[build-system]
requires = ["setuptools>=61.0", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "rapid-mlx"
version = "0.6.49"
description = "Rapid-MLX — AI inference for Apple Silicon. Drop-in OpenAI API, 2-4x faster than Ollama."
readme = "README.md"
license = {text = "Apache-2.0"}
requires-python = ">=3.10"
authors = [
    {name = "vllm-mlx contributors"}
]
keywords = ["llm", "mlx", "apple-silicon", "vllm", "inference", "transformers"]
classifiers = [
    "Development Status :: 4 - Beta",
    "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 = [
    # Core — these are all you need for `rapid-mlx serve <text-model>`
    "mlx>=0.29.0",
    "mlx-lm>=0.31.0",  # 0.31+ required for ArraysCache native batching (hybrid models)
    # mlx-vlm is opt-in via [vision] extras (saves ~322 MB for text-only users)
    "transformers>=5.0.0",  # mlx-lm 0.30.5+ requires transformers 5.0 (rc3 bug fixed in stable)
    "tokenizers>=0.19.0",
    "huggingface-hub>=0.23.0",
    "numpy>=1.24.0",
    "pillow>=10.0.0",
    "tqdm>=4.66.0",
    "pyyaml>=6.0",
    "requests>=2.28.0",
    "tabulate>=0.9.0",
    "psutil>=5.9.0",
    "fastapi>=0.100.0",
    "uvicorn>=0.23.0",
    "mcp>=1.0.0",
    "jsonschema>=4.0.0",
]

[project.optional-dependencies]
# Vision/multimodal models (Gemma 4, Qwen-VL, etc.) — adds ~322 MB
# Required for any model with vision input. Text-only models work without this.
# 0.5.0+ also unlocks DFlash speculative decoding (see [dflash] extras).
vision = [
    "mlx-vlm>=0.5.0",  # 0.5.0: gemma4 multi-image + tool-parser fixes, TurboQuant race-condition fix, continuous-batching guard. Also unlocks DFlash spec-decode hooks (see [dflash] extras).
    "opencv-python>=4.8.0",
    "torch>=2.3.0",
    "torchvision>=0.18.0",
]
# DFlash speculative decoding for Qwen3.5/3.6 dense 8-bit models
# (issue #264). Adds ~1-4 GB at runtime (drafter weights) and depends on
# mlx-vlm's spec-decode runtime. Text-only — does NOT pull torch/cv2 like
# the vision extras do. Only enable if you serve a DFlash-eligible alias
# (e.g. qwen3.5-27b-8bit) with --enable-dflash.
dflash = [
    "mlx-vlm>=0.5.0",
]
# Embedding endpoint
embeddings = [
    "mlx-embeddings>=0.0.5",
]
# Gradio chat UI
chat = [
    "gradio>=4.0.0",
    "pytz>=2024.1",
]
# All extras — union of vision + chat + embeddings.
# Expanded directly to avoid a self-dependency that breaks
# pip install .[all] and editable installs.
all = [
    # vision
    "mlx-vlm>=0.5.0",
    "opencv-python>=4.8.0",
    "torch>=2.3.0",
    "torchvision>=0.18.0",
    # chat
    "gradio>=4.0.0",
    "pytz>=2024.1",
    # embeddings
    "mlx-embeddings>=0.0.5",
    # dflash (text-only path covered by mlx-vlm above; listed for clarity)
]
dev = [
    "pytest>=7.0.0",
    "pytest-asyncio>=0.21.0",
    "black>=23.0.0",
    "ruff>=0.1.0",
    "mypy>=1.0.0",
]
vllm = [
    "vllm>=0.4.0",
]
# Guided decoding with outlines for structured JSON output
guided = [
    "outlines[mlxlm]>=1.0.0",
]
# Audio dependencies for TTS/STT (mlx-audio)
audio = [
    "mlx-audio>=0.2.9",
    "sounddevice>=0.4.0",
    "soundfile>=0.12.0",
    "scipy>=1.10.0",
    "numba>=0.57.0",
    "tiktoken>=0.5.0",
    "misaki[zh,ja]>=0.5.0",  # Chinese (zh) and Japanese (ja) support
    "spacy>=3.7.0",
    "num2words>=0.5.0",
    "loguru>=0.7.0",
    "phonemizer>=3.2.0",
    # Additional multilingual dependencies
    "ordered_set>=4.1.0",    # Required for Chinese TTS
    "cn2an>=0.5.0",          # Chinese number conversion
    "fugashi>=1.3.0",        # Japanese tokenizer
    "unidic-lite>=1.0.0",    # Japanese dictionary for fugashi
    "jieba>=0.42.0",         # Chinese word segmentation
]

[project.urls]
Homepage = "https://github.com/raullenchai/Rapid-MLX"
Documentation = "https://github.com/raullenchai/Rapid-MLX#readme"
Repository = "https://github.com/raullenchai/Rapid-MLX"

[project.entry-points."vllm.platform_plugins"]
mlx = "vllm_mlx.plugin:mlx_platform_plugin"

[project.scripts]
rapid-mlx = "vllm_mlx.cli:main"
vllm-mlx = "vllm_mlx.cli:main"
vllm-mlx-chat = "vllm_mlx.gradio_app:main"
vllm-mlx-bench = "vllm_mlx.benchmark:main"

[tool.setuptools.package-data]
vllm_mlx = [
    "aliases.json",
    "agents/profiles/*.yaml",
    # Integration tests bundled inside the package so `rapid-mlx agents
    # <name> --test` works on pip/brew installs (the source layout keeps
    # them at tests/integrations/, mirrored here via symlinks). Without
    # this, the test runner shows a confusing path-error skip.
    "_integration_tests/*.py",
    "_integration_tests/*.sh",
]

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

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

[tool.ruff]
line-length = 88

[tool.ruff.format]
# Vendored upstream files — keep formatting identical to the source PR
# so future syncs produce a clean diff.
exclude = ["vllm_mlx/models/deepseek_v4.py"]

[tool.ruff.lint]
select = ["E", "F", "W", "I", "N", "UP", "B", "SIM"]
ignore = [
    "E402",   # module-level import not at top
    "E501",   # line too long
    "E731",   # lambda assignment
    "F811",   # redefined unused
    "F841",   # unused variable
    "B004",   # unreliable callable check
    "B007",   # unused loop variable
    "B008",   # function call in default arg
    "B011",   # assert false
    "B017",   # assert raises exception
    "B905",   # zip without strict
    "N801",   # class name casing
    "N806",   # variable in function casing
    "SIM103", # needless bool
    "SIM108", # ternary instead of if-else
    "SIM110", # reimplemented builtin
    "SIM222", # expr or true
    "UP028",  # yield in for loop
    "SIM102", # collapsible if
    "SIM105", # suppressible exception
    "SIM115", # open file with context handler
    "B904",   # raise without from inside except
]

[tool.ruff.lint.per-file-ignores]
# Vendored from ml-explore/mlx-lm PR #1192 (Blaizzy). Keep as-is so
# upstream syncs remain a clean diff. Drop this once mlx-lm 0.32+
# ships native deepseek_v4 support and we delete the vendored copy.
"vllm_mlx/models/deepseek_v4.py" = ["UP", "B", "SIM", "N", "F", "I", "E"]
# Tensor-shape parameter names (B, H, N, D) follow ML literature
# convention; N803 lowercase rule is incompatible.
"scripts/bench_attention.py" = ["N803"]

[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"