[BUG] 加载 LoRA 权重报错

[pixelock]

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

LoRA SFT 得到 adapter 权重后, 使用下面的方法加载权重时报错:

from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(
    # path to the output directory
    'output/minicpmv2_lora',
    device_map="auto",
    trust_remote_code=True
).eval()

错误:

Traceback (most recent call last):
  File "/MiniCPM-V/minicpm_infer.py", line 17, in <module>
    model = AutoPeftModelForCausalLM.from_pretrained(
  File "/usr/local/lib/python3.10/dist-packages/peft/auto.py", line 126, in from_pretrained
    base_model.resize_token_embeddings(len(tokenizer))
  File "/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py", line 1901, in resize_token_embeddings
    model_embeds = self._resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
  File "/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py", line 1915, in _resize_token_embeddings
    old_embeddings = self.get_input_embeddings()
  File "/usr/local/lib/python3.10/dist-packages/transformers/modeling_utils.py", line 1668, in get_input_embeddings
    raise NotImplementedError
NotImplementedError

表现为缺少 get_input_embeddings 函数. 看 finetune.py 中, get_input_embeddings 方法是动态加入的, 是不是这里存在问题?

期望行为 | Expected Behavior

LoRA 权重可以正常加载, 不报错.

复现方法 | Steps To Reproduce

1. 训练得到 lora adapter 权重
2. 使用 peft 中的 AutoPeftModelForCausalLM 方法加载 adapter 目录, 得到完整的模型
3. 在这一步会遇到 get_input_embeddings 的 NotImplementedError 错误

运行环境 | Environment

- OS:
- Python: 3.11
- Transformers: 4.40.0
- PyTorch: 2.2.0
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`): 12.3

备注 | Anything else?

No response

[isongxw]

hugging face的最新版代码更新了 modeling_minicpmv.py，把get_input_embeddings的实现加进去了，亲测更新后lora模型加载就不会报这个错误了

[pixelock]

hugging face的最新版代码更新了 modeling_minicpmv.py，把get_input_embeddings的实现加进去了，亲测更新后lora模型加载就不会报这个错误了

官方已解决, 更新测试没有问题了