Arctic-Platform/pyproject.toml at main · Snowflake-AI-Research/Arctic-Platform · 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
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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.metadata]
# this is needed when we want to point to a github branch in `dependencies` - leave it enabled
allow-direct-references = true
license-files = [
  "LICENSE",
]

[tool.hatch.build]
include = [
  "arctic_platform",
  "arctic_platform/integrations/verl/config/**/*.yaml",
  "arctic_platform/integrations/verl/examples/*.sh",
  "arctic_platform/integrations/verl/README.md",
]

[project.urls]
Homepage = "https://github.com/Snowflake-AI-Research/Arctic-Platform"
Repository = "https://github.com/Snowflake-AI-Research/Arctic-Platform"
Issues = "https://github.com/Snowflake-AI-Research/Arctic-Platform/issues"

[project]
name = "arctic_platform"
version = "0.1.4.dev0"
description = "Snowflake LLM training library w/ RL"
authors = [
    {author = "Tunji Ruwase", email = "tunji.ruwase@snowflake.com"},
    {author = "Stas Bekman", email = "stas.bekman@snowflake.com"},
    {author = "Michael Wyatt", email = "michael.wyatt@snowflake.com"},
    {author = "Jeff Rasley", email = "jeff.rasley@snowflake.com"},
    {author = "Samyam Rajbhandari", email = "samyam.rajbhandari@snowflake.com"},
]
license = "Apache-2.0"
readme = "README.md"
keywords = ["llm", "training", "rl"]

classifiers = [
  # How mature is this project?  4 - Beta / 5 - Production/Stable
  "Development Status :: 4 - Beta",

  "Programming Language :: Python :: 3",
  "Programming Language :: Python :: 3.12",
]

requires-python = ">=3.12"

dependencies = [
    "pydantic>=2.10",
    "pydantic-settings>=2.0",  # CortexConfig reads ARCTIC_CORTEX_* via BaseSettings
    "typing-extensions",  # Self, per the pydantic model_validator convention
]

# Optional dependency groups
[project.optional-dependencies]
all = [
    "arctic_platform[dev]",
    "arctic_platform[testing]",
    "arctic_platform[cortex]",
    "arctic_platform[onprem]",
    "arctic_platform[verl]",
]

dev = [
    "arctic_platform[formatting]",
    "arctic_platform[testing]",
]

testing = [
    "parameterized",
    "pytest-instafail",
    "pytest-flakefinder",
    "pytest-sugar",
    "pytest-timeout",
    "pytest-xdist",
    "pytest",
]

formatting = [
    "pre-commit",
    "autoflake",
]

# qwen3.5 and other models that use GDN attention layers need this for much better perf
gdn = [
    "flash-linear-attention",
    "causal-conv1d",
]

# Shared DSSST1 tensor wire format.
wire = [
    "numpy",
    "safetensors",
    "torch",
]

# HTTP clients shared by Cortex and on-prem transports.
transport = [
    "arctic_platform[wire]",
    "aiohttp",
    "requests",
]

# The client, talking to Cortex over SnowAPI.
cortex = [
    "arctic_platform[transport]",
    "tenacity",
    "urllib3",
]

# A training-only on-prem server, plus the client that drives it. Deliberately
# excludes the sampling stack: both servers import arctic_inference lazily so
# this install can serve training without vLLM.
sft = [
    "arctic_platform[transport]",
    "deepspeed>=0.19.2",  # fix the static buffers in mixed precision
    "fastapi",
    "liger-kernel>=0.8.0",
    "nvidia-ml-py",
    "psutil",
    "ray",
    "tensordict",
    "transformers",
    "uvicorn",
]

# [sft] plus sampling. arctic-inference[vllm] pulls vLLM and applies that extra's
# pin; [rl] does not declare vLLM. 0.3.0 is the first release of that extra
# contract.
rl = [
    "arctic_platform[sft]",
    "arctic-inference[server,vllm]>=0.3.0",
]

onprem = [
    "arctic_platform[rl]",
]

# verl adapter (arctic_platform/integrations/verl). verl itself is user-supplied and
# version-pinned in launchers; here we only carry the small set of shared runtime
# dependencies the adapter itself needs beyond [rl].
verl = [
    "arctic_platform[rl]",
    "codetiming",
    "omegaconf",
]

[tool.pytest.ini_options]
# Only auto-collect the maintained suite.
testpaths = ["tests"]

# Per-test wall-clock cap
timeout = 300

# Cooperative whole-suite cap (pytest-timeout): checked between tests, so it bounds total wall-clock for a normal
# run. Disable it for intentionally long runs that exceed it, e.g. flakefinder: `pytest --session-timeout=0 ...`.
session_timeout = 3600

timeout_method = "signal"

markers = [
    "vllm: marks a GPU test that drives the vLLM sampling engine; it cannot run concurrently with other GPU tests, so it is pinned to a single xdist worker via xdist_group. Exclude it from a parallel pool with '-m \"not vllm\"' and run it on its own.",
    "gpu_serial: marks a heavyweight GPU test whose body must hold the host-wide GPU lock (serialized across xdist workers by the _serialize_gpu_work autouse fixture).",
]