
    "performance: marks tests as performance tests (skip in CI) - no timeout",
    "integration: marks tests as integration tests",
    "unit: marks tests as unit tests",
    "run_explicitly: mark test to only run explicitly",
    "schema_fuzz: hypothesis CLI fuzz tests (run in CI on schema changes)",
]
filterwarnings = [
    "ignore:Session timeout reached:RuntimeWarning",
]

[tool.coverage.run]
source = ["src"]
concurrency = ["multiprocessing", "thread"]
parallel = true
sigterm = true
omit = [
    "*/tests/*",
    "*/test_*",
    "*/__pycache__/*",
    "*/venv/*",
    "*/\\.venv/*",
]

[tool.coverage.report]
exclude_lines = [
    "pragma: no cover",
    "def __repr__",
    "if self.debug:",
    "if settings.DEBUG",
    "raise AssertionError",
    "raise NotImplementedError",
    "if 0:",
    "if __name__ == .__main__.:",
    "class .*\\\\bProtocol\\):",
    "@(abc\\\\.)?abstractmethod",
]

[tool.mypy]
ignore_missing_imports = true
exclude = [
    "openai_types_gen\\.py$",
    "metrics/reporter\\.py$",
    "legacy_mlperf_deepseek_r1/mlperf_eval/",
]

[[tool.mypy.overrides]]
module = [
    "inference_endpoint.openai.openai_types_gen",
    "inference_endpoint.metrics.reporter",
]
ignore_errors = true