muVLA/pyproject.toml at main · CognitiveAISystems/muVLA · 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
118
[build-system]
requires = ["setuptools"]
build-backend = "setuptools.build_meta"

[project]
name = "mu-vla"
authors = [
    {name = "the muVLA authors"},
]
description = "mu-VLA: recurrent memory for Vision-Language-Action models in partially observable manipulation"
version = "1.0.0"
readme = "README.md"
requires-python = ">=3.9,<3.12"
keywords = ["vision-language-action models", "recurrent memory", "POMDP", "robot learning"]
license = {file = "LICENSE"}
classifiers = [
    "Development Status :: 3 - Alpha",
    "Intended Audience :: Developers",
    "Intended Audience :: Education",
    "Intended Audience :: Science/Research",
    "License :: OSI Approved :: MIT License",
    "Operating System :: OS Independent",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.9",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3 :: Only",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
    "accelerate>=0.25.0",
    "draccus==0.8.0",
    "einops",
    # flash-attn is deliberately absent: nothing in mu-VLA requests
    # attn_implementation="flash_attention_2" (the one call site is commented out),
    # and the published results were produced without it. See SETUP.md.
    "huggingface_hub",
    "json-numpy",
    "jsonlines",
    "matplotlib",
    "peft==0.11.1",
    "protobuf==4.25.9",
    "rich",
    "sentencepiece==0.1.99",
    "timm==0.9.10",
    "tokenizers==0.19.1",
    "torch==2.2.1",
    "torchvision==0.17.1",
    "torchaudio==2.2.1",
    # NOTE: this is a REQUIRED fork, not a convenience pin. It adds the memory-token
    # attention mask that mu-VLA depends on (`LlamaConfig.use_mu_vla_memory_mask`).
    # Substituting the PyPI `transformers` package installs cleanly but silently
    # disables memory: mem-tokens are injected and then attended to as ordinary
    # prefix tokens, so training runs and produces meaningless memory metrics.
    "transformers @ git+https://github.com/CognitiveAISystems/transformers-mu-openvla-oft.git@9dbc09f574912a45dd0d71354c035e3c37bcce9e",
    "wandb",
    "tensorflow==2.15.0",
    "tensorflow_datasets==4.9.3",
    # Must stay <1.18: newer releases ship protobuf-5.27-generated code,
    # which cannot load under the protobuf 4.x that tensorflow 2.15 requires.
    "tensorflow_metadata==1.17.3",
    "tensorflow_graphics==2021.12.3",
    "dlimp @ git+https://github.com/moojink/dlimp_openvla@040105d256bd28866cc6620621a3d5f7b6b91b46",
    "diffusers==0.30.3",
    "imageio[ffmpeg,pyav]",
    "opencv-python==4.10.0.84",
    "gymnasium==0.29.1",
    # Pinned to the tag the published results were produced against. v1.0.0 renamed
    # two observation keys (`prompt` -> `task_cue`, `joints` -> `proprio`) and
    # experiments/robot/mikasa_robo/ is written against that spelling, so this pin
    # must not float. Note the upstream packaging gap documented in SETUP.md:
    # `low_poly_light_bulb.glb` is missing from every built distribution, affecting
    # the 8 lamp environments (none of which are in the 32-env benchmark).
    "mikasa-robo-suite @ git+https://github.com/CognitiveAISystems/MIKASA-Robo.git@v1.0.0",
]

[project.optional-dependencies]
dev = [
    "black>=24.2.0",
    "gpustat",
    "ipython",
    "pre-commit",
    "ruff>=0.2.2",
]

[project.urls]
homepage = "https://github.com/CognitiveAISystems/muVLA"
repository = "https://github.com/CognitiveAISystems/muVLA"

[tool.uv]
# Resolution is restricted to the platform the published results were produced on.
# torch 2.2.1 + tensorflow 2.15.0 have no working combination on macOS or on
# Python != 3.10, so a wider range would only produce locks that fail at install
# time. See SETUP.md.
environments = ["sys_platform == 'linux' and platform_machine == 'x86_64' and python_full_version == '3.10.*'"]
prerelease = "allow"

[tool.setuptools.packages.find]
where = ["."]
exclude = ["cache"]

[tool.setuptools.package-data]
"prismatic" = ["py.typed"]

[tool.black]
line-length = 121
target-version = ["py39", "py310"]
preview = true

[tool.ruff]
line-length = 121
target-version = "py39"

[tool.ruff.lint]
select = ["A", "B", "E", "F", "I", "RUF", "W"]
ignore = ["F722"]

[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["E402", "F401"]