forked from mosaicml/llm-foundry
-
Notifications
You must be signed in to change notification settings - Fork 0
/
setup.py
144 lines (129 loc) · 4.81 KB
/
setup.py
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
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
"""MosaicML LLM Foundry package setup."""
import os
import re
import setuptools
from setuptools import setup
_PACKAGE_NAME = "llm-foundry"
_PACKAGE_DIR = "llmfoundry"
_REPO_REAL_PATH = os.path.dirname(os.path.realpath(__file__))
_PACKAGE_REAL_PATH = os.path.join(_REPO_REAL_PATH, _PACKAGE_DIR)
# Read the repo version
# We can't use `.__version__` from the library since it's not installed yet
with open(os.path.join(_PACKAGE_REAL_PATH, "__init__.py")) as f:
content = f.read()
# regex: '__version__', whitespace?, '=', whitespace, quote, version, quote
# we put parens around the version so that it becomes elem 1 of the match
expr = re.compile(r"""^__version__\W+=\W+['"]([0-9\.]*)['"]""", re.MULTILINE)
repo_version = expr.findall(content)[0]
# Use repo README for PyPi description
with open("README.md", "r", encoding="utf-8") as fh:
long_description = fh.read()
# Hide the content between <!-- SETUPTOOLS_LONG_DESCRIPTION_HIDE_BEGIN --> and
# <!-- SETUPTOOLS_LONG_DESCRIPTION_HIDE_END --> tags in the README
while True:
start_tag = "<!-- SETUPTOOLS_LONG_DESCRIPTION_HIDE_BEGIN -->"
end_tag = "<!-- SETUPTOOLS_LONG_DESCRIPTION_HIDE_END -->"
start = long_description.find(start_tag)
end = long_description.find(end_tag)
if start == -1:
assert end == -1, "there should be a balanced number of start and ends"
break
else:
assert end != -1, "there should be a balanced number of start and ends"
long_description = (
long_description[:start] + long_description[end + len(end_tag) :]
)
classifiers = [
"Programming Language :: Python :: 3",
"Programming Language :: Python :: 3.8",
"Programming Language :: Python :: 3.9",
"Programming Language :: Python :: 3.10",
]
install_requires = [
"mosaicml[libcloud,wandb,mlflow,oci,gcs]>=0.16.4,<0.17",
"accelerate>=0.20,<0.21", # for HF inference `device_map`
"transformers>=4.33,<4.34",
"mosaicml-streaming>=0.6,<0.7",
"torch>=1.13.1,<2.1.1",
"datasets>=2.14.5,<2.15",
"fsspec==2023.6.0", # newer version results in a bug in datasets that duplicates data
"sentencepiece==0.1.97",
"einops==0.5.0",
"omegaconf>=2.2.3,<3",
"slack-sdk<4",
"mosaicml-cli>=0.3,<1",
"onnx==1.14.0",
"onnxruntime==1.15.1",
"boto3>=1.21.45,<2",
"huggingface-hub>=0.17.0,<1.0",
]
extra_deps = {}
extra_deps["dev"] = [
"pre-commit>=2.18.1,<3",
"pytest>=7.2.1,<8",
"pytest_codeblocks>=0.16.1,<0.17",
"pytest-cov>=4,<5",
"pyright==1.1.256",
"toml>=0.10.2,<0.11",
"packaging>=21,<23",
"hf_transfer==0.1.3",
]
extra_deps["tensorboard"] = [
"mosaicml[tensorboard]>=0.16.1,<0.17",
]
extra_deps["gpu"] = [
"flash-attn==1.0.9",
"cmake>=3.25.0,<=3.26.3", # required for triton-pre-mlir below
# PyPI does not support direct dependencies, so we remove this line before uploading from PyPI
"triton-pre-mlir@git+https://github.com/vchiley/triton.git@triton_pre_mlir_sm90#subdirectory=python",
"mosaicml-turbo==0.0.4",
# PyPI does not support direct dependencies, so we remove this line before uploading from PyPI
"xentropy-cuda-lib@git+https://github.com/HazyResearch/[email protected]#subdirectory=csrc/xentropy",
]
extra_deps["gpu-flash2"] = [
"flash-attn==2.3.2",
"mosaicml-turbo==0.0.4",
# PyPI does not support direct dependencies, so we remove this line before uploading from PyPI
"xentropy-cuda-lib@git+https://github.com/HazyResearch/[email protected]#subdirectory=csrc/xentropy",
]
extra_deps["peft"] = [
"loralib==0.1.1", # lora core
"bitsandbytes==0.39.1", # 8bit
"scipy>=1.10.0,<=1.11.0", # bitsandbytes dependency; TODO: eliminate when incorporated to bitsandbytes
# TODO: pin peft when it stabilizes.
# PyPI does not support direct dependencies, so we remove this line before uploading from PyPI
"peft==0.4.0",
]
extra_deps["openai"] = [
"openai==0.27.8",
"tiktoken==0.4.0",
]
extra_deps["all-cpu"] = set(
dep for key, deps in extra_deps.items() for dep in deps if "gpu" not in key
)
extra_deps["all"] = set(
dep for key, deps in extra_deps.items() for dep in deps if key != "gpu-flash2"
)
extra_deps["all-flash2"] = set(
dep for key, deps in extra_deps.items() for dep in deps if key != "gpu"
)
setup(
name=_PACKAGE_NAME,
version=repo_version,
author="MosaicML",
author_email="[email protected]",
description="LLM Foundry",
long_description=long_description,
long_description_content_type="text/markdown",
url="https://github.com/mosaicml/llm-foundry/",
package_data={
"llmfoundry": ["py.typed"],
},
packages=setuptools.find_packages(
exclude=[".github*", "mcli*", "scripts*", "tests*"]
),
classifiers=classifiers,
install_requires=install_requires,
extras_require=extra_deps,
python_requires=">=3.7",
)