pyproject.toml

[build-system]
requires = [
    "setuptools>=77,<82",
    "setuptools-scm>=8",
]
build-backend = "build_backend"
backend-path = ["."]

[project]
name = "xinference"
description = "Model Serving Made Easy"
readme = "README.md"
authors = [
    { name = "Chris Qin", email = "chris.qin@xinference.co" },
]
maintainers = [
    { name = "Chris Qin", email = "chris.qin@xinference.co" },
]
license = "Apache-2.0"
license-files = ["LICENSE"]
requires-python = ">=3.10"
classifiers = [
    "Operating System :: OS Independent",
    "Programming Language :: Python",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: Implementation :: CPython",
    "Topic :: Software Development :: Libraries",
]
dynamic = ["version"]
dependencies = [
    "xoscar>=0.9.10",
    "psutil>=5.9.0",
    "torch",
    "pillow",
    "click<8.2.0",
    "tqdm>=4.27",
    "tabulate",
    "requests",
    "aiohttp",
    "pydantic",
    "fastapi>=0.110.3,<0.137",
    "python-multipart>=0.0.18",  # required by FastAPI's Form/File/UploadFile params; used to come in transitively via gradio
    "orjson",  # used by core.utils.json_dumps for all REST API JSON responses; used to come in transitively via gradio
    "uvicorn",
    "PyYAML",
    "huggingface-hub>=0.19.4",
    "filelock>=3.30.0",  # preserve_lock_file keeps Wan MLX conversion locks stable on Windows
    "typing_extensions",
    "modelscope>=1.21.0",
    "sse_starlette>=1.6.5",  # ensure_bytes API break change: https://github.com/sysid/sse-starlette/issues/65
    "openai>=1.40.0",  # For typing
    "anthropic>=0.25.0",  # For Anthropic protocol typing & route registration
    "python-jose[cryptography]",
    "bcrypt>=4.0.0",
    "aioprometheus[starlette]>=23.12.0",
    "httpx",
    "nvidia-ml-py",
    "pynvml>=12",
    "async-timeout",
    # peft 0.17.1 and below import HybridCache at module level, which was
    # removed in transformers 5.0.0. peft declares an unbounded transformers
    # requirement, so pinning it low silently breaks any transformers 5.x
    # install (see the sentence-transformers rerank import chain).
    "peft>=0.18.0",
    "timm",
    "setproctitle",
    "uv",
]

[project.urls]
Homepage = "https://github.com/xorbitsai/inference"

[project.scripts]
xinference = "xinference.deploy.cmdline:cli"
xinference-local = "xinference.deploy.cmdline:local"
xinference-supervisor = "xinference.deploy.cmdline:supervisor"
xinference-worker = "xinference.deploy.cmdline:worker"
xinference-router = "xinference.deploy.router:main"
xinference-router-agent = "xinference.deploy.router_agent:main"
xinference-migrate-auth = "xinference.api.oauth2.advanced.migrate:main"
xinference-reset-auth-password = "xinference.api.oauth2.advanced.reset_password:main"

[project.entry-points."vllm.general_plugins"]
xinference_navidc_ocr = "xinference.thirdparty.navidc_ocr:register"

[project.optional-dependencies]
dev = [
    "cython>=0.29",
    "pytest>=3.5.0",
    "pytest-cov>=2.5.0",
    "pytest-timeout>=1.2.0",
    "pytest-forked>=1.0",
    "pytest-asyncio>=0.14.0",
    "pytest-mock>=3.11.1",
    "ipython>=6.5.0",
    "sphinx>=3.0.0",
    "pydata-sphinx-theme>=0.3.0",
    "sphinx-intl>=0.9.9",
    "jieba>=0.42.0",
    "ruff",
    "black",
    "openai>=1.40.0",
    "anthropic",
    "langchain",
    "langchain-community",
    "langchain-openai",
    "sphinx-tabs",
    "sphinx-design",
    "pypdfium2",  # For PDF OCR tests
    "tomli; python_version < '3.11'",
]
otel = [
    "opentelemetry-api>=1.20.0",
    "opentelemetry-sdk>=1.20.0",
    "opentelemetry-exporter-otlp-proto-http>=1.20.0",
    "opentelemetry-exporter-otlp-proto-grpc>=1.20.0",
    "opentelemetry-instrumentation-fastapi>=0.41b0",
    "opentelemetry-instrumentation-httpx>=0.41b0",
]
router = [
    "tokenizers>=0.21.0,<0.23.0",
]
all = [
    "xinference[anthropic]",
    "xinference[llama_cpp]",
    "xinference[transformers]",
    "xinference[vllm]",
    "xinference[mlx]",
    "xinference[embedding]",
    "xinference[rerank]",
    "xinference[image]",
    "xinference[video]",
    "xinference[audio]",
    "xinference[router]",
]
intel = [
    "torch==2.1.0a0",
    "intel_extension_for_pytorch==2.1.10+xpu",
]
llama_cpp = [
    "xllamacpp>=0.2.0",
]
transformers = [
    "transformers>=4.53.3",
    "torch",
    "accelerate>=0.28.0",
    "sentencepiece",
    "transformers_stream_generator",
    "bitsandbytes ; sys_platform=='linux'",
    "protobuf",
    "einops",
    "tiktoken",
    "optimum",
    "attrdict",  # For deepseek VL
    "timm>=0.9.16",  # For deepseek VL
    "torchvision",  # For deepseek VL
    "peft",
    "eva-decord",  # For video in VL
    "jj-pytorchvideo",  # For CogVLM2-video
    "qwen-vl-utils!=0.0.9",  # For qwen2-vl
    "qwen_omni_utils",  # For qwen2.5-omni
    "audioread",  # Undeclared runtime dependency of qwen_omni_utils
    "datamodel_code_generator",  # for minicpm-4B
    "jsonschema",  # for minicpm-4B
    "blobfile",  # for moonlight-16b-a3b
]
transformers_quantization = [
    "bitsandbytes ; sys_platform=='linux'",
    "gptqmodel",
    "autoawq!=0.2.6 ; sys_platform!='darwin'",  # autoawq 0.2.6 pinned torch to 2.3
    "datasets>=3.4.0",  # see GH#3934
]
vllm = [
    "vllm>=0.2.6 ; sys_platform=='linux'",
    "xxhash",
]
sglang = [
    "sglang[srt]>=0.4.2.post4 ; sys_platform=='linux'",
]
mlx = [
    "mlx-lm>=0.21.5 ; sys_platform=='darwin' and platform_machine=='arm64'",
    "mlx-vlm>=0.3.4 ; sys_platform=='darwin' and platform_machine=='arm64'",
    "mlx-whisper ; sys_platform=='darwin' and platform_machine=='arm64'",
    "f5-tts-mlx ; sys_platform=='darwin' and platform_machine=='arm64'",
    "mlx-audio ; sys_platform=='darwin' and platform_machine=='arm64'",
    "qwen_vl_utils!=0.0.9",
    "tomli",
]
embedding = [
    "sentence-transformers>=3.1.0,<6",
    "FlagEmbedding",
    "datasets>=3.4.0",  # see GH#3934
]
rerank = [
    "FlagEmbedding",
    "datasets>=3.4.0",  # see GH#3934
]
image = [
    "diffusers>=0.32.0",  # fix conflict with matcha-tts
    "controlnet_aux",
    "deepcache",
    "verovio>=4.3.1",  # For got_ocr2
    "transformers>=4.53.3",  # For got_ocr2
    "tiktoken>=0.6.0",  # For got_ocr2
    "accelerate>=0.28.0",  # For got_ocr2
    "torch",  # For got_ocr2
    "torchvision",  # For got_ocr2
    "gguf",  # For flux and sd3.5 series
    "pypdfium2",  # For PDF input on the OCR API
]
video = [
    "diffusers>=0.32.0",
    "imageio-ffmpeg",
]
audio = [
    "funasr==1.2.7",
    "omegaconf~=2.3.0",
    "nemo_text_processing<=1.1.0; sys_platform == 'linux'",  # 1.1.0 requires pynini==2.1.6.post1
    "WeText",
    "librosa",
    "xxhash",
    "torch>=2.0.0",  # for CosyVoice, matcha
    "torchaudio>=2.0.0",
    "ChatTTS>=0.2.1",
    "tiktoken",  # For CosyVoice, openai-whisper
    "lightning>=2.0.0",  # For CosyVoice, matcha
    "hydra-core>=1.3.2",  # For CosyVoice, matcha
    "inflect",  # For CosyVoice, matcha
    "conformer",  # For CosyVoice, matcha
    "diffusers>=0.32.0",  # For CosyVoice, matcha
    "gdown",  # For CosyVoice, matcha
    "pyarrow",  # For CosyVoice, matcha
    "HyperPyYAML",  # For CosyVoice
    "onnxruntime>=1.16.0",  # For CosyVoice, use onnxruntime-gpu==1.16.0 if possible
    "pyworld>=0.3.4",  # For CosyVoice
    "loguru",  # For Fish Speech
    "natsort",  # For Fish Speech
    "loralib",  # For Fish Speech
    "ormsgpack",  # For Fish Speech
    "cachetools",  # For Fish Speech
    "silero-vad",  # For Fish Speech
    "vector-quantize-pytorch<=1.17.3,>=1.14.24",  # For Fish Speech
    "torchdiffeq",  # For F5-TTS
    "x_transformers>=1.31.14",  # For F5-TTS
    "pypinyin",  # For F5-TTS
    "tomli",  # For F5-TTS
    "vocos",  # For F5-TTS
    "librosa",  # For F5-TTS
    "jieba",  # For F5-TTS
    "pydub",  # For F5-TTS & MegaTTS3; used to come in transitively via gradio
    "soundfile",  # For F5-TTS & MeloTTS & Kokoro
    "cached_path",  # For MeloTTS
    "unidic-lite",  # For MeloTTS, unidic requires manually download
    "cn2an",  # For MeloTTS
    "mecab-python3",  # For MeloTTS
    "num2words",  # For MeloTTS
    "pykakasi",  # For MeloTTS
    "fugashi",  # For MeloTTS
    "g2p_en",  # For MeloTTS
    "anyascii",  # For MeloTTS
    "gruut[de,es,fr]",  # For MeloTTS
    "kokoro>=0.7.15",  # Kokoro, already included misaki[en]
    "misaki[en,zh]>=0.7.15",  # Kokoro
    "langdetect",  # MegaTTS3
    "pyloudnorm",  # MegaTTS3
    "json5",  # IndexTTS2
    "munch",  # IndexTTS2
    "matplotlib",  # IndexTTS2
    "flatten_dict",  # IndexTTS2
    "julius",  # IndexTTS2
    "tensorboard",  # IndexTTS2
    "randomname",  # IndexTTS2
    "argbind",  # IndexTTS2
    "ffmpy",  # IndexTTS2 (audiotools' core.ffmpeg imports this unconditionally); used to come in transitively via gradio
]
doc = [
    "ipython>=6.5.0",
    "sphinx>=3.0.0",
    "pydata-sphinx-theme>=0.3.0",
    "sphinx-intl>=0.9.9",
    "sphinx-tabs",
    "sphinx-design",
    "prometheus_client",
    "timm",
]
musa = [
    "mthreads-ml-py>=2.2.8",
    "torchada>=0.1.11",
]
benchmark = [
    "psutil",
]
anthropic = [
    "anthropic",  # no-op: anthropic is now a core dependency, kept for backward compatibility
]

[tool.setuptools]
zip-safe = false
include-package-data = true

[tool.setuptools.packages.find]
# Plain packages only; namespace discovery would sweep in top-level
# directories such as doc/, frontend/ and benchmark/.
namespaces = false
include = ["xinference*"]
exclude = [
    "*.conftest*",
    "*.tests.*",
    "*.tests",
]

[tool.setuptools_scm]
version_file = "xinference/_version.py"
# must stay a numeric three-part release: bdist_wheel fails on values
# like "0+unknown" when building without any git metadata (e.g. the
# Dockerfile's dependency-skeleton layer)
fallback_version = "0.0.0+unknown"

[tool.black]
include = '\.pyi?$'
extend-exclude = '''
/(
| ^/xinference/(_version.py)
| thirdparty
)/
'''

[tool.pytest.ini_options]
asyncio_mode = "auto"

[tool.coverage.run]
branch = true
relative_files = true
cover_pylib = false
plugins = ["Cython.Coverage"]
include = ["xinference/*"]
omit = [
    "xinference/_version.py",
    "*.pxd",
    "*/tests/*",
]

[tool.coverage.report]
exclude_lines = [
    "pragma: no cover",
    "def __repr__",
    "raise AssertionError",
    "raise NotImplementedError",
    "return NotImplemented",
]

[tool.isort]
profile = "black"
skip = ["thirdparty"]

[tool.mypy]
ignore_missing_imports = true
follow_imports = "skip"
exclude = "thirdparty"

[tool.codespell]
ignore-words-list = "hist,rcall,fpr,ser,nd,inout,ot,Ba,ba,asend,hart,coo,splitted,datas,fro,te"
skip = ".idea,.git,./build,./docs/build,node_modules,static,generated,*.po,*.ts,*.json,*.c,*.cpp,*.cfg,thirdparty"

[tool.ruff]
line-length = 100
# Mirrors the old flake8 `exclude` list from setup.cfg.
extend-exclude = [
    "__init__.py",
    "xinference/thirdparty",
    "xinference/_version.py",
    ".github",
    "build",
    "ci",
    "dist",
    "docs",
]

[tool.ruff.lint]
preview = true
# Equivalent of the flake8 `select` list previously in setup.cfg.
# E901/E999 are omitted (ruff always reports syntax errors); E127 and the
# long-deprecated W601-W604 checks have no ruff counterpart.
select = [
    "E101",
    "E111",
    "E117",
    "E201",
    "E202",
    "E223",
    "E224",
    "E225",
    "E231",
    "E242",
    "E251",
    "E273",
    "E274",
    "E275",
    "E301",
    "E302",
    "E303",
    "E304",
    "E305",
    "E401",
    "E703",
    "E902",
    "F401",
    "F63",
    "F7",
    "F811",
    "F821",
    "F822",
    "F823",
    "F841",
    "W191",
    "W291",
    "W292",
    "W293",
    "W391",
    "W605",
]

[tool.ruff.lint.per-file-ignores]
# Sphinx injects the `tags` global when executing conf.py.
"doc/source/conf.py" = ["F821"]