使用p-tuning、adapter、qlora对llama7b进行微调，loss相差比较大，会不会是因为sft不支持p_tuning和daapter导致的

[PowerDispatch]

adapter和p-tuning的代码如下： `import os os.environ['CUDA_VISIBLE_DEVICES'] = '3'

os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128'

from modelscope import Model, AutoModelForSequenceClassification, AutoTokenizer, MsDataset from swift import Swift, LoRAConfig, AdapterConfig, Trainer, TrainingArguments, PromptConfig, PrefixTuningConfig import torch torch.cuda.empty_cache() from transformers import default_data_collator import conf as co

model = Model.from_pretrained(co.base_model_path, torch_dtype=torch.bfloat16) adapter_config = AdapterConfig( dim=model.config.hidden_size, target_modules=['mlp'], method_name='forward', hidden_pos=0, adapter_length=32,

        )

prompt_config = PromptConfig(

dim=768, # hidden states的维度

target_modules=r'.*layer.\d+$', # 要使用正则表达式替换的模块

embedding_pos=0, # embedding张量的位置

prompt_length=10, # 提示符token的长度

attach_front=False, # 是否将提示符附加在embedding前面

)

prompt_config = PrefixTuningConfig( num_virtual_tokens=20, # hidden states的维度 task_type='CAUSAL_LM' )

if co.train_type == 2: train_type = adapter_config else: train_type = prompt_config

model = Swift.prepare_model(model, train_type) tokenizer = AutoTokenizer.from_pretrained('/tf/model/Llama-2-7b-chat-ms/')

train_dataset = MsDataset.load(dataset_name = '/tf/user/lgq/data',data_files={'train': co.train_data_path}).to_hf_dataset() val_dataset = MsDataset.load(dataset_name = '/tf/user/lgq/data',data_files={'val': co.val_data_path}).to_hf_dataset().select(range(4))

def tokenize_function(examples): return tokenizer(examples['instruction'], examples['input'],examples['output'], padding='max_length', truncation=True, max_length=4096)

train_dataset = train_dataset.map(tokenize_function, batched=True) val_dataset = val_dataset.map(tokenize_function, batched=True)

arguments = TrainingArguments( output_dir= co.output_dir, per_device_train_batch_size=1, learning_rate=4.5e-07 eval_steps=50

)

trainer = Trainer(model, arguments, train_dataset=train_dataset, eval_dataset=val_dataset, data_collator=default_data_collator,) trainer.train()`