pyproject.toml

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

[project]
name = "onellm"
dynamic = ["version"]
description = "A unified interface for interacting with large language models from various providers - a complete drop-in replacement for OpenAI's client"
readme = "README.md"
requires-python = ">=3.10"
license = {text = "Apache-2.0"}
authors = [
    {name = "Ran Aroussi", email = "ran@aroussi.com"}
]
dependencies = [
    "requests>=2.32.5",
    # httpx replaces aiohttp as the async HTTP client (see CHANGELOG
    # 0.20260427.0). The [http2] extra pulls in `h2`, which we negotiate
    # via TLS ALPN inside ``onellm/http_pool.py``.
    "httpx[http2]>=0.27.0",
    "pydantic>=2.12.5",
    "PyYAML>=6.0.3",
    "openai>=2.15.0",
    "tiktoken>=0.12.0",
    "huggingface-hub>=1.3.4",
    # Transitive pins — minimum safe versions for known CVEs
    "urllib3>=2.6.3",
    "filelock>=3.20.3",
    "pyasn1>=0.6.2",
]
classifiers = [
    "Development Status :: 5 - Production/Stable",
    "Intended Audience :: Developers",
    "License :: OSI Approved :: Apache Software License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Topic :: Software Development :: Libraries :: Python Modules",
]

[project.urls]
"Homepage" = "https://github.com/muxi-ai/onellm"
"Documentation" = "https://docs.muxi.io"
"Issues" = "https://github.com/muxi-ai/onellm/issues"

[project.optional-dependencies]
# Lean local-embedding backend: onnxruntime (CPU) + transformers tokenizer.
# Does NOT pull torch. Repos that ship only PyTorch weights need the
# [local-pytorch] extra in addition. This is a HARD BREAK from the v0.20260415
# shape of [cache] which pulled sentence-transformers; see CHANGELOG.
cache = [
    "onnxruntime>=1.20.0",
    "transformers>=4.44.0",
    "numpy>=1.26.0",
    "faiss-cpu>=1.13.2",
]
# CUDA-accelerated end-to-end GPU stack for local embeddings:
#   - onnxruntime-gpu: CUDAExecutionProvider auto-selected at session build
#   - faiss-gpu-cu12:  GPU vector index (community-maintained PyPI wheel;
#                      the official FAISS project ships GPU via conda only)
# Replaces the earlier half-baked [local-gpu] shape that paired GPU ONNX
# with CPU FAISS. Installs the same transformers/numpy baseline as [cache].
# CUDA-only: Apple Silicon should use [cache] (CoreML EP auto-activates);
# AMD/Intel GPUs and DirectML are not supported today (no matching FAISS
# GPU wheel). See CHANGELOG 0.20260422.2.
local-cuda = [
    "onnxruntime-gpu>=1.20.0",
    "transformers>=4.44.0",
    "numpy>=1.26.0",
    "faiss-gpu-cu12>=1.13.0",
]
# Deprecated alias kept for back-compat. Resolves to [local-cuda] via the
# transitive-extra mechanism, so pinning onellm[local-gpu] still works and
# now gets the faiss-gpu-cu12 upgrade transparently. Scheduled for removal
# in a future major release; please migrate to onellm[local-cuda].
local-gpu = [
    "onellm[local-cuda]",
]
# Fallback backend for repos that don't ship ONNX weights. Pulls torch +
# sentence-transformers (~1 GB install). Use only when you're pinned to a
# PyTorch-only repo; prefer ONNX-ready repos (e.g. nomic-ai/*) otherwise.
local-pytorch = [
    "sentence-transformers>=5.2.2",
]
all = [
    "anthropic>=0.76.0",
    "google-generativeai>=0.8.6",
    "boto3>=1.42.35",
    "llama-cpp-python>=0.3.16",
    "google-auth>=2.48.0",
    "google-cloud-aiplatform>=1.134.0",
    # Default to the CPU onnxruntime + faiss stack for [all] so it stays
    # installable on every platform (including macOS, which has no CUDA).
    # Callers who want GPU should use onellm[local-cuda] explicitly.
    "onnxruntime>=1.20.0",
    "transformers>=4.44.0",
    "sentence-transformers>=5.2.2",
    "numpy>=1.26.0",
    "faiss-cpu>=1.13.2",
]
bedrock = [
    "boto3>=1.42.35",
]
vertexai = [
    "google-auth>=2.48.0",
    "google-cloud-aiplatform>=1.134.0",
]
llama = [
    "llama-cpp-python>=0.3.16",
]
dev = [
    "pytest>=9.0.2",
    "pytest-asyncio>=1.3.0",
    "pytest-cov>=7.0.0",
    "black>=26.1.0",
    "isort>=7.0.0",
    "mypy>=1.19.1",
    "ruff>=0.14.14",
    "python-dotenv>=1.2.1",
]

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

[tool.setuptools]
include-package-data = true

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

[tool.setuptools.dynamic]
version = {file = ["onellm/.version"]}

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

[tool.ruff.lint]
select = ["E", "F", "N", "W", "C90", "I", "B", "UP", "A"]
ignore = [
    "E501",   # Line too long (handled by black)
    "A002",   # Argument shadowing builtin (e.g. `input`) — OpenAI API compat
    "C901",   # Function too complex — tracked for incremental cleanup
    "E402",   # Module-level import not at top — integration test path setup
    "B904",   # raise ... from err — tracked for incremental cleanup
]

[tool.pytest.ini_options]
asyncio_mode = "auto"
testpaths = ["tests"]
markers = [
    "slow: marks tests as slow (real model downloads, cold starts); deselect with '-m \"not slow\"'",
    "integration: marks tests that require real external services or API keys",
]

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

[tool.isort]
profile = "black"
line_length = 100