使用 llama3 的 lora 微调报错：NotImplementedError: Cannot copy out of meta tensor; no data!

[chenmonster]

在 colab 环境，执行 llama3 lora 微调报错：NotImplementedError: Cannot copy out of meta tensor; no data! 。微调代码如下：

import pandas as pd
from modelscope import snapshot_download
import torch
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, TrainingArguments, Trainer, DataCollatorForSeq2Seq
from peft import LoraConfig, TaskType, get_peft_model


def process_func(example):
    MAX_LENGTH = 384    # Llama分词器会将一个中文字切分为多个token，因此需要放开一些最大长度，保证数据的完整性
    input_ids, attention_mask, labels = [], [], []
    instruction = tokenizer(f"<|start_header_id|>user<|end_header_id|>\n\n{example['instruction'] + example['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", add_special_tokens=False)  # add_special_tokens 不在开头加 special_tokens
    response = tokenizer(f"{example['output']}<|eot_id|>", add_special_tokens=False)
    input_ids = instruction["input_ids"] + response["input_ids"] + [tokenizer.pad_token_id]
    attention_mask = instruction["attention_mask"] + response["attention_mask"] + [1]  # 因为eos token咱们也是要关注的所以 补充为1
    labels = [-100] * len(instruction["input_ids"]) + response["input_ids"] + [tokenizer.pad_token_id]
    if len(input_ids) > MAX_LENGTH:  # 做一个截断
        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
    }


# 下载模型
model_dir = snapshot_download('LLM-Research/Meta-Llama-3-8B-Instruct', cache_dir='.', revision='master')
print('模型路径：', model_dir)
tokenizer = AutoTokenizer.from_pretrained(model_dir, use_fast=False, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
model = AutoModelForCausalLM.from_pretrained(model_dir, device_map='auto', torch_dtype=torch.bfloat16)
model.enable_input_require_grads() # 开启梯度检查点
# 下载数据集
df = pd.read_json('https://github.com/datawhalechina/self-llm/raw/master/dataset/huanhuan.json')
ds = Dataset.from_pandas(df)
print('数据集：', ds[:3])
# 处理数据集
tokenized_id = ds.map(process_func, remove_columns=ds.column_names)
# 定义 LoraConfig
config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    inference_mode=False, # 训练模式
    r=8, # Lora 秩
    lora_alpha=32, # Lora alaph，具体作用参见 Lora 原理
    lora_dropout=0.1# Dropout 比例
)
model = get_peft_model(model, config)
model.print_trainable_parameters()
# 定义 TrainingArguments 参数
args = TrainingArguments(
    output_dir='./output/llama3',
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    logging_steps=10,
    num_train_epochs=3,
    save_steps=100,
    learning_rate=1e-4,
    save_on_each_node=True,
    gradient_checkpointing=True
)
# 训练
trainer = Trainer(
    model=model,
    args=args,
    train_dataset=tokenized_id,
    data_collator=DataCollatorForSeq2Seq(tokenizer=tokenizer, padding=True),
)
trainer.train()
# 保存 lora 模型
lora_path='./llama3_lora'
trainer.model.save_pretrained(lora_path)
tokenizer.save_pretrained(lora_path)

[KMnO4-zx]

emmm 教程没有在colab环境下做过测试，你可以调试一下 看看bug在哪，也可以使用与教程相同的autodl环境

[chenmonster]

model = AutoModelForCausalLM.from_pretrained(model_dir, device_map='auto', torch_dtype=torch.bfloat16)

将上面的代码改为

model = AutoModelForCausalLM.from_pretrained(model_dir, device_map='cuda', torch_dtype=torch.half, trust_remote_code=True)

重新运行，报错：

OutOfMemoryError: CUDA out of memory.

[KMnO4-zx]

torch.half 是自动选择半精度加载，有些显卡不支持bf16，但全部显卡应该都支持fp16，这个应该不是导致oom的原因