sway Public

Watch 0 Fork 0 Star 0
TOML · 5481 bytes Raw Blame History
1 [project]
2 name = "dlm-sway"
3 version = "0.1.0.dev0"
4 description = "Differential testing for fine-tuned causal LMs: did LoRA/QLoRA training actually change behavior, or is the model defaulting to the pretrained base?"
5 readme = "README.md"
6 requires-python = ">=3.11"
7 license = { text = "MIT" }
8 authors = [{ name = "Matt Wolffe", email = "mfwolffe@outlook.com" }]
9 keywords = [
10 "lora",
11 "qlora",
12 "peft",
13 "fine-tuning",
14 "evaluation",
15 "llm",
16 "differential-testing",
17 ]
18 classifiers = [
19 "Development Status :: 3 - Alpha",
20 "Intended Audience :: Developers",
21 "Intended Audience :: Science/Research",
22 "License :: OSI Approved :: MIT License",
23 "Programming Language :: Python :: 3",
24 "Programming Language :: Python :: 3.11",
25 "Programming Language :: Python :: 3.12",
26 "Topic :: Scientific/Engineering :: Artificial Intelligence",
27 ]
28
29 # Core deps: spec loading, orchestration, reporting. No torch — a user
30 # who only defines specs or writes a custom backend shouldn't pull 3 GB
31 # of CUDA wheels.
32 dependencies = [
33 "pydantic>=2.9",
34 "pyyaml>=6.0",
35 "typer>=0.12",
36 "rich>=13.7",
37 "numpy>=1.26",
38 "packaging>=24.0",
39 ]
40
41 [project.optional-dependencies]
42 # HuggingFace + PEFT scoring backend. The canonical path.
43 hf = [
44 "torch>=2.4",
45 "transformers>=4.45",
46 "peft>=0.13",
47 "safetensors>=0.4",
48 ]
49 # Apple Silicon inference. Env markers keep `uv sync --extra mlx` a no-op
50 # on non-Apple hosts so Linux/CUDA contributors' wheel resolution stays
51 # sane.
52 mlx = [
53 "mlx>=0.18; sys_platform == 'darwin' and platform_machine == 'arm64'",
54 "mlx-lm>=0.19; sys_platform == 'darwin' and platform_machine == 'arm64'",
55 ]
56 # Stylistic fingerprinting (C1). spaCy models pull at runtime via
57 # `python -m spacy download`.
58 style = [
59 "spacy>=3.7",
60 "textstat>=0.7",
61 "nlpaug>=1.1",
62 ]
63 # Semantic similarity (A2). MiniLM ~80 MB, CPU-friendly.
64 semsim = [
65 "sentence-transformers>=3.0",
66 ]
67 # Optional .dlm integration. Only imported inside dlm_sway.integrations.dlm.
68 dlm = [
69 "dlm>=0.9",
70 ]
71 # Visualization (P9).
72 viz = [
73 "matplotlib>=3.8",
74 ]
75 all = [
76 "torch>=2.4",
77 "transformers>=4.45",
78 "peft>=0.13",
79 "safetensors>=0.4",
80 "mlx>=0.18; sys_platform == 'darwin' and platform_machine == 'arm64'",
81 "mlx-lm>=0.19; sys_platform == 'darwin' and platform_machine == 'arm64'",
82 "spacy>=3.7",
83 "textstat>=0.7",
84 "nlpaug>=1.1",
85 "sentence-transformers>=3.0",
86 "matplotlib>=3.8",
87 ]
88
89 [project.scripts]
90 dlm-sway = "dlm_sway.cli.app:main"
91
92 [project.urls]
93 Homepage = "https://github.com/tenseleyFlow/DocumentLanguageModel"
94 Issues = "https://github.com/tenseleyFlow/DocumentLanguageModel/issues"
95
96 [dependency-groups]
97 dev = [
98 "pytest>=8.0",
99 "pytest-cov>=5.0",
100 "mypy>=1.11",
101 "ruff>=0.6",
102 "types-pyyaml>=6.0",
103 "hypothesis>=6.152.1",
104 ]
105
106 [build-system]
107 requires = ["hatchling"]
108 build-backend = "hatchling.build"
109
110 [tool.hatch.build.targets.wheel]
111 packages = ["src/dlm_sway"]
112
113 # -------- ruff --------
114 [tool.ruff]
115 line-length = 100
116 target-version = "py311"
117 src = ["src", "tests"]
118
119 [tool.ruff.lint]
120 select = [
121 "E", # pycodestyle errors
122 "F", # pyflakes
123 "W", # pycodestyle warnings
124 "I", # isort
125 "UP", # pyupgrade
126 "B", # bugbear
127 "N", # pep8-naming
128 "C4", # comprehensions
129 "SIM", # simplify
130 "PT", # pytest
131 "RET", # return
132 "ARG", # unused args
133 "PTH", # use pathlib
134 "TID", # tidy imports
135 ]
136 ignore = [
137 "E501", # handled by formatter
138 ]
139
140 [tool.ruff.lint.per-file-ignores]
141 "tests/**/*.py" = ["ARG", "PT011", "SIM117"]
142 # PyTorch's canonical `import torch.nn.functional as F` is universally
143 # read, so we allow the naming exception in the HF backend only.
144 "src/dlm_sway/backends/hf.py" = ["N812"]
145 # The .dlm bridge is the one place allowed to import the ``dlm`` package.
146 "src/dlm_sway/integrations/dlm/*.py" = ["TID251"]
147
148 [tool.ruff.lint.flake8-tidy-imports.banned-api]
149 # Hard architectural boundary: the `dlm` package is only importable
150 # from inside the optional integration shim. This keeps dlm-sway
151 # usable for anyone with just a HuggingFace base + PEFT adapter.
152 "dlm".msg = "Import `dlm` only from dlm_sway.integrations.dlm (the optional extra)."
153
154 [tool.ruff.format]
155 quote-style = "double"
156 indent-style = "space"
157
158 # -------- mypy --------
159 [tool.mypy]
160 strict = true
161 python_version = "3.11"
162 packages = ["dlm_sway"]
163 mypy_path = "src"
164 warn_return_any = true
165 warn_unused_ignores = true
166 warn_redundant_casts = true
167 no_implicit_optional = true
168 disallow_untyped_decorators = true
169 plugins = ["pydantic.mypy"]
170
171 [tool.pydantic-mypy]
172 init_forbid_extra = true
173 init_typed = true
174 warn_required_dynamic_aliases = true
175
176 # Stubless ML ecosystem packages. Narrow boundaries in backends/* import
177 # them explicitly; the rest of the codebase stays strict.
178 [[tool.mypy.overrides]]
179 module = [
180 "torch",
181 "torch.*",
182 "transformers.*",
183 "peft.*",
184 "safetensors.*",
185 "mlx.*",
186 "mlx_lm.*",
187 "sentence_transformers.*",
188 "spacy.*",
189 "textstat.*",
190 "nlpaug.*",
191 "huggingface_hub.*",
192 "dlm.*",
193 ]
194 ignore_missing_imports = true
195 disable_error_code = ["no-untyped-call"]
196
197 # -------- pytest --------
198 [tool.pytest.ini_options]
199 testpaths = ["tests"]
200 addopts = [
201 "-ra",
202 "-m", "not slow and not gpu and not online",
203 ]
204 markers = [
205 "slow: expensive; deselected by default",
206 "gpu: requires CUDA; skipped on CPU/MPS runners",
207 "online: touches the network; skipped in offline CI",
208 ]