博客

参考：https://github.com/huggingface/peft

一般用于大模型微调，下面案例是用于文本nlp模型

**可以安装peft库，里面封装了lora

import argparse
import os

import torch
from torch.optim import AdamW
from torch.utils.data import DataLoader
from peft import (
    get_peft_config,
    get_peft_model,
    get_peft_model_state_dict,
    set_peft_model_state_dict,
    LoraConfig,
    PeftType,
    PrefixTuningConfig,
    PromptEncoderConfig,
)

import evaluate
from datasets import load_dataset
from transformers import AutoModelForSequenceClassification, AutoTokenizer, get_linear_schedule_with_warmup, set_seed
from tqdm import tqdm



batch_size = 32
model_name_or_path = "roberta-large"
task = "mrpc"
peft_type = PeftType.LORA
device = "cuda"
num_epochs = 10


if any(k in model_name_or_path for k in ("gpt", "opt", "bloom")):
    padding_side = "left"
else:
    padding_side = "right"
##加载tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, padding_side=padding_side)
if getattr(tokenizer, "pad_token_id") is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id
    
##加载微调数据
datasets = load_dataset("glue", task)
metric = evaluate.load("glue", task)


def tokenize_function(examples):
    # max_length=None => use the model max length (it's actually the default)
    outputs = tokenizer(examples["sentence1"], examples["sentence2"], truncation=True, max_length=None)
    return outputs


tokenized_datasets = datasets.map(
    tokenize_function,
    batched=True,
    remove_columns=["idx", "sentence1", "sentence2"],
)

# We also rename the 'label' column to 'labels' which is the expected name for labels by the models of the
# transformers library
tokenized_datasets = tokenized_datasets.rename_column("label", "labels")


def collate_fn(examples):
    return tokenizer.pad(examples, padding="longest", return_tensors="pt")


# Instantiate dataloaders.
train_dataloader = DataLoader(tokenized_datasets["train"], shuffle=True, collate_fn=collate_fn, batch_size=batch_size)
eval_dataloader = DataLoader(
    tokenized_datasets["validation"], shuffle=False, collate_fn=collate_fn, batch_size=batch_size
)

##lore模型配置及融合进原模型
peft_config = LoraConfig(task_type="SEQ_CLS", inference_mode=False, r=8, lora_alpha=16, lora_dropout=0.1)
lr = 3e-4

model = AutoModelForSequenceClassification.from_pretrained(model_name_or_path, return_dict=True)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()

##加优化器进行训练
optimizer = AdamW(params=model.parameters(), lr=lr)

# Instantiate scheduler
lr_scheduler = get_linear_schedule_with_warmup(
    optimizer=optimizer,
    num_warmup_steps=0.06 * (len(train_dataloader) * num_epochs),
    num_training_steps=(len(train_dataloader) * num_epochs),
)

# model.to(device)  ##GPU上运行
for epoch in range(num_epochs):
    model.train()
    for step, batch in enumerate(tqdm(train_dataloader)):
        # batch.to(device)  ##GPU上运行
        outputs = model(**batch)
        loss = outputs.loss
        loss.backward()
        optimizer.step()
        lr_scheduler.step()
        optimizer.zero_grad()



    model.eval()
    for step, batch in enumerate(tqdm(eval_dataloader)):
        # batch.to(device) ##GPU上运行
        with torch.no_grad():
            outputs = model(**batch)
        predictions = outputs.logits.argmax(dim=-1)
        predictions, references = predictions, batch["labels"]
        metric.add_batch(
            predictions=predictions,
            references=references,
        )

    eval_metric = metric.compute()
    print(f"epoch {epoch}:", eval_metric)

##保存lore模型
tokenizer.save_pretrained('roberta-large-lore')
model.save_pretrained('roberta-large-lore')

训练过程

保存的模型

加载lore推理

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "roberta-large-lore"  ##上面保存lore模型的文件夹
config = PeftConfig.from_pretrained(peft_model_id)
inference_model = AutoModelForSequenceClassification.from_pretrained(config.base_model_name_or_path)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
inference_model = PeftModel.from_pretrained(inference_model, peft_model_id)

# inference_model.to(device)  ##GPU上运行
inference_model.eval()
for step, batch in enumerate(tqdm(eval_dataloader)):
    # batch.to(device) ##GPU上运行
    with torch.no_grad():
        outputs = inference_model(**batch)
    predictions = outputs.logits.argmax(dim=-1)
    predictions, references = predictions, batch["labels"]
    metric.add_batch(
        predictions=predictions,
        references=references,
    )

eval_metric = metric.compute()
print(eval_metric)

加载案例2：

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
from peft import PeftModel


tokenizer = AutoTokenizer.from_pretrained("baichuan-inc/baichuan-7B", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("baichuan-inc/baichuan-7B", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, "hiyouga/baichuan-7b-sft")
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

query = "晚上睡不着怎么办"

inputs = tokenizer(["<human>:{}\n<bot>:".format(query)], return_tensors="pt")
inputs = inputs.to("cuda")
generate_ids = model.generate(**inputs, max_new_tokens=256, streamer=streamer)

参考：https://zhuanlan.zhihu.com/p/638058537


from peft import PeftModel
from transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer
import torch

# create tokenizer
base_model = "timdettmers/guanaco-33b-merged"
tokenizer = LlamaTokenizer.from_pretrained(base_model)

# base model
model = LlamaForCausalLM.from_pretrained(
        base_model,
        torch_dtype=torch.float16,
        device_map="auto",
    )

# LORA PEFT adapters
adapter_model = "lyogavin/Anima33B"

model = PeftModel.from_pretrained(
        model,
        adapter_model,
        #torch_dtype=torch.float16,
    )
model.eval()

# prompt
prompt = "中国的首都是哪里？"
inputs = tokenizer(prompt, return_tensors="pt")

# Generate
generate_ids = model.generate(**inputs, max_new_tokens=30)
print(tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0])

# output: '中国的首都是哪里？\n中国的首都是北京。\n北京位于中国北部，是中国历史悠'