题意：尝试在一个名为 DummyOptim 的对象上调用 .step() 方法，但是这个对象并没有定义这个方法

问题背景：

I want to use deepspeed for training LLMs along with Huggingface Trainer. But when I use deepspeed along with trainer I get error "AttributeError: 'DummyOptim' object has no attribute 'step'". Below is my code

import argparse
import numpy as np
import torch
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM

from trl import DPOTrainer, DPOConfig
def preprocess_data(item):
    return {
        'prompt': 'Instruct: ' + item['prompt'] + '\n',
        'chosen': 'Output: ' + item['chosen'],
        'rejected': 'Output: ' + item['rejected']
    }        

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--epochs", type=int, default=1)
    parser.add_argument("--beta", type=float, default=0.1)
    parser.add_argument("--batch_size", type=int, default=4)
    parser.add_argument("--lr", type=float, default=1e-6)
    parser.add_argument("--seed", type=int, default=2003)
    parser.add_argument("--model_name", type=str, default="EleutherAI/pythia-14m")
    parser.add_argument("--dataset_name", type=str, default="jondurbin/truthy-dpo-v0.1")
    parser.add_argument("--local_rank", type=int, default=0)

    args = parser.parse_args()

    # Determine device based on local_rank
    device = torch.device("cuda", args.local_rank) if torch.cuda.is_available() else torch.device("cpu")


    tokenizer = AutoTokenizer.from_pretrained(args.model_name)
    tokenizer.pad_token = tokenizer.eos_token
    model = AutoModelForCausalLM.from_pretrained(args.model_name).to(device)
    ref_model = AutoModelForCausalLM.from_pretrained(args.model_name).to(device)

    dataset = load_dataset(args.dataset_name, split="train")
    dataset = dataset.map(preprocess_data)

    # Split the dataset into training and validation sets
    dataset = dataset.train_test_split(test_size=0.1, seed=args.seed)
    train_dataset = dataset['train']
    val_dataset = dataset['test']

    training_args = DPOConfig(
        learning_rate=args.lr,
        num_train_epochs=args.epochs,
        per_device_train_batch_size=args.batch_size,
        logging_steps=10,
        remove_unused_columns=False,
        max_length=1024,
        max_prompt_length=512,
        deepspeed="ds_config.json"       
    )

    

    # Verify and print embedding dimensions before finetuning
    print("Base model embedding dimension:", model.config.hidden_size)

    model.train()
    ref_model.eval()

    dpo_trainer = DPOTrainer(
        model,
        ref_model,
        beta=args.beta,
        train_dataset=train_dataset,
        eval_dataset=val_dataset,
        tokenizer=tokenizer,
        args=training_args,
    )

    dpo_trainer.train()
    # Evaluate
    evaluation_results = dpo_trainer.evaluate()
    print("Evaluation Results:", evaluation_results)

    save_model_name = 'finetuned_model'
    model.save_pretrained(save_model_name)

if __name__ == "__main__":
    main()

The config file used is the below one

{
"zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "offload_param": {
            "device": "cpu",
            "pin_memory": true
        },
        "overlap_comm": true,
        "contiguous_gradients": true,
        "sub_group_size": 1e9,
        "reduce_bucket_size": "auto",
        "stage3_prefetch_bucket_size": "auto",
        "stage3_param_persistence_threshold": "auto",
        "stage3_max_live_parameters": 1e9,
        "stage3_max_reuse_distance": 1e9,
        "stage3_gather_16bit_weights_on_model_save": true
    },
"bf16": {
    "enabled": "auto"
},
"fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "initial_scale_power": 32,
    "loss_scale_window": 1000,
    "hysteresis": 2,
    "min_loss_scale": 1
},

"gradient_accumulation_steps": "auto",
"gradient_clipping": "auto",
"train_batch_size": "auto",
"train_micro_batch_size_per_gpu": "auto",
"wall_clock_breakdown": false,
"flops_profiler": {
    "enabled": false,
    "detailed": false
},
"optimizer": {
    "type": "Lamb",
    "params": {
    "lr": "auto",
    "betas": [0.9, 0.999],
    "eps": "auto",
    "weight_decay": "auto"
    }
},
"zero_allow_untested_optimizer": true
}

The code works with out deepspeed. I have torch=2.3.1, deepspeed =0.14.5, trl=0.9.4 and CUDA Version: 12.5.

Appreciate any hint on this !

问题解决：

from accelerate.utils import DistributedType

training_args.distributed_state.distributed_type = DistributedType.DEEPSPEED

adding this solves the issue