pyproject.toml

[build-system]
requires = ["setuptools>=68"]
build-backend = "setuptools.build_meta"

[project]
name = "hunyuan-ocr"
version = "0.1.3"
description = "Evaluation-backed AMD ROCm port of HunyuanOCR-1.5 for OmniDocBench v1.6"
readme = "README.md"
requires-python = ">=3.11"
# Mixed license: see NOTICE and LICENSES/. Original packaging/tooling is Apache-2.0;
# code ported from HunyuanOCR and the model weights are under the Tencent Hunyuan
# Community License (not OSI open source; excludes EU/UK/KR).
license = { text = "Mixed: Apache-2.0 (original packaging/tooling) AND Tencent Hunyuan Community License (code ported from HunyuanOCR + model weights). See NOTICE and LICENSES/." }
authors = [{ name = "AIwork4me" }]
keywords = [
  "hunyuanocr", "ocr", "document-ai", "amd-gpu", "rocm", "rdna3",
  "llama-cpp", "vllm", "omnidocbench", "vision-language-model",
]
classifiers = [
  "Development Status :: 4 - Beta",
  "Intended Audience :: Science/Research",
  "Operating System :: POSIX :: Linux",
  "Programming Language :: Python :: 3",
  "Programming Language :: Python :: 3.11",
  "Programming Language :: Python :: 3.12",
  "Programming Language :: Python :: 3.13",
  "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
# Core deps are intentionally GPU-free: importing the package, validating,
# scoring-wrapper config, and the CLI all work on a plain CPU. The heavy / GPU
# deps are opt-in extras. Plain-PyPI torch is NEVER a dependency — on ROCm you
# must install torch from a verified ROCm wheel source separately (see README).
dependencies = [
  "pillow",
  "pyyaml",
  "tqdm",
  "requests",
]

[project.optional-dependencies]
# OpenAI-compatible client (vLLM / llama-server / any OAI server).
client = ["openai>=1.0"]
# Hugging Face model/artifact downloader (provides the `hf` CLI used in Quick Start).
download = ["huggingface_hub>=0.25"]
# transformers backend (Phase 1 oracle). Pinned to the benchmark version for
# reproducibility; see reproducibility.lock.yaml. Requires a separate ROCm torch.
transformers = ["transformers==5.13.0"]
# Convenience alias; pyyaml is already core. Reserved for future eval-only deps.
eval = ["pyyaml"]
# OmniDocBench-ROCm platform integration (conformance, score, publish).
platform = ["omnidocbench-rocm>=0.2.0"]
# Development / CI tooling.
dev = ["pytest>=8", "pytest-mock", "ruff>=0.16,<0.17", "build", "pytest-cov>=4"]

[project.urls]
Homepage = "https://github.com/AIwork4me/HunyuanOCR-ROCm"
Repository = "https://github.com/AIwork4me/HunyuanOCR-ROCm"
Issues = "https://github.com/AIwork4me/HunyuanOCR-ROCm/issues"

[project.scripts]
hunyuan-ocr = "hunyuan_ocr.cli:main"

[tool.setuptools.packages.find]
where = ["src"]

[tool.setuptools.package-data]
hunyuan_ocr = ["data/*.yaml"]

[tool.ruff]
line-length = 120
target-version = "py311"
extend-exclude = ["REPRO.yaml", "reproduce.md"]

[tool.ruff.lint]
# The eval pipeline is defensive by design: per-page failures are caught and the
# run continues (never raises — see the adapter contract). Broad blind-except
# (BLE001) and silent try/except/pass (S110) are intentional, not bugs.
ignore = ["BLE001", "S110"]

[tool.ruff.lint.per-file-ignores]
# Vendored / verbatim upstream-derived algorithm files: lint only, never reformat,
# so diffs against Tencent HunyuanOCR stay trackable. The multi-statement (E701/
# E702) and import-layout (E401) styles and the placeholder f-string (F541) are
# preserved verbatim from upstream on purpose.
"src/hunyuan_ocr/postprocess.py" = ["E501", "E741", "E701", "E702", "E401", "F541"]
"src/hunyuan_ocr/tasks.py" = ["E501"]
"src/hunyuan_ocr/contract.py" = ["E501"]

[tool.ruff.format]
# per-file-ignores is lint-only; the formatter must exclude the same vendored
# files explicitly so `ruff format` never rewrites them.
exclude = [
  "src/hunyuan_ocr/postprocess.py",
  "src/hunyuan_ocr/tasks.py",
  "src/hunyuan_ocr/contract.py",
]

[tool.pytest.ini_options]
testpaths = ["tests"]
markers = [
  "gpu: end-to-end test requiring a real gfx1100 ROCm GPU + model weights (deselected in CI)",
]