PEFT LoRA适配器模型与基础模型合并报错求助

尝试将PEFT LoRA适配器模型ArcturusAI/Crystalline-1.1B-v23.12-tagger与基础模型TinyLlama/TinyLlama-1.1B-Chat-v0.6合并为完整模型时，运行以下代码出现错误：

from transformers import AutoModel, AutoConfig

# Load the pretrained model and LoRA adapter
pretrained_model_name = "TinyLlama/TinyLlama-1.1B-Chat-v0.6"

pretrained_model = AutoModel.from_pretrained(pretrained_model_name)
lora_adapter = AutoModel.from_pretrained("ArcturusAI/Crystalline-1.1B-v23.12-tagger")

# Assuming the models have the same architecture (encoder, decoder, etc.)
# Get the weights of each model
pretrained_weights = pretrained_model.state_dict()
lora_adapter_weights = lora_adapter.state_dict()

# Combine the weights (adjust the weights based on your preference)
combined_weights = {}
for key in pretrained_weights:
    combined_weights[key] = 0.8 * pretrained_weights[key] + 0.2 * lora_adapter_weights[key]

# Load the combined weights into the pretrained model
pretrained_model.load_state_dict(combined_weights)

# Save the integrated model
pretrained_model.save_pretrained("ArcturusAI/Crystalline-1.1B-v23.12-tagger-fullmodel")

错误信息：

OSError: ArcturusAI/Crystalline-1.1B-v23.12-tagger does not appear to have a file named pytorch_model.bin, tf_model.h5, model.ckpt or flax_model.msgpack.

错误原因 #

1. LoRA模型加载方式错误：AutoModel.from_pretrained()仅支持加载完整Transformer模型的权重文件，而LoRA适配器仅包含增量权重（如lora_A、lora_B这类参数），没有完整模型的权重文件，因此触发找不到文件的报错。
2. 权重合并逻辑错误：LoRA并非通过简单加权基础模型与适配器权重实现合并，而是需要将LoRA的增量权重注入到基础模型的对应层结构中。

正确解决方案 #

使用PEFT库提供的原生工具完成LoRA与基础模型的合并，步骤如下：

1. 安装依赖库 #

确保安装所需工具库：

pip install transformers peft torch accelerate

2. 合并模型代码 #

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

# 加载基础模型与配套Tokenizer
base_model_name = "TinyLlama/TinyLlama-1.1B-Chat-v0.6"
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    device_map="auto",
    torch_dtype="auto"
)
tokenizer = AutoTokenizer.from_pretrained(base_model_name)

# 加载LoRA适配器并合并到基础模型
lora_model_name = "ArcturusAI/Crystalline-1.1B-v23.12-tagger"
merged_model = PeftModel.from_pretrained(base_model, lora_model_name)
merged_model = merged_model.merge_and_unload()

# 保存合并后的完整模型与Tokenizer
merged_model.save_pretrained("Crystalline-1.1B-v23.12-tagger-fullmodel")
tokenizer.save_pretrained("Crystalline-1.1B-v23.12-tagger-fullmodel")

代码说明 #

• AutoModelForCausalLM：针对对话类模型（如TinyLlama Chat）使用因果语言模型类加载，比通用AutoModel更适配任务场景。
• PeftModel.from_pretrained()：正确关联基础模型与LoRA适配器，自动识别适配器的权重结构。
• merge_and_unload()：将LoRA增量权重注入基础模型的对应层，生成无适配器依赖的完整模型。
• 同步保存Tokenizer：确保后续使用模型时，分词逻辑与训练阶段一致。

内容的提问来源于stack exchange，提问作者arcturusvx