Create config_loader.py
Browse files- config_loader.py +174 -0
config_loader.py
ADDED
@@ -0,0 +1,174 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
from typing import Dict, Any, Optional
|
4 |
+
import torch
|
5 |
+
from transformers import (
|
6 |
+
AutoModelForCausalLM,
|
7 |
+
AutoTokenizer,
|
8 |
+
PreTrainedTokenizerFast,
|
9 |
+
TrainingArguments,
|
10 |
+
Trainer
|
11 |
+
)
|
12 |
+
|
13 |
+
class ConfigLoader:
|
14 |
+
"""A utility class to load configs and instantiate transformers objects."""
|
15 |
+
|
16 |
+
def __init__(self, config_path: str, default_dir: str = "../configs"):
|
17 |
+
"""Initialize with a config file path."""
|
18 |
+
self.config_path = os.path.join(default_dir, config_path) if not os.path.isabs(config_path) else config_path
|
19 |
+
self.config = {}
|
20 |
+
self.default_dir = default_dir
|
21 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
22 |
+
self._load_config()
|
23 |
+
|
24 |
+
def _load_config(self) -> None:
|
25 |
+
"""Load the configuration from a JSON file."""
|
26 |
+
if not os.path.exists(self.config_path):
|
27 |
+
raise FileNotFoundError(f"Config file not found: {self.config_path}")
|
28 |
+
|
29 |
+
try:
|
30 |
+
with open(self.config_path, "r", encoding="utf-8") as f:
|
31 |
+
self.config = json.load(f)
|
32 |
+
print(f"✅ Loaded config from {self.config_path}")
|
33 |
+
except json.JSONDecodeError as e:
|
34 |
+
raise ValueError(f"Invalid JSON in {self.config_path}: {e}")
|
35 |
+
except Exception as e:
|
36 |
+
raise RuntimeError(f"Error loading config: {e}")
|
37 |
+
|
38 |
+
def get(self, key: str, default: Any = None) -> Any:
|
39 |
+
"""Get a value from the config with an optional default."""
|
40 |
+
return self.config.get(key, default)
|
41 |
+
|
42 |
+
def validate(self, required_keys: list = None):
|
43 |
+
"""Validate required keys in the config."""
|
44 |
+
if required_keys:
|
45 |
+
missing = [key for key in required_keys if key not in self.config]
|
46 |
+
if missing:
|
47 |
+
raise KeyError(f"Missing required keys in config: {missing}")
|
48 |
+
|
49 |
+
def save(self, save_path: Optional[str] = None) -> None:
|
50 |
+
"""Save the current config to a file."""
|
51 |
+
path = save_path or self.config_path
|
52 |
+
os.makedirs(os.path.dirname(path), exist_ok=True)
|
53 |
+
try:
|
54 |
+
with open(path, "w", encoding="utf-8") as f:
|
55 |
+
json.dump(self.config, f, indent=4)
|
56 |
+
print(f"✅ Config saved to {path}")
|
57 |
+
except Exception as e:
|
58 |
+
raise RuntimeError(f"Error saving config: {e}")
|
59 |
+
|
60 |
+
def load_model(self, model_path: Optional[str] = None) -> AutoModelForCausalLM:
|
61 |
+
"""Load a transformers model based on config or path."""
|
62 |
+
try:
|
63 |
+
model_name_or_path = model_path or self.config.get("model_name", "mistralai/Mixtral-8x7B-Instruct-v0.1")
|
64 |
+
model_config = self.config.get("model_config", {})
|
65 |
+
if model_path and not model_config: # Local path without config
|
66 |
+
model = AutoModelForCausalLM.from_pretrained(
|
67 |
+
model_name_or_path,
|
68 |
+
torch_dtype=torch.bfloat16,
|
69 |
+
device_map="auto",
|
70 |
+
low_cpu_mem_usage=True
|
71 |
+
)
|
72 |
+
else: # Use config for custom model
|
73 |
+
from transformers import MistralConfig
|
74 |
+
config = MistralConfig(**model_config)
|
75 |
+
model = AutoModelForCausalLM.from_config(config)
|
76 |
+
return model.to(self.device)
|
77 |
+
except Exception as e:
|
78 |
+
raise RuntimeError(f"Error loading model: {e}")
|
79 |
+
|
80 |
+
def load_tokenizer(self, tokenizer_path: Optional[str] = None) -> PreTrainedTokenizerFast:
|
81 |
+
"""Load a tokenizer based on config or path."""
|
82 |
+
try:
|
83 |
+
tokenizer_path = tokenizer_path or self.config.get("tokenizer_path", "../finetuned_charm15/tokenizer.json")
|
84 |
+
if tokenizer_path.endswith(".json"):
|
85 |
+
tokenizer = PreTrainedTokenizerFast(tokenizer_file=tokenizer_path)
|
86 |
+
else:
|
87 |
+
tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
|
88 |
+
|
89 |
+
if tokenizer.pad_token is None:
|
90 |
+
tokenizer.pad_token = tokenizer.eos_token
|
91 |
+
tokenizer.pad_token_id = tokenizer.eos_token_id
|
92 |
+
print(f"✅ Loaded tokenizer from {tokenizer_path}")
|
93 |
+
return tokenizer
|
94 |
+
except Exception as e:
|
95 |
+
raise RuntimeError(f"Error loading tokenizer: {e}")
|
96 |
+
|
97 |
+
def get_training_args(self) -> TrainingArguments:
|
98 |
+
"""Create TrainingArguments from config."""
|
99 |
+
try:
|
100 |
+
training_config = self.config.get("training_config", {
|
101 |
+
"output_dir": "../finetuned_charm15",
|
102 |
+
"per_device_train_batch_size": 1,
|
103 |
+
"num_train_epochs": 3,
|
104 |
+
"learning_rate": 5e-5,
|
105 |
+
"gradient_accumulation_steps": 8,
|
106 |
+
"bf16": True,
|
107 |
+
"save_strategy": "epoch",
|
108 |
+
"evaluation_strategy": "epoch",
|
109 |
+
"save_total_limit": 2,
|
110 |
+
"logging_steps": 100,
|
111 |
+
"report_to": "none"
|
112 |
+
})
|
113 |
+
return TrainingArguments(**training_config)
|
114 |
+
except Exception as e:
|
115 |
+
raise RuntimeError(f"Error creating TrainingArguments: {e}")
|
116 |
+
|
117 |
+
@staticmethod
|
118 |
+
def get_default_config() -> Dict[str, Any]:
|
119 |
+
"""Return a default config combining model, tokenizer, and training settings."""
|
120 |
+
return {
|
121 |
+
"model_name": "mistralai/Mixtral-8x7B-Instruct-v0.1",
|
122 |
+
"tokenizer_path": "../finetuned_charm15/tokenizer.json",
|
123 |
+
"model_config": {
|
124 |
+
"architectures": ["MistralForCausalLM"],
|
125 |
+
"hidden_size": 4096,
|
126 |
+
"num_hidden_layers": 8,
|
127 |
+
"vocab_size": 32000,
|
128 |
+
"max_position_embeddings": 4096,
|
129 |
+
"torch_dtype": "bfloat16"
|
130 |
+
},
|
131 |
+
"training_config": {
|
132 |
+
"output_dir": "../finetuned_charm15",
|
133 |
+
"per_device_train_batch_size": 1,
|
134 |
+
"num_train_epochs": 3,
|
135 |
+
"learning_rate": 5e-5,
|
136 |
+
"gradient_accumulation_steps": 8,
|
137 |
+
"bf16": True,
|
138 |
+
"save_strategy": "epoch",
|
139 |
+
"evaluation_strategy": "epoch",
|
140 |
+
"save_total_limit": 2,
|
141 |
+
"logging_steps": 100,
|
142 |
+
"report_to": "none"
|
143 |
+
},
|
144 |
+
"generation_config": {
|
145 |
+
"max_length": 2048,
|
146 |
+
"temperature": 0.7,
|
147 |
+
"top_p": 0.9,
|
148 |
+
"top_k": 50,
|
149 |
+
"repetition_penalty": 1.2,
|
150 |
+
"do_sample": True
|
151 |
+
}
|
152 |
+
}
|
153 |
+
|
154 |
+
if __name__ == "__main__":
|
155 |
+
# Example usage
|
156 |
+
config_loader = ConfigLoader("charm15_config.json")
|
157 |
+
|
158 |
+
# Load model and tokenizer
|
159 |
+
model = config_loader.load_model()
|
160 |
+
tokenizer = config_loader.load_tokenizer()
|
161 |
+
|
162 |
+
# Get training args
|
163 |
+
training_args = config_loader.get_training_args()
|
164 |
+
|
165 |
+
# Validate
|
166 |
+
config_loader.validate(["model_name", "training_config"])
|
167 |
+
|
168 |
+
# Test generation
|
169 |
+
inputs = tokenizer("Hello, Charm 15!", return_tensors="pt").to(config_loader.device)
|
170 |
+
outputs = model.generate(**inputs, **config_loader.get("generation_config", {}))
|
171 |
+
print(f"Generated: {tokenizer.decode(outputs[0], skip_special_tokens=True)}")
|
172 |
+
|
173 |
+
# Save updated config
|
174 |
+
config_loader.save("../finetuned_charm15/config.json")
|