Qwen2VL+Lora微调后合并模型丢失文件bug

[gxlover0625]

Reminder

• I have read the README and searched the existing issues.

System Info

[2024-10-19 11:31:58,444] [INFO] [real_accelerator.py:219:get_accelerator] Setting ds_accelerator to cuda (auto detect)

• llamafactory version: 0.9.1.dev0
• Platform: Linux-5.10.134-010.ali5000.al8.x86_64-x86_64-with-glibc2.32
• Python version: 3.10.15
• PyTorch version: 2.3.1+cu118 (GPU)
• Transformers version: 4.45.2
• Datasets version: 2.21.0
• Accelerate version: 0.34.2
• PEFT version: 0.12.0
• TRL version: 0.9.6
• GPU type: NVIDIA H20
• DeepSpeed version: 0.15.2

Reproduction

按照官网执行Qwen2VL+Lora微调代码

llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
llamafactory-cli export examples/merge_lora/qwen2vl_lora_sft.yaml

其中的qwen2vl_lora_sft.yaml如下

### model
model_name_or_path: /home/admin/workspace/aop_lab/llm/qwen/Qwen2-VL-7B-Instruct

### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all

### dataset
dataset: paco_all  # video: mllm_video_demo
template: qwen2_vl
cutoff_len: 1024
max_samples: 10000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: /home/admin/workspace/aop_lab/sft_results/paco_all/qwen2_vl-7b/lora/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 8
gradient_accumulation_steps: 2
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000

### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

qwen2vl_lora_sft.yaml如下

### Note: DO NOT use quantized model or quantization_bit when merging lora adapters

### model
model_name_or_path: /home/admin/workspace/aop_lab/llm/qwen/Qwen2-VL-7B-Instruct
adapter_name_or_path: /home/admin/workspace/aop_lab/sft_results/paco_all/qwen2_vl-7b/lora/sft
template: qwen2_vl
finetuning_type: lora

### export
export_dir: /home/admin/workspace/aop_lab/save_models/paco_all/qwen2_vl_lora_sft
export_size: 2
export_device: cpu
export_legacy_format: false

Expected behavior

Bug

微调过程没有出错，在合并lora的时候也没有报错，但是观察原始文件和微调模型文件发现丢失chat_template.json，具体如下

• 这是原始模型文件夹
  
• 这是微调后模型文件夹
  
  经过两张图的对比，发现合并后的文件夹中丢失了以下文件
• chat_template.json
• configuration.json
  多了以下文件
• added_tokens.json
• speicial_tokens_map.json

其中的chat_template.json是很重要的文件，模型缺失了这个就无法推理报错
官方推理的推理代码

from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
model_dir = snapshot_download("qwen/Qwen2-VL-7B-Instruct")

model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_dir, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_dir)
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

报错如下

ValueError: No chat template is set for this processor. Please either set the `chat_template` attribute, or provide a chat template as an argument. See https://huggingface.co/docs/transformers/main/en/chat_templating for more information

Others

额外信息

1. Qwen2VL是7B模型，在modelscope官网下的，使用的是以下代码

from modelscope import snapshot_download
model_dir = snapshot_download("qwen/Qwen2-VL-7B-Instruct")

[gxlover0625]

哪怕我把chat_template.json复制到了微调后模型后依然报错，怀疑是整个Processor保存有问题，只能通过processor.chat_template=xx强行设定才行

[aliencaocao]

同样问题。现在只能用预训练的qwen/Qwen2-VL-7B-Instruct的processor，然后weight用自己的local file

[gxlover0625]

同样问题。现在只能用预训练的qwen/Qwen2-VL-7B-Instruct的processor，然后weight用自己的local file

我现在用的方法如下，运行后看起来没问题，不过不知道对不对，是否正确加载了微调后的权重。

processor = AutoProcessor.from_pretrained("sft_dir")
processor.chat_template = "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"

[aliencaocao]

这样也可以，你这个和权重没关系，权重是用automodel加载

[gxlover0625]

这样也可以，你这个和权重没关系，权重是用automodel加载

好的感谢

[aliencaocao]

感觉你还是open比较好，这个是一个明显的bug

[Wiselnn570]

same issue

[Syazvinski]

same issue

[gxlover0625]

same issue

It has been fixed，and i have try new version of llamafactory for two-stage sft and there is no problems

[PangziZhang523]

用lora训练完的模型，merge以后用官方代码推理：processor = AutoProcessor.from_pretrained(model_path) 还是报错Exception: data did not match any variant of untagged enum ModelWrapper at line 757371 column 3

[gxlover0625]

用lora训练完的模型，merge以后用官方代码推理：processor = AutoProcessor.from_pretrained(model_path) 还是报错Exception: data did not match any variant of untagged enum ModelWrapper at line 757371 column 3

我也碰到过这个问题，后来按照qwen2vl官网的安装引导重新开了一个环境，好像就没这个问题了。我的transformers版本是4.45.0，不知道能不能解决您的问题

[PangziZhang523]

感谢回复，我的transformers版本也是4.45.0.dev0，用官方代码测试官方的模型正常推理，把模型换成lora微调后的模型就出现这个问题了，看报错是这句processor = AutoProcessor.from_pretrained(model_path)出现的问题。同时看合并以后的文件是有chat_template.json的

[PangziZhang523]

llamafactory-cli webchat examples/inference/qwen2_vl.yaml是正常的吗？为什么我的会是这样的页面：

[binhoul]

用lora训练完的模型，merge以后用官方代码推理：processor = AutoProcessor.from_pretrained(model_path) 还是报错Exception: data did not match any variant of untagged enum ModelWrapper at line 757371 column 3

请问这个解决了吗？