tuning

from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, DataCollatorForSeq2Seq, TrainingArguments, Trainer
ds = Dataset.load_from_disk("../data/alpaca_data_zh/")
from modelscope import snapshot_download
# path = snapshot_download('langboat/bloom-1b4-zh')
path = snapshot_download('AI-ModelScope/bloom-3b')
tokenizer = AutoTokenizer.from_pretrained(path)
def process_func(example):
    MAX_LENGTH = 256
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer("\n".join(["Human: " + example["instruction"], example["input"]]).strip() + "\n\nAssistant: ")
    response = tokenizer(example["output"] + tokenizer.eos_token)
    input_ids = instruction["input_ids"] + response["input_ids"]
    attention_mask = instruction["attention_mask"] + response["attention_mask"]
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"]
    if len(input_ids) > MAX_LENGTH:
        input_ids = input_ids[:MAX_LENGTH]
        attention_mask = attention_mask[:MAX_LENGTH]
        labels = labels[:MAX_LENGTH]
    return {
        "input_ids": input_ids,
        "attention_mask": attention_mask,
        "labels": labels
    }
tokenized_ds = ds.map(process_func, remove_columns=ds.column_names)
model = AutoModelForCausalLM.from_pretrained(path, low_cpu_mem_usage=True)
sum(param.numel() for param in model.parameters())
# bitfit
# 选择模型参数里面的所有bias部分

num_param = 0
for name, param in model.named_parameters():
    if "bias" not in name:
        param.requires_grad = False
    else:
        num_param += param.numel()

num_param
num_param / sum(param.numel() for param in model.parameters())
args = TrainingArguments(
    output_dir="./chatbot",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    logging_steps=10,
    num_train_epochs=1
)
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_ds,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
)
trainer.train()


from modelscope import snapshot_download
# path = snapshot_download('langboat/bloom-1b4-zh')
path = snapshot_download('AI-ModelScope/bloom-3b')
tokenizer = AutoTokenizer.from_pretrained(path)
path = '/mnt/workspace/transformers-code/03-PEFT/17-bitfit/chatbot/checkpoint-3000'

model = AutoModelForCausalLM.from_pretrained(path, low_cpu_mem_usage=True)
model = model.cuda()
ipt = tokenizer("Human: {}\n{}".format("考试有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt").to(model.device)
tokenizer.decode(model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True)

from peft import PromptTuningConfig, get_peft_model, TaskType, PromptTuningInit

# Soft Prompt
# config = PromptTuningConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10)
# config
# Hard Prompt
config = PromptTuningConfig(task_type=TaskType.CAUSAL_LM,
                            prompt_tuning_init=PromptTuningInit.TEXT,
                            prompt_tuning_init_text="下面是一段人与机器人的对话。",
                            num_virtual_tokens=len(tokenizer("下面是一段人与机器人的对话。")["input_ids"]),
                            tokenizer_name_or_path=path)


model = get_peft_model(model, config)
print(model.print_trainable_parameters())
args = TrainingArguments(
    output_dir="./chatbot",
    per_device_train_batch_size=1,
    gradient_accumulation_steps=8,
    logging_steps=10,
    num_train_epochs=1,
    save_steps=20,
)
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_ds,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
)
trainer.train()

加载

from peft import PeftModel
peft_model = PeftModel.from_pretrained(model=model, model_id="./chatbot/checkpoint-800/")
peft_model = peft_model.cuda()
ipt = tokenizer("Human: {}\n{}".format("考试有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt").to(peft_model.device)
print(tokenizer.decode(peft_model.generate(**ipt, max_length=128, do_sample=True)[0], skip_special_tokens=True))

P-tuning

from peft import PromptEncoderConfig, TaskType, get_peft_model, PromptEncoderReparameterizationType

config = PromptEncoderConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10,
                             encoder_reparameterization_type=PromptEncoderReparameterizationType.MLP,
                             encoder_dropout=0.1, encoder_num_layers=5, encoder_hidden_size=1024)

prefix tuning

from peft import PrefixTuningConfig, get_peft_model, TaskType

config = PrefixTuningConfig(task_type=TaskType.CAUSAL_LM, num_virtual_tokens=10, prefix_projection=True)
model = get_peft_model(model, config)

model.print_trainable_parameters()

trainable params: 399,951,360 || all params: 3,402,508,800 || trainable%: 11.754601780897671

lora

for name, parameter in model.named_parameters():
    print(name)

from peft import LoraConfig, TaskType, get_peft_model

config = LoraConfig(task_type=TaskType.CAUSAL_LM, target_modules=".*\.1.*k_proj", modules_to_save=["word_embeddings"])
config = LoraConfig(task_type=TaskType.CAUSAL_LM, target_modules=["k_proj","v_proj"])
config

lora推理

from transformers import AutoModelForCausalLM, AutoTokenizer

from peft import PeftModel

from modelscope import snapshot_download
path = snapshot_download('AI-ModelScope/bloom-3b')
model = AutoModelForCausalLM.from_pretrained(path, low_cpu_mem_usage=True)
tokenizer = AutoTokenizer.from_pretrained(path)
p_model = PeftModel.from_pretrained(model, model_id="./chatbot/checkpoint-500/")
ipt = tokenizer("Human: {}\n{}".format("考试有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt")
tokenizer.decode(p_model.generate(**ipt, do_sample=False)[0], skip_special_tokens=True)
#模型合并
merge_model = p_model.merge_and_unload()
ipt = tokenizer("Human: {}\n{}".format("考试有哪些技巧？", "").strip() + "\n\nAssistant: ", return_tensors="pt")
tokenizer.decode(merge_model.generate(**ipt, do_sample=False)[0], skip_special_tokens=True)
#保存模型
merge_model.save_pretrained("./chatbot/merge_model")