Spaces:
Sleeping
Sleeping
File size: 6,825 Bytes
26d66c1 |
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 |
import torch
import os
import locale
import math
import mlflow
import pandas as pd
from trl import SFTTrainer
from datasets import Dataset
from transformers import AutoTokenizer, AutoModel, LlamaForCausalLM, TrainingArguments, BitsAndBytesConfig
from peft import get_peft_model, LoraConfig
from src.logging import logger
from src.models.entity import (ModelTrainerConfig,
LoraParameters,
BitsAndBytesParameters,
TrainingArgumentsParameters)
class ModelTrainer:
def __init__(self, model_trainer_config: ModelTrainerConfig, lora_parameters: LoraParameters, bits_and_bytes_parameters: BitsAndBytesParameters, training_arguments: TrainingArgumentsParameters):
self.model_trainer_config = model_trainer_config
self.lora_parameters = lora_parameters
self.bits_and_bytes_parameters = bits_and_bytes_parameters
self.training_arguments = training_arguments
def __load_data(self):
train_dataset = pd.read_csv(os.path.join(self.model_trainer_config.data_path, "train_dataset.csv"))
eval_dataset = pd.read_csv(os.path.join(self.model_trainer_config.data_path, "eval_dataset.csv"))
train_dataset = Dataset.from_pandas(train_dataset)
eval_dataset = Dataset.from_pandas(eval_dataset)
self.train_dataset = train_dataset
self.eval_dataset = eval_dataset
logger.info("Data loaded")
def __initialize_tokenizer(self, model_name: str):
self.tokenizer = AutoTokenizer.from_pretrained(self.model_trainer_config.base_model)
self.tokenizer.add_special_tokens({'pad_token': '[PAD]'})
logger.info("Tokenizer initialized")
def __initialize_lora(self):
self.lora_cofig = LoraConfig(
r = self.lora_parameters.r,
target_modules = self.lora_parameters.target_modules,
lora_alpha = self.lora_parameters.lora_alpha,
lora_dropout = self.lora_parameters.lora_dropout,
bias = self.lora_parameters.bias,
task_type = self.lora_parameters.task_type
)
logger.info("Lora initialized")
def __initialize_bits_and_bytes(self):
self.nf4_config = BitsAndBytesConfig(
load_in_4bit = self.bits_and_bytes_parameters.load_in_4bit,
bnb_4bit_quant_type = self.bits_and_bytes_parameters.bnb_4bit_quant_type,
bnb_4bit_use_double_quant = self.bits_and_bytes_parameters.bnb_4bit_use_double_quant,
bnb_4bit_compute_dtype = torch.bfloat16
)
logger.info("Bits and bytes initialized")
def __initialize_training_arguments(self):
self.training_args = TrainingArguments(
output_dir = self.training_arguments.output_dir,
evaluation_strategy = self.training_arguments.evaluation_strategy,
save_strategy = self.training_arguments.save_strategy,
num_train_epochs = self.training_arguments.num_train_epochs,
per_device_train_batch_size = self.training_arguments.per_device_train_batch_size,
gradient_accumulation_steps = self.training_arguments.gradient_accumulation_steps,
optim = self.training_arguments.optim,
learning_rate = self.training_arguments.learning_rate,
fp16 = self.training_arguments.fp16,
max_grad_norm = self.training_arguments.max_grad_norm,
warmup_ratio = self.training_arguments.warmup_ratio,
group_by_length = self.training_arguments.group_by_length,
lr_scheduler_type = self.training_arguments.lr_scheduler_type
)
logger.info("Training arguments initialized")
def __create_model(self):
self.model = LlamaForCausalLM.from_pretrained(
self.model_trainer_config.base_model, device_map='auto', quantization_config=self.nf4_config,
)
self.model = get_peft_model(self.model, self.lora_config)
#self.model.print_trainable_parameters()
logger.info("Model created")
def __evaluate(self, trainer):
evaluation_results = trainer.evaluate()
logger.info(f"Perplexity: {math.exp(evaluation_results['eval_loss']):.2f}")
def __save_model(self):
self.model.save_pretrained(os.path.join(self.config.root_dir, f"{self.model_trainer_config.base_model}-math-model"))
logger.info("Model saved")
def __save_tokenizer(self):
self.tokenizer.save_pretrained(os.path.join(self.config.root_dir,"tokenizer"))
logger.info("Tokenizer saved")
def train(self):
if self.model_trainer_config.upload_from_hf:
logger.info("Uploading model from HuggingFace")
self.__initialize_tokenizer(self.model_trainer_config.hf_model_name)
self.__initialize_bits_and_bytes()
self.model = AutoModel.from_pretrained(self.model_trainer_config.hf_model_name, device_map='auto', quantization_config=self.nf4_config)
self.__save_model()
self.__save_tokenizer()
return None
if torch.cuda.is_available():
try:
locale.getpreferredencoding = lambda: "UTF-8"
self.__load_data()
self.__initialize_tokenizer(self.model_trainer_config.base_model)
self.__initialize_lora()
self.__initialize_bits_and_bytes()
self.__initialize_training_arguments()
self.__create_model()
trainer = SFTTrainer(self.model,
train_dataset=self.train_dataset,
eval_dataset=self.eval_dataset,
dataset_text_field="text",
max_seq_length=256,
args=self.training_args,
)
logger.info("Trainer created")
#Upcast layer norms to float 32 for stability
for name, module in trainer.model.named_modules():
if "norm" in name:
module = module.to(torch.float32)
logger.info("Layer norms upcasted to float32")
logger.info(">>>>>>> Training started <<<<<<<<")
with mlflow.start_run(run_name=self.model_trainer_config.training_name):
trainer.train()
logger.info(">>>>>>> Training completed <<<<<<<<")
self.__evaluate(trainer)
self.__save_model()
self.__save_tokenizer()
except Exception as e:
raise e
else:
raise Exception("No GPU found") |