turboquant_plus/pyproject.toml at main · TheTom/turboquant_plus · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
[project]
name = "refract-llm"
version = "0.3.2.3"
description = "REFRACT — Reference-anchored Robust Acid-test for Compressed Transformers. Multi-axis KV-cache fidelity scoring for LLMs across llama.cpp, MLX, vLLM, and SGLang."
readme = "refract/README.md"
license = { text = "MIT" }
requires-python = ">=3.9"
authors = [
    { name = "Tom Turney (TheTom)" },
]
keywords = [
    "llm",
    "kv-cache",
    "quantization",
    "evaluation",
    "fidelity",
    "llama-cpp",
    "mlx",
    "vllm",
    "sglang",
    "perplexity",
    "kld",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: MIT License",
    "Operating System :: MacOS",
    "Operating System :: POSIX :: Linux",
    "Operating System :: Microsoft :: Windows",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.9",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]

# REFRACT itself has zero non-stdlib runtime deps for the framework — kept
# that way deliberately so `pip install refract-llm` is small and doesn't
# pull torch / transformers / vllm by default. Backend extras opt into
# their own dependencies.
dependencies = []

[project.optional-dependencies]
# Apple Silicon (Metal). MLX backend uses native Python — no patches needed.
refract-mlx = [
    "mlx>=0.20",
    "mlx-lm>=0.31",  # qwen3.5 + gemma-4 support requires 0.31+
]
# CUDA / ROCm. vLLM backend uses an in-process LLM via vllm.LLM.
# Caches one LLM at a time and evicts on KV-config change so hybrid models
# (Qwen3.6-35B-A3B etc.) that don't fit two simultaneous instances can
# still be scored.
refract-vllm = [
    "vllm>=0.5",
]
# SGLang backend is HTTP-based; the user runs the SGLang server separately
# (typically via the published Docker image). REFRACT only needs a HTTP
# client and a tokenizer to construct prompt token IDs.
refract-sglang = [
    "requests>=2.28",
    "transformers>=4.40",
]
# Convenience: install all backend extras at once.
full = [
    "mlx>=0.20",
    "mlx-lm>=0.31",
    "vllm>=0.5",
    "requests>=2.28",
    "transformers>=4.40",
]
# Development.
dev = [
    "pytest>=7.0",
    "pytest-cov>=4.0",
    "coverage>=7.0",
    "build>=1.0",
    "twine>=5.0",
]

[project.urls]
Homepage = "https://github.com/TheTom/turboquant_plus/tree/main/refract"
Repository = "https://github.com/TheTom/turboquant_plus"
Issues = "https://github.com/TheTom/turboquant_plus/issues"
Changelog = "https://github.com/TheTom/turboquant_plus/blob/main/refract/CHANGELOG.md"
"Cross-engine bench" = "https://github.com/TheTom/turboquant_plus/blob/main/docs/papers/cross-engine-mi300x.md"

[project.scripts]
refract = "refract.cli:main"

[tool.setuptools.packages.find]
# Only ship the refract framework. The `turboquant` Python package in the
# same repo is dev-only (TurboQuant kernels research workspace) and is NOT
# part of this wheel.
include = ["refract*"]

[tool.setuptools.package-data]
# Ship prompts, llama.cpp patches, and example reports inside the wheel so
# users can run REFRACT immediately and apply the trajectory patch to their
# own llama.cpp build without cloning the source repo.
refract = [
    "prompts/*.jsonl",
    "examples/*.json",
    "examples/*.html",
    "examples/*.md",
    "llama-cpp-patches/*.patch",
]

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

[tool.pytest.ini_options]
# Refract's tests are the ones gating wheel releases.
testpaths = ["refract/tests"]