pyproject.toml

[build-system]
requires = [
    "setuptools>=61.0",
    "wheel",
    "cmake>=3.27",
    "nanobind",
    "mlx==0.31.2",
]
build-backend = "setuptools.build_meta"

[project]
name = "omlx"
dynamic = ["version"]
description = "LLM inference server, optimized for your Mac"
readme = "README.md"
license = {text = "Apache-2.0"}
requires-python = ">=3.11"
authors = [
    {name = "omlx contributors"}
]
keywords = ["llm", "mlx", "apple-silicon", "vllm", "inference", "transformers"]
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: Apache Software License",
    "Operating System :: MacOS",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]

dependencies = [
    # Vanilla MLX v0.31.2; optional GLM-5.2 kernels live under
    # omlx.custom_kernels.glm_moe_dsa when built for release.
    "mlx==0.31.2",
    # mlx-lm from commit (2c008fd, v0.31.3) - trust_remote_code load gate,
    # transformers 5.7.0 compatibility, and prior BatchGenerator/cache fixes.
    "mlx-lm @ git+https://github.com/ml-explore/mlx-lm@2c008fd0252b2c569227d12568356ab88ab0560a",
    # regex for mlx-lm's Gemma 4 tool parser (uses recursive patterns)
    "regex",
    # mlx-embeddings from latest commit (32981fa)
    "mlx-embeddings @ git+https://github.com/Blaizzy/mlx-embeddings@32981fa4e8064ed664b52071789dd18271fe4206",
    # mlx-vlm custom processors bypass HF AutoProcessor, so torch is not required
    "transformers>=5.7.0",
    # transformers 5.x's tokenization_mistral_common module imports
    # ReasoningEffort from mistral_common (added in 1.10). The audio extras
    # transitively pull mistral-common via mlx-audio[stt], which historically
    # resolved to 1.9.x and broke WhisperProcessor.from_pretrained() with a
    # silent "Processor not found" on every Whisper model. Pinning the
    # minimum here ensures resolution lands on a compatible version even
    # without the audio extra installed.
    # See pmarreck/omlx#1 for full repro and root cause.
    "mistral-common>=1.10",
    "tokenizers>=0.19.0",
    "huggingface-hub>=0.23.0",
    "numpy>=1.24.0,<2.4",
    "tqdm>=4.66.0",
    "pyyaml>=6.0",
    "itsdangerous>=2.0",
    "jinja2>=3.0",
    # mlx-lm 0.31.3 imports rich from mlx_lm.cli_ui for chat/lora entrypoints.
    "rich>=13.0.0",
    "sentencepiece",
    "tiktoken",
    "protobuf",
    "requests>=2.28.0",
    # SOCKS proxy support (used by httpx via huggingface-hub)
    "socksio>=1.0.0",
    "tabulate>=0.9.0",
    # Resource monitoring
    "psutil>=5.9.0",
    # Native process title support (shows as omlx-server instead of python)
    "setproctitle>=1.3.3",
    # Server
    "fastapi>=0.108.0",
    "uvicorn>=0.23.0",
    # python-multipart is needed by audio_routes.py (File/Form uploads), which
    # server.py unconditionally imports regardless of the [audio] extra.
    "python-multipart>=0.0.5",
    # JSON Schema validation for structured output
    "jsonschema>=4.0.0",
    # Harmony format parser for gpt-oss models
    "openai-harmony",
    # Cohere Command/Cohere2 output parser for reasoning and tool calls
    "cohere_melody>=0.9.0",
    # mlx-vlm from commit e390667. MiniMax M3 support is vendored under
    # omlx.patches.mlx_vlm_minimax_m3_compat because upstream removed it.
    "mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm@e3906673d54df42cf573b837a76e65d58a2b865c",
    "Pillow>=9.0.0",
    # dflash-mlx v0.1.10 (9ca0028) - bstnxbt repo. Apple G17 NAX verify,
    # prefix snapshot metrics, CopySpec mode, and full-context draft-layer cache checks.
    "dflash-mlx @ git+https://github.com/bstnxbt/dflash-mlx@9ca002898b48e14c9727dec17299f497e8467870",
    "markitdown[pdf,docx,pptx]==0.1.6",
]

[project.optional-dependencies]
grammar = [
    # Grammar-constrained decoding for structured output (requires torch ~2GB)
    "xgrammar>=0.1.32",
]
mcp = [
    "mcp>=1.5.0",
]
modelscope = [
    "modelscope>=1.10.0",
]
audio = [
    # mlx-audio from commit (5175326) with tts/stt/sts extras
    # tts extra: misaki, num2words, spacy, phonemizer-fork, espeakng-loader, sentencepiece
    # stt extra: tiktoken, mistral-common[audio]
    # sts extra: tts deps + stt deps + webrtcvad
    "mlx-audio[tts,stt,sts] @ git+https://github.com/Blaizzy/mlx-audio@51753266e0a4f766fd5e6fbc46652224efc23981",
]
paroquant = [
    # ParoQuant runtime loader for paroquant-quantized models. Installed
    # with --no-deps semantics here because the official [mlx] extra pulls
    # torchvision, which the load path doesn't actually use (verified end
    # to end on 0.1.14). The DMG build installs the same pin via build.py.
    "paroquant==0.1.15; python_version >= '3.11'",
]
# Internal extra consumed by packaging/build.py to populate the
# venvstacks mlx-base layer's requirements list. Listed here as the
# single source of truth so the layer-template venvstacks.toml stays
# version-free. End-users don't `pip install omlx[bundle]`; this is for
# the macOS .app build pipeline.
#
# mlx-audio[tts,stt,sts]'s 12 transitive deps appear explicitly because
# venvstacks 0.7.0's internal `uv pip compile` runs with --no-config and
# doesn't propagate pyproject.toml's `tool.uv.override-dependencies`,
# so it can't resolve mlx-audio's `mlx-lm==0.31.1` pin against our
# git-pinned mlx-lm (v0.31.3). build.py installs mlx-audio + paroquant
# separately via --no-deps post-resolution.
bundle = [
    "mcp>=1.5.0",
    "modelscope>=1.10.0",
    # mlx-audio[tts,stt,sts] transitives:
    "scipy>=1.11.0",
    "librosa>=0.10.0",
    "miniaudio>=1.59",
    "numba>=0.59.0",
    "pyloudnorm>=0.1.0",
    "sounddevice>=0.4.6",
    "misaki>=0.9.4",
    "num2words>=0.5.14",
    "spacy>=3.8.4",
    "phonemizer-fork>=3.3.2",
    "espeakng-loader>=0.2.4",
    "webrtcvad>=2.0.10",
    # [audio] extra on top of pyproject's plain mistral-common>=1.10
    "mistral-common[audio]>=1.10",
]
dev = [
    "pytest>=7.0.0",
    "pytest-asyncio>=0.21.0",
    "black>=23.0.0",
    "ruff>=0.1.0",
    "mypy>=1.0.0",
    "mcp>=1.5.0",
    "venvstacks>=0.7.0",
    # The torch-stub smoke test (tests/test_torch_stub.py) gates
    # xgrammar / tvm-ffi version bumps. It skips when these aren't
    # importable, which silently hides regressions — install no-deps
    # in the dev environment so the test actually runs.
    "xgrammar==0.2.0",
    "apache-tvm-ffi==0.1.11",
]
# PEP 735 dependency groups — consumed by `uv sync --dev`.
# Keep in sync with [project.optional-dependencies] dev above
# (pip/legacy toolchains use that section instead).
[dependency-groups]
dev = [
    "pytest>=7.0.0",
    "pytest-asyncio>=0.21.0",
    "black>=23.0.0",
    "ruff>=0.1.0",
    "mypy>=1.0.0",
    "mcp>=1.5.0",
    "venvstacks>=0.7.0",
    "xgrammar==0.2.0",
    "apache-tvm-ffi==0.1.11",
]

[project.urls]
Homepage = "https://github.com/jundot/omlx"
Documentation = "https://github.com/jundot/omlx#readme"
Repository = "https://github.com/jundot/omlx"

[project.scripts]
omlx = "omlx.cli:main"

[tool.setuptools.dynamic]
version = {attr = "omlx._version.__version__"}

[tool.setuptools.packages.find]
where = ["."]
include = ["omlx*"]

[tool.setuptools.package-data]
"omlx" = ["oq_calibration_data.json"]
"omlx.admin" = ["templates/**/*.html", "static/**/*", "i18n/*.json"]
"omlx.eval" = ["data/*.jsonl"]
"omlx.custom_kernels.glm_moe_dsa" = ["*.metallib", "*.dylib", "*.so"]

[tool.uv]
# mlx and mlx-lm are git-pinned; override transitive pins
# (e.g. mlx-audio -> mlx-lm==0.31.1) so the resolver accepts them.
override-dependencies = [
    "mlx==0.31.2",
    "mlx-lm @ git+https://github.com/ml-explore/mlx-lm@2c008fd0252b2c569227d12568356ab88ab0560a",
]

[tool.black]
line-length = 88
target-version = ["py310", "py311", "py312", "py313"]

[tool.ruff]
line-length = 88
select = ["E", "F", "W", "I", "N", "UP", "B", "SIM"]
ignore = ["E501", "B905"]

[tool.mypy]
python_version = "3.10"
warn_return_any = true
warn_unused_configs = true
ignore_missing_imports = true

[tool.pytest.ini_options]
testpaths = ["tests"]
python_files = ["test_*.py"]
asyncio_mode = "auto"