deepseek-glm4-grpo训练

一、目录

1.grpo 重新训练已经微调的glm4模型

二、实现

1.grpo 重新训练已经微调的glm4模型
1.1 指令：

 CUDA_VISIBLE_DEVICES=1 nohup python test.py --model_name_or_path /home/LLaMA-Factory/saves/glm4-9b-lora-alpaca_reference_train20250115_01_merge \
        --dataset_name /home/LLaMA-Factory/data/alpca_reference_train20250115_1.json \
        --learning_rate 5.0e-6 \
        --num_train_epochs 1   \
        --per_device_train_batch_size 6  \
        --per_device_eval_batch_size 6  \
        --num_generations 6 \
        --gradient_accumulation_steps 4 \
         --logging_steps 25 \
        --eval_strategy steps \
        --eval_steps 50 \
        --use_peft 1 \
        --lora_r 32 \
        --lora_alpha 16 \
        --output_dir /saves/glm4-9b-grpo >grop_output.log 2>&1 &
        
#注意： num_generations取值范围 : 2 ~ batch/2

1.2 遇到问题及解决

1. tokenizer no padding_side 字段
解决：脚本中添加该字段  padding_side: Optional[str] = None,
 def _pad(
            self,
            encoded_inputs: Union[Dict[str, EncodedInput], BatchEncoding],
            max_length: Optional[int] = None,
            padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
            pad_to_multiple_of: Optional[int] = None,
            padding_side: Optional[str] = None,
            return_attention_mask: Optional[bool] = None,
    ) -> dict:
2. model no logits_to_keep 字段
修改模型脚本，进行添加该字段，以及相关功能。
 def forward(
            self,
            input_ids: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.Tensor] = None,
            attention_mask: Optional[torch.Tensor] = None,
            past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            labels: Optional[torch.Tensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            return_last_logit: Optional[bool] = False,
            logits_to_keep: int = 0
    ):
        lm_logits = self.transformer.output_layer(hidden_states[:, -logits_to_keep:, :])

1.3 脚本

#coding="utf8"
import json
import argparse
from typing import Optional
from dataclasses import dataclass, field
from transformers import AutoModelForCausalLM, AutoTokenizer
from torch.utils.data import Dataset
from trl import GRPOConfig, GRPOTrainer, ModelConfig, ScriptArguments, TrlParser, get_peft_config


@dataclass
class GRPOScriptArguments(ScriptArguments):
    """
    Script arguments for the GRPO training script.

    Args:
        reward_model_name_or_path (`str` or `None`):
            Reward model id of a pretrained model hosted inside a model repo on huggingface.co or local path to a
            directory containing model weights saved using [`~transformers.PreTrainedModel.save_pretrained`].
    """

    reward_model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": "Reward model id of a pretrained model hosted inside a model repo on huggingface.co or "
            "local path to a directory containing model weights saved using `PreTrainedModel.save_pretrained`."
        },
    )


class MyDataset(Dataset):
    def __init__(self, dataset):
        self.dataset = dataset
        

    def __len__(self):
        return len(self.dataset)

    def __getitem__(self, idx):
        text = self.dataset[idx]["instruction"]
        return {"prompt": text}

def get_dataset(path):
    import json
    with open(path, 'r', encoding='utf-8') as f:
        data = json.load(f)
    dataset = MyDataset(data[:-200])
    evaldataset = MyDataset(data[-200:])
    return dataset, evaldataset


def main(script_args, training_args, model_args):
    # Load a pretrained model
    print(model_args.model_name_or_path)
    model = AutoModelForCausalLM.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=True
    )
    tokenizer = AutoTokenizer.from_pretrained(
        model_args.model_name_or_path, trust_remote_code=True
    )

    def reward_len(completions, **kwargs):
        #输出奖励
        data = []
        for completion in completions:
           
            try:
                completion = json.loads(completion)
                data.append(1.0)
            except:
                
                data.append(0.0)
         
        return data
    
    # Load the dataset
    dataset, evaldataset = get_dataset(script_args.dataset_name)

    # Initialize the GRPO trainer
    trainer = GRPOTrainer(
        model=model,
        reward_funcs = reward_len,
        args=training_args,
        train_dataset = dataset,
        eval_dataset = evaldataset if training_args.eval_strategy != "no" else None,
        processing_class=tokenizer,
        peft_config=get_peft_config(model_args),
    )

    # Train and push the model to the Hub
    trainer.train()

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    # if training_args.push_to_hub:
    #     trainer.push_to_hub(dataset_name=script_args.dataset_name)


def make_parser(subparsers: argparse._SubParsersAction = None):
    dataclass_types = (GRPOScriptArguments, GRPOConfig, ModelConfig)
    if subparsers is not None:
        parser = subparsers.add_parser("grpo", help="Run the GRPO training script", dataclass_types=dataclass_types)
    else:
        parser = TrlParser(dataclass_types)
    return parser


if __name__ == "__main__":
    parser = make_parser()
    script_args, training_args, model_args = parser.parse_args_and_config()
    main(script_args, training_args, model_args)