How to load the model and the checkpoint after trained the model?

[ccwdb]

I trained my model using the code in the sft_trainer.py. And I save the checkpoint and the model in the same dir.
But I don't know how to load the model with the checkpoint. Or I just want to konw that trainer.save_model(script_args.output_dir) means I have save a trained model, not just a checkpoint?
I try many ways to load the trained model but errors like

RuntimeError: Error(s) in loading state_dict for PrefixEncoder:
	Missing key(s) in state_dict: "embedding.weight". 

So, how to load the model???

[lvwerra]

Are you using PEFT for fine-tuning? And can you the code you are using to load the model?

[ccwdb]

Are you using PEFT for fine-tuning? And can you the code you are using to load the model?

Following are my code.

from datasets import load_dataset
from trl import SFTTrainer
from transformers import AutoModel, DataCollatorForLanguageModeling, AutoTokenizer, TrainingArguments, AutoModelForCausalLM
from peft import LoraConfig

# 加载模型和tokenizer
MODEL_PATH = "/home/qiji/chatglm2-6b"
model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, trust_remote_code=True).half().cuda()
# model = AutoModel.from_pretrained(MODEL_PATH, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
# tokenizer.padding_side = 'right'
# 设置微调参数
training_arguments = TrainingArguments(
    output_dir='/home/qiji/Container/jinkundong/SFT/results',
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    save_steps=5000,
    logging_steps=1000,
    learning_rate=2e-4,
    fp16=True,
    max_grad_norm=0.3,
    max_steps=5000,
    warmup_ratio=0.03,
    group_by_length=True,
    lr_scheduler_type='constant',
)
model.config.use_cache = False

peft_config = LoraConfig(
    r=64,
    lora_alpha=16,
    lora_dropout=0.1,
    bias="none",
    task_type="CAUSAL_LM",
)

data_collator = DataCollatorForLanguageModeling(
    tokenizer=tokenizer,
    mlm=False,
)
no_deprecation_warning=True
dataset = load_dataset("/home/qiji/Container/jinkundong/SFT/SFT_dataset", split="train")
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    dataset_text_field="input",
    max_seq_length=512,
    peft_config=peft_config,
    args=training_arguments,
    data_collator=data_collator,
    packing=False,
)

trainer.train()

model_save_path = "/home/qiji/Container/jinkundong/SFT_2"
trainer.save_model(model_save_path)

As I konw, output_dir is the path that save the checkpoint and output_dir save the model. But I don't know what they can do .

[younesbelkada]

hi @ccwdb
Can you try to run (after training)

import torch
from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(output_dir, torch_dtype=torch.float16)

Make sure to install peft>=0.4.0

[ccwdb]

hi @ccwdb Can you try to run (after training)

import torch
from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(output_dir, torch_dtype=torch.float16)

Make sure to install peft>=0.4.0

I tried your code, but it says that I didn't have a file named config.json.
Here is my file.

I don't konw the difference between pytorch_model,bin and adapter_model.bin

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

[louieworth]

After checking the source code. I think you should load the model from:

from transformers.trainer_utils import EvalPrediction, get_last_checkpoint

last_checkpoint = get_last_checkpoint(script_args.output_dir)
trainer.train(resume_from_checkpoint=last_checkpoint)

Hello, may you provide how to save the model? Can I directly use

trainer.save