model-quantization-recipes/pyproject.toml at main · VinRobotics/model-quantization-recipes · 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
[project]
name = "model-quantization-recipes"
version = "0.1.0"
description = "Quantization recipes and deployment workflows for LLM and ASR models."
requires-python = ">=3.10,<3.13"
authors = [
  { name = "quangnd58" },
  { name = "vrfai" },
]
license = { text = "BSD-3-Clause" }

[project.optional-dependencies]
gemma4 = [
  "torch>=2.3",
  "transformers==4.57.6",
  "accelerate==1.12.0",
  "datasets==2.19.0",
  "huggingface-hub==0.36.2",
  "tokenizers==0.22.2",
  "safetensors==0.7.0",
  "nvidia-modelopt[all]==0.39.0",
  "peft==0.18.1",
  "diffusers==0.37.0",
  "sentencepiece==0.2.1",
  "protobuf==6.33.6",
  "numpy==2.2.6",
  "tqdm==4.67.3",
]
qwen3-asr = [
  "torch>=2.8.0",
  "torchaudio>=2.8.0",
  "transformers>=4.40.0",
  "safetensors>=0.4.0",
  "huggingface-hub>=0.23.0",
  "nvidia-modelopt>=0.39.0",
  "soundfile>=0.12.0",
  "jiwer>=3.0.0",
  "datasets==2.19.0",
  "tqdm>=4.66.0",
  "requests>=2.31.0",
  "numpy>=1.24.0",
  "pyyaml>=6.0.0",
]
qwen36-27b = [
  "torch>=2.10.0",
  "torchvision>=0.25.0",
  "transformers==5.6.0",       # 5.6.x required; patches in quantize.sh target this
  "accelerate==1.12.0",
  "datasets==2.19.0",
  "huggingface-hub==0.36.2",
  "tokenizers==0.22.2",
  "safetensors==0.7.0",
  "llmcompressor==0.10.0.1",
  "compressed-tensors==0.14.0.1",
  "sentencepiece==0.2.1",
  "numpy==2.2.6",
  "tqdm==4.67.3",
  "pyyaml>=6.0.2",
]
qwen38-27b = [
  "torch>=2.10.0",
  "torchvision>=0.25.0",
  "transformers==5.14.1",      # >=5.8 is a hard floor; `qwen3_5` does not exist before it
  "accelerate==1.14.0",
  "datasets==5.0.1",
  "huggingface-hub==1.27.0",
  "tokenizers==0.22.2",
  "safetensors==0.8.0",
  "llmcompressor==0.13.0",     # 0.13 resolves SmoothQuant via `match_modules_set`
  "compressed-tensors==0.18.0",
  "numpy==2.4.6",
  "pandas==3.0.5",
  "pillow==12.3.0",
  "tqdm==4.70.0",
]
qwen36-moe-35b-nvfp4 = [
  "torch>=2.3",
  "transformers==5.5.4",
  "accelerate==1.12.0",
  "datasets==2.19.0",
  "nvidia-modelopt[all]==0.39.0",
  "safetensors==0.7.0",
]
cosmos-reason2 = [
  "torch>=2.10.0",
  "torchvision>=0.25.0",
  "transformers==4.57.6",
  "accelerate==1.12.0",
  "datasets==2.19.0",
  "huggingface-hub==0.36.2",
  "tokenizers==0.22.2",
  "safetensors==0.7.0",
  "llmcompressor==0.10.0.1",
  "compressed-tensors==0.14.0.1",
  "peft==0.18.1",
  "diffusers==0.37.0",
  "sentencepiece==0.2.1",
  "numpy==2.2.6",
  "tqdm==4.67.3",
]

# `build_calib_jsonl.py` is stdlib-only (argparse, gzip, json, random). The one real
# dependency is the `huggingface-cli` entrypoint used to fetch the calibration source data.
# Quantize/export/build run against TensorRT-Edge-LLM's own environment, not this one --
# see recipes/internvla-n1-dualvln/README.md.
internvla-n1-dualvln = [
  "huggingface-hub>=0.36.0",
]

[tool.ruff]
target-version = "py312"

[tool.mypy]
python_version = "3.10"
warn_unused_configs = true

[[tool.mypy.overrides]]
module = [
  "datasets",
  "modelopt",
  "modelopt.*",
  "torch",
  "transformers",
  "llmcompressor",
]
ignore_missing_imports = true