merge模型

[arbani369]

感谢您的微调脚本，还想请问一下，我已经训练完成，如何merge成一个完整的模型呢，感谢！

[xiez22]

你好！我还没有尝试过merge ChatTS模型，你可以尝试按照https://huggingface.co/docs/peft/developer_guides/lora#merge-lora-weights-into-the-base-model 里面的方法来试试看

[arbani369]

我尝试用merge模型的方式和直接加载lora权重的方式推理，但是都没有看到微调效果，下边是我merge模型的代码
`import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

def merge_and_save_model(base_model_path, lora_path, output_path):
# 1. 加载基础模型和tokenizer
print(f"Loading base model from {base_model_path}")
base_model = AutoModelForCausalLM.from_pretrained(
base_model_path,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(
base_model_path,
trust_remote_code=True
)

# 2. 加载LoRA适配器
print(f"Loading LoRA adapter from {lora_path}")
lora_model = PeftModel.from_pretrained(
    base_model,
    lora_path,
    torch_dtype=torch.float16
)

# 3. 合并权重并卸载适配器
print("Merging LoRA weights with base model")
merged_model = lora_model.merge_and_unload()

# 4. 保存合并后的模型
print(f"Saving merged model to {output_path}")
merged_model.save_pretrained(output_path)
tokenizer.save_pretrained(output_path)
print("Merge completed successfully!")`

下面是我直接加载lora权重推理的代码
base_model_name = "/mnt/data/cg/code/space_mllm/ckpts/ChatTS-14B" lora_adapter_name = "/mnt/data/cg/code/space_mllm/ChatTS-Training/ChatTS-Training-main/output/0826_test_print/checkpoint-600" base_model = AutoModelForCausalLM.from_pretrained(base_model_name,trust_remote_code=True, device_map=0, torch_dtype='float16') tokenizer = AutoTokenizer.from_pretrained(lora_adapter_name,trust_remote_code=True) processor = AutoProcessor.from_pretrained(lora_adapter_name, trust_remote_code=True,tokenizer=tokenizer) model = PeftModel.from_pretrained(base_model, lora_adapter_name) model.eval()

输出的结果都像是没有微调过的原始模型的输出，但我在训练过程中打印中间输出是符合微调数据集的输出的，loss也在下降，请问这个问题您有什么见解嘛，感谢！

[Alan-2018]

我尝试用merge模型的方式和直接加载lora权重的方式推理，但是都没有看到微调效果，下边是我merge模型的代码 `import torch from peft import PeftModel from transformers import AutoModelForCausalLM, AutoTokenizer

def merge_and_save_model(base_model_path, lora_path, output_path): # 1. 加载基础模型和tokenizer print(f"Loading base model from {base_model_path}") base_model = AutoModelForCausalLM.from_pretrained( base_model_path, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True ) tokenizer = AutoTokenizer.from_pretrained( base_model_path, trust_remote_code=True )

# 2. 加载LoRA适配器
print(f"Loading LoRA adapter from {lora_path}")
lora_model = PeftModel.from_pretrained(
    base_model,
    lora_path,
    torch_dtype=torch.float16
)

# 3. 合并权重并卸载适配器
print("Merging LoRA weights with base model")
merged_model = lora_model.merge_and_unload()

# 4. 保存合并后的模型
print(f"Saving merged model to {output_path}")
merged_model.save_pretrained(output_path)
tokenizer.save_pretrained(output_path)
print("Merge completed successfully!")`

下面是我直接加载lora权重推理的代码 base_model_name = "/mnt/data/cg/code/space_mllm/ckpts/ChatTS-14B" lora_adapter_name = "/mnt/data/cg/code/space_mllm/ChatTS-Training/ChatTS-Training-main/output/0826_test_print/checkpoint-600" base_model = AutoModelForCausalLM.from_pretrained(base_model_name,trust_remote_code=True, device_map=0, torch_dtype='float16') tokenizer = AutoTokenizer.from_pretrained(lora_adapter_name,trust_remote_code=True) processor = AutoProcessor.from_pretrained(lora_adapter_name, trust_remote_code=True,tokenizer=tokenizer) model = PeftModel.from_pretrained(base_model, lora_adapter_name) model.eval()

输出的结果都像是没有微调过的原始模型的输出，但我在训练过程中打印中间输出是符合微调数据集的输出的，loss也在下降，请问这个问题您有什么见解嘛，感谢！

Hi；想请教一下；这个问题解决了嘛？