pyproject.toml

[project]
name = "libreyolo"
version = "1.7.4"
description = "MIT-licensed computer vision library: one Python API for YOLO, RF-DETR, RT-DETR and dozens more model families, with training and export where a family supports them."
keywords = [
    "yolo",
    "object-detection",
    "computer-vision",
    "pytorch",
    "deep-learning",
    "instance-segmentation",
    "pose-estimation",
    "rf-detr",
    "rt-detr",
    "yolov9",
    "yolox",
    "onnx",
    "mit-license",
]
classifiers = [
    "Development Status :: 5 - Production/Stable",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3 :: Only",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
    "Topic :: Scientific/Engineering :: Image Recognition",
    "Topic :: Software Development :: Libraries :: Python Modules",
]
readme = "README.md"
requires-python = ">=3.10"
license = "MIT"
license-files = [
    "LICENSE",
    "NOTICE",
    "THIRD_PARTY_NOTICES.txt",
    "weights/LICENSE_NOTICE.txt",
    "licenses/Apache-2.0.txt",
]
authors = [
    {name = "LibreYOLO Team"}
]
dependencies = [
    "numpy>=1.19.0",
    "Pillow>=9.1.0",
    "torch>=2.4.0",
    "torchvision>=0.19.0",
    "PyYAML>=6.0",
    "requests>=2.25.0",
    "opencv-python>=4.11.0.86",
    "mss>=9.0.1",
    "tqdm>=4.65.0",
    "pycocotools>=2.0.0",
    "typer>=0.9.0",
    "click>=8.0.0",
    "safetensors>=0.4.0",
    "scipy>=1.7.0",
    # Ephemeral automatic-DDP job transport. Standard pickle validation still
    # rejects lambdas/closures; cloudpickle carries transportable __main__
    # callbacks without importing and re-running the user's training script.
    "cloudpickle>=3.0.0",
    # LibreYOLO("owner/repo") loads from the Hugging Face Hub out of the box,
    # which the Hub's "Use this model" snippet relies on. Only list_repo_files
    # and hf_hub_download are needed, both in 0.x; the bound stays below 1.0
    # so the molmo2 extra (transformers 4.57.1 needs huggingface_hub<1.0)
    # still resolves. The hf extra keeps >=1.0 for gated-download metadata.
    "huggingface_hub>=0.34.0",
]

[dependency-groups]
dev = [
    # Below 9 so the dev group resolves next to the executorch extra:
    # ExecuTorch 1.2 requires pytest<9 at runtime.
    "pytest>=8.3",
    "pytest-timeout>=2.3.1",
    "pytest-xdist>=3.6",
    "ruff>=0.14.10",
    "ty>=0.0.1",
]

[project.optional-dependencies]
marigold = [
    "diffusers==0.38.0",
    "peft==0.18.1",
    "accelerate==1.13.0",
    "transformers==5.4.0",
    "bitsandbytes==0.49.2; platform_system == 'Linux' or platform_system == 'Windows'",
]
stream = [
    # YouTube page URLs are resolved lazily to direct media URLs. Webcam,
    # RTSP/RTMP/TCP/UDP, HLS, and local multi-stream inputs need no extra.
    "yt-dlp",
]
onnx = [
    "onnx>=1.14.0",
    "onnxsim>=0.4.0",
    # Opset-21 runtime floor required by the embedded OpenCV Zoo LaMa graph.
    "onnxruntime>=1.18.0",
]
triton = [
    # HTTP/HTTPS V2 inference only. gRPC and shared-memory extras are excluded.
    "tritonclient[http]>=2.40.0",
]
executorch = [
    # Fixed-shape FP32 .pte export and Python runtime validation. Kept out of
    # aggregate extras because ExecuTorch constrains the compatible Torch pair.
    # Pinned to the validated minor: an unpinned install can resolve
    # ExecuTorch 1.0 with an older Torch instead. ExecuTorch 1.2's compiled
    # runtime is built against torch 2.11 but only asks for torch>=2.11, so
    # pip would pair it with a newer torch whose libc10 the runtime cannot
    # load ("Symbol not found"); pin the pair it was built for.
    # ExecuTorch 1.2 has no Python 3.14 wheels. The marker skips the stack
    # there so pip installs the current LibreYOLO without it instead of
    # backtracking to an old release; the export raises a clear error.
    "executorch>=1.2,<1.3; python_version < '3.14'",
    "torch>=2.11,<2.12; python_version < '3.14'",
    "torchvision>=0.26,<0.27; python_version < '3.14'",
]
coreai = [
    # Core AI (.aimodel) export, macOS only: the toolchain neither converts
    # nor runs elsewhere. Kept out of every aggregate extra on purpose,
    # because coreai-torch pins torch to 2.11.x and would drag the whole
    # environment onto that version. coreai-torch requires Python 3.11+.
    "coreai-torch>=0.4.1; sys_platform == 'darwin' and python_version >= '3.11'",
]
rfdetr = [
    "transformers>=5.1.0",
]
eomt = [
    "transformers>=5.1.0",
]
midas = [
    # MiDaS's official ViT-L/16 and EfficientNet-Lite3 encoders. The bounded
    # 1.0 series preserves the checkpoint keys and bitwise parity validated by
    # the native port without relying on MiDaS's historical Torch Hub fetch.
    "timm>=1.0.28,<1.1",
]
lora = [
    # LoRA / PEFT fine-tuning (lora=True) for transformer backbones (RF-DETR).
    "libreyolo[rfdetr]",
    "peft>=0.12.0",
]
plots = [
    "matplotlib>=3.7.0",
]
fast-eval = [
    # C++ COCO eval backend. Preferred by default (faster_coco_eval=True) and
    # transparently skipped when absent, so this stays an extra rather than a
    # hard dependency: it ships compiled wheels, and a source build on a
    # platform without one must not be able to break a plain `pip install`.
    "faster-coco-eval>=1.7.0",
]
vla = [
    # LibreVLA loads vision-language-action policies through the Apache-2.0
    # lerobot package (policy classes, processor pipelines, LeRobot datasets).
    # lerobot requires Python >= 3.12, so the extra installs nothing on older
    # interpreters and the tier raises an ImportError that says so. Kept out
    # of `all` for the same reason.
    "lerobot[smolvla,diffusion,dataset]>=0.6.1; python_version >= '3.12'",
]
gen = [
    # LibreGen loads generative image pipelines (FLUX.2 klein, SANA-Sprint)
    # through the Apache-2.0 diffusers package; text encoders come from
    # transformers and accelerate backs low-memory loading and offload=True.
    # Speridlabs Idol reads DINOv3 through transformers and the FLUX.2 VAE
    # through diffusers.
    "diffusers>=0.38.0",
    "transformers>=5.1.0",
    "accelerate>=1.7.0",
]
llm = [
    # LibreLLM talks to OpenAI-compatible Responses / Chat Completions
    # endpoints through the official Apache-2.0 OpenAI Python SDK.
    "openai>=1.66.0",
]
ground = [
    # LibreGround (instruction → click) shares the VLM extra stack.
    "libreyolo[vlm]",
]
molmo2 = [
    # Pinned Molmo2 remote code uses Transformers 4 processor and RoPE APIs.
    "transformers==4.57.1",
    "einops>=0.8.0",
    "accelerate>=0.26.0",
]
vlm = [
    # VLM-as-detector and LibreGround families load through transformers.
    # huggingface_hub (used for snapshot download) ships with transformers.
    "transformers>=5.1.0",
    # SmolVLM2's processor requires num2words.
    "num2words>=0.5.0",
    # LocateAnything's HF remote-code processor/model imports these at load time.
    # decord 0.6.0 has no sdist and wheels only for Linux x86_64 and Windows
    # x64, so LocateAnything is unavailable elsewhere; the rest of vlm installs.
    "decord==0.6.0; (sys_platform == 'linux' and platform_machine == 'x86_64') or (sys_platform == 'win32' and platform_machine == 'AMD64')",
    "lmdb>=1.7.0",
    "peft>=0.12.0",
]
vlm-train = [
    # VLM detection fine-tuning: the vlm inference stack plus PEFT for LoRA.
    # peft>=0.17 is the floor the LoRA-adapter save/load contract was
    # validated against.
    "libreyolo[vlm]",
    "peft>=0.17.0",
]
sam = [
    # Promptable-segmentation families load through transformers;
    # EdgeTAM's RepViT backbone uses Transformers' TimmWrapper.
    "transformers>=5.3.0",
    "timm>=1.0.23",
]
hub-kernels = [
    # Optional runtime loader for compiled Hub kernels (e.g. the
    # ms_deform_attn slot). Installing this extra is the opt-in; disable
    # with LIBREYOLO_HUB_KERNELS=0. See docs/kernels.md.
    "kernels>=0.6.0",
]
openvocab = [
    # Discriminative open-vocabulary detectors load through transformers
    # (GroundingDinoForObjectDetection / Owlv2ForObjectDetection /
    # OmDetTurboForObjectDetection). OMDet-Turbo's configured Swin backbone
    # is instantiated through transformers' TimmBackbone.
    "transformers>=5.1.0",
    "timm>=1.0.0",
    # OV-DEIM embeds text prompts with the vendored CLIP BPE tokenizer, which
    # needs these at predict time (Unicode \p{L}/\p{N} classes + ftfy fixing).
    "regex>=2023.0.0",
    "ftfy>=6.0.0",
]
sensenova = [
    # SenseNova-Vision (Bagel-MoT): the vendored architecture loads through
    # accelerate's big-model dispatch; the tokenizer and activation registry
    # come from transformers. flash-attn is optional (SDPA fallback built in).
    "transformers>=5.1.0",
    "accelerate>=1.7.0",
    # 4-bit NF4 loading on consumer GPUs; no macOS wheels.
    "bitsandbytes>=0.45.0; sys_platform != 'darwin'",
]
modus = [
    # LibreMODUS uses the vendored Bagel/Qwen2/SigLIP architecture and
    # accelerate big-model dispatch. No remote code or feature VQ tokenizer
    # package is loaded.
    "transformers>=5.1.0",
    "accelerate>=1.7.0",
]
label = [
    # AI-assist for `libreyolo label`: SAM click-to-mask via the LibreSAM tier.
    # The manual labeller and model-assisted box auto-label work without this.
    "libreyolo[sam]",
]
tensorrt = [
    # No CUDA on macOS; marker drops these. Version pin must match the
    # [[tool.uv.dependency-metadata]] blocks below. No aarch64 wheels on PyPI:
    # Jetson gets TensorRT from JetPack. TensorRT 10.16 publishes no Python
    # 3.14 bindings; without the version bound pip backtracks to an old
    # LibreYOLO release on 3.14 instead of skipping TensorRT.
    "tensorrt-cu12==10.16.1.11; sys_platform != 'darwin' and platform_machine != 'aarch64' and python_version < '3.14'",
    # No pycuda: it is sdist only and needs the CUDA toolkit headers to build,
    # which broke pip installs. Export and inference use torch.cuda. INT8
    # calibration imports cuda.bindings, which PyPI torch already installs on
    # Linux; elsewhere it asks for `pip install cuda-python`. Pinning
    # cuda-python here would fight torch's own cuda-bindings major version.
]
openvino = [
    "openvino>=2024.0.0",
]
paddle = [
    "libreyolo[onnx]",
    # paddlepaddle 2.6.2 publishes wheels up to Python 3.12 only. The markers
    # skip the Paddle stack on 3.13+ so pip installs the current LibreYOLO
    # without it instead of backtracking to an old release.
    # X2Paddle 1.6.0 imports onnx.mapping, removed after ONNX 1.17.
    "onnx>=1.14.0,<1.18; python_version < '3.13'",
    # X2Paddle 1.6.0 generates a valid YOLO9 static graph with Paddle 2.6.2.
    # Paddle 3.x rejects the generated mixed scalar arithmetic during capture.
    "paddlepaddle==2.6.2; python_version < '3.13'",
    "x2paddle==1.6.0; python_version < '3.13'",
    # Imported by X2Paddle but omitted from its wheel metadata.
    "six>=1.16.0; python_version < '3.13'",
]
mnn = [
    "libreyolo[onnx]",
    "MNN>=3.6.1",
]
ncnn = [
    "pnnx",
    "ncnn",
]
tflite = [
    "libreyolo[onnx]",
    # onnx2tf 2.4.3 includes the RF-DETR conversion fixes required by this path.
    # Requires Python 3.12+ (onnx2tf 2.4.x wheel constraint). Every onnx2tf
    # release pins ai-edge-litert==2.1.2 and numpy==2.2.6 (or 1.26.4), which
    # publish no Python 3.14 wheels, so the converter is skipped there:
    # without the upper bound pip backtracks to an old LibreYOLO release.
    "onnx2tf>=2.4.3; python_version >= '3.12' and python_version < '3.14'",
    # Runtime used by TFLiteBackend to execute exported .tflite artifacts.
    # Without onnx2tf's pin it resolves to a release with 3.14 wheels.
    "ai-edge-litert>=2.1.2; python_version >= '3.12'",
    # Required for the RF-DETR GridSample ONNX graph rewrite (fallback path).
    "onnx-graphsurgeon>=0.3.27",
    # The RF-DETR onnxsim preprocessing step imports onnxsim, which the onnx
    # extra above provides. Not onnx-simplifier: that is an sdist-only
    # duplicate of the same module, compiled from C++ and overwriting onnxsim.
]
# LiteRT is Google's current name for TensorFlow Lite; same export path,
# same .tflite output. Separate extra so both names install.
litert = [
    "libreyolo[tflite]",
]
# Intentionally empty: all tracking dependencies (e.g. scipy) are core deps,
# so `libreyolo[tracking]` installs nothing extra. Kept for a stable extra name.
tracking = []
rtdetr = []
gaze = [
    # Enables best-effort auto-download of the L2CS Gaze360 checkpoint from
    # the official Google Drive (large-file confirm-token handling). Without
    # it, LibreL2CS prints manual download instructions instead.
    "gdown>=5.1.0",
]
clip = [
    # LibreCLIP towers are native torch (no open_clip at runtime). Only the
    # vendored BPE text tokenizer needs these to reproduce CLIP's exact token
    # IDs (Unicode \p{L}/\p{N} classes + ftfy text fixing).
    "regex>=2023.0.0",
    "ftfy>=6.0.0",
]
clip-convert = [
    # Convert-/parity-only: re-host OpenCLIP LAION-2B weights as LibreCLIP
    # checkpoints (weights/convert_clip_weights.py) and run the parity tests.
    "libreyolo[clip]",
    "open_clip_torch>=2.24.0",
]
siglip2 = [
    # LibreSigLIP2 towers are native torch (no transformers at runtime). Only the
    # multilingual SentencePiece text tokenizer (Gemma vocab) needs this.
    "sentencepiece>=0.1.99",
]
siglip2-convert = [
    # Convert-/parity-only: re-host Apache-2.0 google/siglip2-* weights as
    # LibreSigLIP2 checkpoints (weights/convert_siglip2_weights.py) and run the
    # parity tests against the reference SiglipModel.
    "libreyolo[siglip2]",
    "transformers>=5.1.0",
]
coreml = [
    # coremltools ships wheels for macOS and Linux x86_64 only (no Windows
    # support); CoreML models run on macOS. No release publishes Python 3.14
    # wheels yet, and its C++ source build fails without the right toolchain.
    "coremltools>=8.0; python_version < '3.14' and (sys_platform == 'darwin' or (sys_platform == 'linux' and platform_machine == 'x86_64'))",
]
fiftyone = [
    # Dataset curation and prediction analysis (Apache-2.0). Deliberately kept
    # out of `all`: fiftyone depends on opencv-python-headless while LibreYOLO
    # depends on opencv-python, and both ship the same `cv2` module, so
    # installing them together leaves whichever landed last in place. See
    # docs/fiftyone.md.
    "fiftyone>=1.0.0",
    # graphql-core 3.3.0 (2026-09-27) removed graphql.execution.ExecutionContext,
    # which strawberry-graphql imports, so `import fiftyone` fails with it.
    "graphql-core<3.3",
]
tensorboard = [
    "tensorboard>=2.13.0",
]
mlflow = [
    "mlflow>=2.9.0",
]
wandb = [
    "wandb>=0.16.0",
]
comet = [
    "comet-ml>=3.47.1",
]
clearml = [
    "clearml>=1.17.0",
]
neptune = [
    "neptune-scale>=0.30.0",
]
dvclive = [
    "dvclive>=3.49.0",
]
dvc = [
    "libreyolo[dvclive]",
]
hf = [
    # Hugging Face Hub loading (LibreYOLO("owner/repo")) and push_to_hub.
    # v1 adds dry-run byte metadata used to bound gated model downloads.
    "huggingface_hub>=1.0.0",
]
all = [
    "libreyolo[stream]",
    "libreyolo[onnx]",
    "libreyolo[triton]",
    "libreyolo[rfdetr]",
    "libreyolo[eomt]",
    "libreyolo[midas]",
    "libreyolo[lora]",
    "libreyolo[plots]",
    "libreyolo[tensorrt]",
    "libreyolo[tflite]",
    "libreyolo[openvino]",
    # paddle is left out: X2Paddle needs onnx<1.18 while onnx2tf (tflite)
    # pins onnx 1.20.1 on Python 3.12+, and paddlepaddle 2.6.2 has no Python
    # 3.13+ wheels. Install libreyolo[paddle] on its own.
    "libreyolo[mnn]",
    "libreyolo[ncnn]",
    "libreyolo[tracking]",
    "libreyolo[rtdetr]",
    "libreyolo[gaze]",
    "libreyolo[clip]",
    "libreyolo[siglip2]",
    "libreyolo[coreml]",
    "libreyolo[vlm]",
    "libreyolo[llm]",
    "libreyolo[gen]",
    "libreyolo[sam]",
    "libreyolo[label]",
    "libreyolo[openvocab]",
    "libreyolo[sensenova]",
    "libreyolo[modus]",
    "libreyolo[tensorboard]",
    "libreyolo[mlflow]",
    "libreyolo[wandb]",
    "libreyolo[comet]",
    "libreyolo[clearml]",
    "libreyolo[dvclive]",
    "libreyolo[hf]",
    # Keep Neptune separate: stable neptune-scale requires protobuf<7 while
    # the TFLite/onnx2tf extra included above requires protobuf 7.
]

[project.scripts]
libreyolo = "libreyolo.cli:entrypoint"

[project.urls]
Homepage = "https://www.libreyolo.com"
Documentation = "https://www.libreyolo.com/docs"
Repository = "https://github.com/LibreYOLO/libreyolo"
Issues = "https://github.com/LibreYOLO/libreyolo/issues"
Changelog = "https://www.libreyolo.com/docs/changelog"

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

[tool.setuptools.package-data]
libreyolo = [
    "assets/*.jpg",
    "config/**/*.yaml",
    "training/*.yaml",
    "data/*.json",
    "data/*.txt",
    "models/clip/*.gz",
    "models/clip/data/*.json",
    "models/siglip2/*.model",
    "models/siglip2/data/*.json",
    "models/darknet/cfgs/*.cfg",
    "models/yolo7/*.yaml",
    "models/**/NOTICE",
    "models/**/LICENSE",
    "models/**/*.md",
]

[tool.setuptools.packages.find]
include = ["libreyolo*"]

[build-system]
requires = ["setuptools>=77", "wheel"]
build-backend = "setuptools.build_meta"

# tensorrt-cu12 is sdist-only and its setup.py refuses to run on macOS, so
# uv's universal lock can't extract its metadata on darwin hosts. Static
# metadata lets uv resolve the non-darwin fork without building. Bump in
# lockstep with the version pin in [project.optional-dependencies].tensorrt.
# Pull torch/torchvision from the PyTorch CUDA 12.8 index on Linux/Windows so
# `uv sync` produces a GPU-enabled environment by default. macOS keeps the
# PyPI default (CPU/MPS wheel) since download.pytorch.org has no darwin builds.
# Requires NVIDIA driver >= 555 (CUDA 12.8 runtime) on the host.
[tool.uv]
conflicts = [
    [{ extra = "molmo2" }, { extra = "hf" }],
    [{ extra = "molmo2" }, { extra = "all" }],
    [{ extra = "molmo2" }, { extra = "eomt" }],
    [{ extra = "molmo2" }, { extra = "gen" }],
    [{ extra = "molmo2" }, { extra = "ground" }],
    [{ extra = "molmo2" }, { extra = "label" }],
    [{ extra = "molmo2" }, { extra = "lora" }],
    [{ extra = "molmo2" }, { extra = "marigold" }],
    [{ extra = "molmo2" }, { extra = "modus" }],
    [{ extra = "molmo2" }, { extra = "openvocab" }],
    [{ extra = "molmo2" }, { extra = "rfdetr" }],
    [{ extra = "molmo2" }, { extra = "sam" }],
    [{ extra = "molmo2" }, { extra = "sensenova" }],
    [{ extra = "molmo2" }, { extra = "siglip2-convert" }],
    [{ extra = "molmo2" }, { extra = "vlm" }],
    [{ extra = "molmo2" }, { extra = "vlm-train" }],
    # lerobot (vla) needs huggingface-hub>=1.6; transformers 4.57.1 needs <1.0.
    [{ extra = "molmo2" }, { extra = "vla" }],

    [{ extra = "neptune" }, { extra = "tflite" }],
    [{ extra = "neptune" }, { extra = "litert" }],
    [{ extra = "neptune" }, { extra = "all" }],

    # lerobot (vla) needs accelerate>=1.14; marigold pins accelerate 1.13.0.
    [{ extra = "marigold" }, { extra = "vla" }],

    # X2Paddle needs onnx<1.18; onnx2tf (tflite, in all) pins onnx 1.20.1 and
    # protobuf 7, which paddlepaddle 2.6.2 rejects on Windows.
    [{ extra = "paddle" }, { extra = "tflite" }],
    [{ extra = "paddle" }, { extra = "litert" }],
    [{ extra = "paddle" }, { extra = "all" }],

    # ExecuTorch 1.2 requires pytest<9 at runtime; onnx2tf (tflite, in all)
    # pins pytest 9.0.2. ExecuTorch 1.3+ drops the pin but needs torch 2.12,
    # which the cu128 index does not ship.
    [{ extra = "executorch" }, { extra = "tflite" }],
    [{ extra = "executorch" }, { extra = "litert" }],
    [{ extra = "executorch" }, { extra = "all" }],
]

[tool.uv.sources]
torch = [
    { index = "pytorch-cu128", marker = "sys_platform != 'darwin'" },
]
torchvision = [
    { index = "pytorch-cu128", marker = "sys_platform != 'darwin'" },
]

[[tool.uv.index]]
name = "pytorch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true

[[tool.uv.dependency-metadata]]
name = "tensorrt-cu12"
version = "10.16.1.11"
requires-dist = [
    "tensorrt-cu12-libs==10.16.1.11",
    "tensorrt-cu12-bindings==10.16.1.11",
]
requires-python = ">=3.8"

[[tool.uv.dependency-metadata]]
name = "tensorrt-cu12-libs"
version = "10.16.1.11"
requires-dist = []

[[tool.uv.dependency-metadata]]
name = "tensorrt-cu12-bindings"
version = "10.16.1.11"
requires-dist = []

[tool.ruff]
# Vendored from upstream DEIMv2. Keep local lint focused on the LibreYOLO
# wrapper/integration code instead of rewriting third-party implementation files.
extend-exclude = [
    "libreyolo/models/deimv2/engine",
    "libreyolo/models/depth_anything3/_vendor",
]

[tool.ty.rules]
# Rules that catch real bugs — keep as errors
invalid-parameter-default = "error"

# Rules that produce false positives with PyTorch nn.Module dynamic attrs
# and C extensions without type stubs (tensorrt, ncnn, pycuda).
# Downgrade to warnings so they don't block CI but remain visible.
unresolved-import = "warn"       # optional deps not installed in dev env
unresolved-attribute = "warn"    # nn.Module.__getattr__, C extension members
invalid-argument-type = "warn"   # torch.onnx.export kwargs, tensor ops
not-subscriptable = "warn"       # tensor[int] operations
unsupported-operator = "warn"    # tensor arithmetic
not-iterable = "warn"            # Module/tensor iterations
call-non-callable = "warn"       # Module.__call__
invalid-method-override = "warn" # template method + nn.Module patterns
invalid-assignment = "warn"      # nn.Module dynamic attrs, tensor ops
invalid-return-type = "warn"     # torch return type inference
index-out-of-bounds = "warn"     # tensor indexing
unknown-argument = "warn"        # dynamic kwargs patterns
no-matching-overload = "warn"    # torch overloaded functions
possibly-missing-attribute = "warn"

[[tool.ty.overrides]]
include = ["vision_analysis_benchmark/**", "tests/**"]

[tool.ty.overrides.rules]
# Benchmark and test code: suppress all to avoid noise
unresolved-import = "ignore"

[tool.pytest.ini_options]
addopts = "-m unit"
testpaths = ["tests"]
# Only the thread method can interrupt a test wedged inside a native call, such
# as a stalled Hugging Face download; the signal method cannot fire until
# control returns to Python. The limit itself comes from PYTEST_TIMEOUT, which
# the e2e target sets, so unit runs stay unbounded.
timeout_method = "thread"
markers = [
    "marigold_v2: Marigold V2 dense prediction tests",
    "detany3d: optional DetAny3D promptable 3D detection integration",
    "fcos3d: native FCOS3D calibrated 3D detection",
    "wilddet3d: optional WildDet3D 3D detection integration",
    "mood3d: optional 3D-MOOD open-set 3D detection integration",
    "unit: fast tests that avoid external weights or large files",
    "vit: tests covering the ViT classification family",
    "swin: tests covering the Swin classification family",
    "vjepa2: tests covering the V-JEPA 2 video embedding and classification family",
    "levjepa: tests covering the LeVJEPA video embedding family",
    "external_data: tests that require external datasets, weights, or staged large files",
    "network: tests that intentionally use non-local network access",
    "e2e: end-to-end tests requiring full model loading and inference",
    "general_nightly: curated native inference checks in the default nightly",
    "flagship_nightly: heavier nightly checks for flagship YOLO9/RF-DETR models",
    "cuda_graph: opt-in tests requiring CUDA graph capture and replay",
    "extended_training: opt-in multi-epoch or non-core task training tests",
    "training_nightly: opt-in GPU checks for training-time CUDA graph capture",
    "export_backend: tests covering an export backend or serialized runtime format",
    "supported_backend: tests covering a release-blocking supported backend",
    "extended_backend: tests covering an extended backend",
    "onnx: tests covering ONNX export or inference",
    "triton: tests covering NVIDIA Triton HTTP inference",
    "torchscript: tests covering TorchScript export or inference",
    "executorch: tests requiring the optional ExecuTorch exporter/runtime",
    "smoke: fast end-to-end pipeline checks on tiny synthetic data (CPU, <60s)",
    "tensorrt: tests requiring TensorRT",
    "trt: alias for TensorRT tests requiring TensorRT",
    "openvino: tests requiring OpenVINO",
    "mnn: tests requiring MNN",
    "ncnn: tests requiring ncnn",
    "tflite: tests requiring TensorFlow Lite export",
    "paddle: tests requiring PaddlePaddle export/runtime",
    "coreml: tests requiring CoreML (macOS only)",
    "fiftyone: tests requiring the optional FiftyOne integration (fiftyone starts a local MongoDB)",
    "yolox: tests covering the YOLOX model family",
    "yolo7: tests covering the YOLOv7 model family",
    "yolo9: tests covering the YOLO9 model family",
    "yolo9_e2e: tests covering the YOLOv9-E2E model family",
    "yolonas: tests covering the YOLO-NAS model family",
    "ppyoloe: tests covering the PP-YOLOE model family",
    "rfdetr: tests covering the RF-DETR model family",
    "detr: tests covering the original DETR model family",
    "deit: tests covering the DeiT image-classification family",
    "convnextv2: tests covering the ConvNeXt V2 classification family",
    "deeplabv3: tests covering the DeepLabv3 semantic segmentation family",
    "lwdetr: tests covering the LW-DETR model family",
    "faster_rcnn: tests covering the Faster R-CNN model family",
    "retinanet: tests covering the RetinaNet model family",
    "ssd: tests covering the SSD model family",
    "mask_rcnn: tests covering the Mask R-CNN model family",
    "fcn: tests covering the FCN semantic-segmentation family",
    "ppliteseg: tests covering the PP-LiteSeg semantic-segmentation family",
    "unet: tests covering the U-Net semantic-segmentation family",
    "fcos: tests covering the FCOS model family",
    "deformable_detr: tests covering the Deformable DETR model family",
    "dinodetr: tests covering the DINO-DETR model family",
    "centernet: tests covering the CenterNet model family",
    "dfine: tests covering the D-FINE model family",
    "domedetr: tests covering the Dome-DETR tiny-object model family",
    "deim: tests covering the DEIM model family",
    "deimv2: tests covering the DEIMv2 model family",
    "tinyformer: tests covering the TinyFormer model family",
    "ec: tests covering the ECDet model family",
    "gtr: tests covering the GTR model family",
    "rtdetr: tests covering the RT-DETR model family",
    "rtdetrv2: tests covering the RT-DETRv2 model family",
    "rtdetrv4: tests covering the RT-DETRv4 model family",
    "picodet: tests covering the PICODET model family",
    "efficientdet: tests covering the EfficientDet model family",
    "rtmdet: tests covering the RTMDet model family",
    "l2cs: tests covering the L2CS gaze estimation family",
    "hrnet: tests covering the HRNet top-down pose family",
    "dekr: tests covering the DEKR bottom-up pose family",
    "fomo: tests covering the LibreFOMO point-localizer family",
    "alexnet: tests covering the LibreAlexNet image-classification family",
    "vgg: tests covering the LibreVGG classification family",
    "vlm: tests covering the LibreVLM (VLM-as-detector) tier",
    "llm: tests covering the LibreLLM OpenAI-compatible client",
    "vla: tests covering the LibreVLA (vision-language-action) tier",
    "gen: tests covering the LibreGen (generative image) tier",
    "sam: tests covering the LibreSAM (promptable-segmentation) tier",
    "openvocab: tests covering the LibreOpenVocab detector tier",
    "sensenova: tests covering the LibreSenseNovaVision unified multimodal family",
    "modus: tests covering the external-weight LibreMODUS any-to-any family",
    "clip: tests covering the LibreCLIP zero-shot open-vocabulary classifier family",
    "siglip2: tests covering the LibreSigLIP2 zero-shot open-vocabulary classifier family",
    "pe: tests covering the LibrePE Perception Encoder Core zero-shot / embedding family",
    "midas: tests covering the MiDaS relative-depth family",
    "quicksrnet: tests covering the QuickSRNet super-resolution family",
    "ddcolor: tests covering the DDColor automatic-colorization family",
    "hvi_cidnet: tests covering the HVI-CIDNet low-light restoration family",
    "lama: tests covering the LaMa mask-guided inpainting family",
    "vitmatte: tests covering the ViTMatte trimap-guided matting family",
    "distributed: multi-rank gloo tests that mp.spawn a process group. Linux-only in the PR gate: they exercise distributed-training math, which is platform-independent, while dist.init_process_group(backend='gloo') costs ~105s per test on the macOS runner (a bare 2-rank scalar all-reduce takes 109s there vs 3.9s on Linux). They also compare a parent-process reference gradient against child-process gradients at a BLAS-sensitive tolerance, so running them on every OS yields flakes rather than coverage. The spawn-plumbing tests are deliberately NOT marked: they cover spawn-only failure modes that Linux's fork never hits.",
    "rf1: RF1 training tests",
    "rf5: RF5 training benchmark tests",
    "slow: slow tests that may take several minutes",
]