[project]
name = "document-language-model"
version = "0.10.0"
description = "Directive-driven local LLM training, retraining, and export from .dlm documents, codebases, and multimodal sources."
readme = "README.md"
requires-python = ">=3.11"
license = { text = "MIT" }
authors = [{ name = "espadonne", email = "mfwolffe@outlook.com" }]
keywords = ["llm", "lora", "fine-tuning", "ollama", "local-ai"]
classifiers = [
    "Development Status :: 2 - Pre-Alpha",
    "Intended Audience :: Developers",
    "License :: OSI Approved :: MIT License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
    # CLI / doc / lightweight plumbing.
    "typer>=0.12",
    "rich>=13.7",
    "prompt-toolkit>=3.0",    # dlm repl interactive prompt (Sprint 24)
    "watchfiles>=0.24",       # dlm train --watch save-to-train loop (Sprint 25)
    "python-ulid>=3.0",       # dlm_id generation (Sprint 13)
    "pydantic>=2.9",          # schema validation (Sprint 03+)
    "pyyaml>=6.0",            # .dlm frontmatter (Sprint 03+)
    "psutil>=6.0",            # hardware doctor (Sprint 05)
    "zstandard>=0.23",        # replay corpus framing (Sprint 08)
    "cbor2>=5.6",             # replay snapshot encoding (Sprint 08)
    "packaging>=24.0",        # lock policy semver parsing (Sprint 15)

    # ML runtime. `dlm train`, `dlm prompt`, `dlm export` all import these;
    # a `pip install dlm` that omits them would ImportError on first call
    # (audit-05 B1).
    "torch>=2.4",
    "transformers>=4.45",
    "peft>=0.13",
    "trl>=0.12",
    "datasets>=3.0",
    "huggingface-hub>=0.25",
    "accelerate>=1.0",        # dlm train --gpus launcher (Sprint 23; audit-08 M2)
    "safetensors>=0.4",       # GGUF conversion + adapter I/O (Sprint 11)
    "sentencepiece>=0.2",     # vendored llama.cpp convert_lora_to_gguf imports it (audit-08 P3)
]

[project.optional-dependencies]
# CUDA-only. `bitsandbytes` won't install on macOS / CPU-only boxes; gate
# behind an extra so `pip install dlm` stays portable and `pip install
# dlm[cuda]` unlocks QLoRA.
cuda = [
    "bitsandbytes>=0.43",
]
# Apple Silicon only (Sprint 21). `mlx` + `mlx-lm` wheels are darwin-arm64
# exclusives; env markers keep `uv sync --extra mlx` a no-op on non-Apple
# hosts so wheel resolution doesn't fail for Linux/CUDA contributors.
mlx = [
    "mlx>=0.18; sys_platform == 'darwin' and platform_machine == 'arm64'",
    "mlx-lm>=0.19; sys_platform == 'darwin' and platform_machine == 'arm64'",
]
# Sprint 26 observability sinks. TensorBoard + W&B are optional — core
# metrics (SQLite) work without them. `uv sync --extra observability`
# unlocks `dlm train --tensorboard` and `dlm train --wandb <project>`.
observability = [
    "tensorboard>=2.15",
    "wandb>=0.17",
]
# Sprint 35.2 audio-language bases. `soundfile` is the only dep needed
# to decode .wav / .flac / .ogg; MP3 support would require libsndfile
# ≥1.1 and is deferred. Keep optional so text-only users don't pull
# the libsndfile C library. `soxr` is the polyphase resampler used by
# `training.audio.auto_resample=True`; without it (or scipy as a
# fallback) the resample path raises `AudioResampleUnavailable`
# rather than training on the wrong sample rate.
audio = [
    "soundfile>=0.12",
    "soxr>=0.3",
]
# Sprint 43 synth teachers. API clients are optional — `dlm synth`
# works with `self` and `hf:` teachers without these. Install via
# `pip install dlm[openai]` or `pip install dlm[anthropic]`.
openai = [
    "openai>=1.0",
]
anthropic = [
    "anthropic>=0.30",
]
# Sprint 26 (X1) cross-repo bridge: `dlm export --emit-sway-json` calls
# into ``dlm_sway.integrations.dlm.autogen`` to write a ready-to-run
# sway.yaml alongside the GGUF. NB: pulls plain `dlm-sway`, NOT
# `dlm-sway[dlm]` — `dlm-sway[dlm]` would round-trip back to this
# package and create a pip resolver cycle. Plain `dlm-sway` is enough
# because `build_spec_dict` operates on data structures the dlm
# bridge already exposes from sway's side.
sway = [
    "dlm-sway>=0.1.0",
]

[project.scripts]
dlm = "dlm.cli.app:main"

[project.urls]
Homepage = "https://github.com/tenseleyFlow/DocumentLanguageModel"
Issues = "https://github.com/tenseleyFlow/DocumentLanguageModel/issues"

[dependency-groups]
dev = [
    # Test + lint tooling only. ML runtime moved to [project].dependencies
    # (audit-05 B1) so `pip install dlm` gives users a working CLI.
    "pytest>=8.0",
    "pytest-cov>=5.0",
    "mypy>=1.11",
    "ruff>=0.6",
    "types-pyyaml>=6.0",
    "types-psutil>=6.0",
    "hypothesis>=6.152.1",
]
docs = [
    # Sprint 16: MkDocs Material site. Separated from `dev` so contributors
    # working on code don't pay the docs dep surface.
    "mkdocs>=1.6",
    "mkdocs-material>=9.5",
]

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.build.targets.wheel]
packages = ["src/dlm"]

# -------- ruff --------
[tool.ruff]
line-length = 100
target-version = "py311"
src = ["src", "tests"]
# Vendored third-party code (llama.cpp) — don't lint or format it.
extend-exclude = ["vendor/llama.cpp"]

[tool.ruff.lint]
select = [
    "E",    # pycodestyle errors
    "F",    # pyflakes
    "W",    # pycodestyle warnings
    "I",    # isort
    "UP",   # pyupgrade
    "B",    # bugbear
    "N",    # pep8-naming
    "C4",   # comprehensions
    "SIM",  # simplify
    "PT",   # pytest
    "RET",  # return
    "ARG",  # unused args
    "PTH",  # use pathlib
    "TID",  # tidy imports
]
ignore = [
    "E501",  # handled by formatter
]

[tool.ruff.lint.per-file-ignores]
# SIM117 (combinable `with`) is noisy in tests that stream tar+zstd
# into a sequence of readers; the nested form is clearer than the
# comma-separated one.
"tests/**/*.py" = ["ARG", "PT011", "SIM117"]
# Typer stub subcommands accept every CLI arg the real implementation
# will take so `--help` reflects the shipping surface — even though the
# stub body discards them.
"src/dlm/cli/commands.py" = ["ARG001"]
# HuggingFace Trainer callbacks MUST accept `args`/`state`/`control`
# positionally even when the implementation only reads some of them —
# HF dispatches them by position. ARG002 for these wrappers is noise.
"src/dlm/train/cpt/embed_warmup.py" = ["ARG002"]
# Modality dispatch uses a polymorphic interface — each subclass uses
# a different subset of the keyword args (text.dispatch_export reads
# none, VL reads gguf_emission_context, audio ignores it). ARG002
# flags the unused ones in each branch; the shared signature is the
# point of the abstraction.
"src/dlm/modality/*.py" = ["ARG002"]

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

# -------- mypy --------
[tool.mypy]
strict = true
python_version = "3.11"
packages = ["dlm"]
mypy_path = "src"
warn_return_any = true
warn_unused_ignores = true
warn_redundant_casts = true
no_implicit_optional = true
disallow_untyped_decorators = true
plugins = ["pydantic.mypy"]

[tool.pydantic-mypy]
init_forbid_extra = true
init_typed = true
warn_required_dynamic_aliases = true

# HF ecosystem packages (Sprint 06/07/09) ship without stubs or py.typed markers.
# Sprint 07 modules touch them through narrow boundaries; annotations are explicit
# at our boundary so the rest of the codebase remains strict.
[[tool.mypy.overrides]]
module = [
    "datasets.*",
    "transformers.*",
    "peft.*",
    "trl.*",
    "huggingface_hub.*",
    "bitsandbytes.*",
]
ignore_missing_imports = true
# Treat calls into these stubless ecosystems as Any-typed; their
# actual return types aren't reflected in stubs.
disable_error_code = ["no-untyped-call"]

# Optional runtime deps used behind feature gates may be absent or
# ship without type metadata on some CI runners. Keep those imports
# narrow and local in code, and suppress missing-stub noise here
# instead of scattering environment-specific inline ignores.
[[tool.mypy.overrides]]
module = [
    "soxr",
    "scipy",
    "scipy.*",
    "mlx_lm",
    "mlx_lm.*",
]
ignore_missing_imports = true
disable_error_code = ["no-untyped-call"]

# -------- pytest --------
[tool.pytest.ini_options]
testpaths = ["tests"]
addopts = [
    "-ra",
    "-m", "not slow and not gpu and not online and not vl and not audio and not ollama",
]
markers = [
    "slow: expensive; deselected by default",
    "gpu: requires CUDA; skipped on CPU/MPS runners",
    "online: touches the network; skipped in offline CI",
    "vl: vision-language; requires a GPU + VL HF weights; deselected by default",
    "audio: audio-language; requires a GPU + audio HF weights; deselected by default",
    "ollama: requires a local Ollama install (0.4+); deselected by default",
]