finetune_quant.py

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

from peft import prepare_model_for_kbit_training
from peft import LoraConfig, get_peft_model

from datasets import load_dataset

import transformers

model_id = "mosaicml/mpt-7b"
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config, device_map={"":0})

model.gradient_checkpointing_enable()
model = prepare_model_for_kbit_training(model)

def print_trainable_parameters(model):
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
    )

config = LoraConfig(
    r=8, 
    lora_alpha=32, 
    target_modules=["query_key_value"], 
    lora_dropout=0.05, 
    bias="none", 
    task_type="CAUSAL_LM"
)

model = get_peft_model(model, config)
print_trainable_parameters(model)

data = load_dataset("Abirate/english_quotes")
data = data.map(lambda samples: tokenizer(samples["quote"]), batched=True)

tokenizer.pad_token = tokenizer.eos_token

trainer = transformers.Trainer(
    model=model,
    train_dataset=data["train"],
    args=transformers.TrainingArguments(
        per_device_train_batch_size=1,
        gradient_accumulation_steps=4,
        warmup_steps=2,
        max_steps=10,
        learning_rate=2e-4,
        fp16=True,
        logging_steps=1,
        output_dir="outputs",
        optim="paged_adamw_8bit"
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)
model.config.use_cache = False  
trainer.train()