pyproject.toml

[build-system]
requires = [
    "setuptools>=61",
    "wheel",
    "setuptools-scm[toml]>=8",
    "pybind11>=2.12",
    "cmake>=3.16",
]
build-backend = "setuptools.build_meta"

[project]
name = "vllm-afd-plugin"
dynamic = ["version"]
description = "vLLM external plugin for Attention-FFN Disaggregation (AFD)"
readme = "README.md"
requires-python = ">=3.10,<3.14"
license = "Apache-2.0"
license-files = ["LICENSE"]
authors = [{ name = "AFD plugin contributors" }]
keywords = ["vllm", "afd", "attention", "ffn", "disaggregation", "inference"]
classifiers = [
    "Development Status :: 2 - Pre-Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "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 = []

# Keep vLLM optional so macOS and CPU-only development can run import/config tests.
# Install the runtime extra only in environments with a compatible vLLM wheel.
[project.optional-dependencies]
vllm = ["vllm==0.26.0"]

[project.urls]
Homepage = "https://github.com/vllm-project/afd-plugin"
Repository = "https://github.com/vllm-project/afd-plugin"
Documentation = "https://github.com/vllm-project/afd-plugin#readme"
"Bug Tracker" = "https://github.com/vllm-project/afd-plugin/issues"

[project.entry-points."vllm.general_plugins"]
afd = "afd_plugin:register_afd"

[dependency-groups]
dev = [
    "datasets>=3.0",
    "httpx>=0.27",
    "huggingface_hub>=0.26",
    "pre-commit>=4",
    "pytest>=7.0",
    "ruff>=0.15.7,<0.16",
    "setuptools-scm>=8",
]
e2e-tests = [
    "lm_eval[api]==0.4.12",
]

[tool.setuptools_scm]

[tool.setuptools.packages.find]
where = ["."]
include = ["afd_plugin*"]
exclude = ["afd_plugin._cann_ops_custom*"]
namespaces = false

[tool.setuptools.package-data]
afd_plugin = ["py.typed", "_cann_ops_custom/**/*"]

[tool.pytest.ini_options]
testpaths = ["tests"]
pythonpath = ["."]
addopts = "-q"
markers = [
    "vllm_runtime: tests requiring an importable vLLM runtime dependency.",
    "gpu: GPU-gated AFD integration tests.",
    "npu: NPU-gated AFD integration tests.",
    "e2e: full end-to-end tests requiring model weights and hardware.",
    "eval: accuracy/evaluation tests requiring datasets.",
    "slow: tests that take more than 120 seconds.",
]

[tool.ruff]
line-length = 88
target-version = "py310"
extend-exclude = [".venv", "venv", "build", "dist"]

[tool.ruff.lint]
select = [
    "E",
    "F",
    "I",
    "N",
    "UP",
    "W",
    "B",
    "G",
    "ISC",
    "SIM",
]

[tool.ruff.lint.per-file-ignores]
"afd_plugin/compat/patches/**/*.py" = [
    # Patch functions and copied upstream code intentionally keep upstream names.
    "E501",
    "N806",
    "N807",
]

[tool.ruff.format]
quote-style = "double"
indent-style = "space"
docstring-code-format = true

[tool.mypy]
python_version = "3.10"
# Silent follow for library code; missing third-party stubs are ignored
# everywhere (including tests). tests/ still get --follow-imports skip from
# tools/pre_commit/mypy.py.
ignore_missing_imports = true
check_untyped_defs = true
follow_imports = "silent"

[tool.ty]

[tool.ty.analysis]
allowed-unresolved-imports = ["torch.**", "vllm.**"]

[tool.typos.default]
extend-ignore-words-re = [
    # Ascend domain terms checked as plain words in docs/comments.
    "(?i)^cann$",
    "(?i)^tbe$",
]
extend-ignore-identifiers-re = [
    ".*_thw",
    ".*thw",
    "ein",
    ".*arange",
    ".*MoBA",
    ".*temperal_downsample",
    ".*nd",
    ".*[Nn][Dd]$",
    ".*_ND_.*",
    ".*NDExpectShape.*",
    ".*EmbedND$",
    ".*nothink.*",
    ".*NOTHINK.*",
    ".*nin.*",
    ".*[Oo]no.*[Aa]nna.*",
    ".*cann.*",
    "alog",
    ".*alog.*",
    ".*[Nn][Dd]3.*",
    ".*updator.*",
    "byt5.*",
    # NemotronVoiceChat vendored triton kernel: dout/DOUT abbreviate d_out
    # (output feature dim), not "doubt".
    ".*dout.*",
    ".*DOUT.*",
    # Vendored CANN/AscendC identifiers under csrc/npu/ascend_kernels/ whose
    # camelCase segments spell a word the dictionary flags. These are part of
    # the CANN API surface (renaming them breaks compilation against the CANN
    # headers) or upstream spellings kept as-is so the vendored tree stays
    # diffable against its origin. The `nd`/`ND` entries above only match a
    # trailing or all-lowercase `nd`, not a mid-identifier camelCase segment.
    # ND->NZ copies and their stride/param names.
    ".*[Nn]d2[Nn]z.*",
    ".*[Nn]d[A-Z].*",
    # ND_ALIGN, ND_WEIGHT_DIM_LIMIT, ND_WEIGHT_MULTI_TENSOR_DIM, MIN_ND_DIM.
    ".*ND_WEIGHT.*",
    ".*MIN_ND_DIM.*",
    ".*ND_ALIGN.*",
    # CANN registration macro.
    ".*REGIST_MATMUL_OBJ.*",
    # Upstream-spelled local names: preGoupList, reformatedTensor,
    # reformatedWeight, isFisrtBlock, copyedL1Left, worskspaceOffset,
    # DataCopyPad2DA8W4NDNew, isND2NZ, enVecND2NZ.
    ".*[Gg]oup.*",
    ".*reformated.*",
    ".*Fisrt.*",
    ".*copyed.*",
    ".*worskspace.*",
    "DataCopyPad2DA8W4NDNew",
    ".*isND2NZ.*",
    ".*enVecND2NZ.*",
]