AI开发平台MODELARTS-迁移适配:多卡分布式执行

时间：2024-09-24 10:50:21

AI开发平台MODELARTS

多卡分布式执行

PyTorch框架下常见的多卡分布式执行主要包括DataParallel（DP）和Distributed Data Parallel （DDP）。torch_npu环境下针对DDP场景的多卡训练有提供支持，具体请参见迁移单卡脚本为多卡脚本。此外，针对deepspeed环境，昇腾有专门的适配环境deepspeed-npu。在此提供一种基于deepspeed的多卡训练脚本，内容如下：

# ds_run_npu.sh
LR=1e-4
TRAIN_FILE=${HOME}/YeungNLPfirefly-train-1.1M/firefly-train-1.1M.jsonl
PER_DEVICE_TRAIN_BATCH_SIZE=4
GRADIENT_ACCUMULATION_STEPS=32
MODEL_DIR=${HOME}/chatglm
OUTPUT_DIR=${HOME}/ChatGLM-6B-main/ptuning/output
DS_CONFIG=${HOME}/ChatGLM-6B-main/ptuning/ds_config.json
APP_SCRIPT=${HOME}/ChatGLM-6B-main/ptuning/main.py
MASTER_PORT=$(shuf -n 1 -i 10000-65535)
 
deepspeed --num_gpus=8 --master_port $MASTER_PORT ${APP_SCRIPT} \
        --deepspeed ${DS_CONFIG} \
        --log_level debug \
        --model_name_or_path ${MODEL_DIR} \
        --train_file ${TRAIN_FILE} \
        --prompt_column input \
        --response_column target \
        --max_source_length 512 \
        --max_target_length 512 \
        --output_dir ${OUTPUT_DIR}/chatglm-6b-${LR} \
        --per_device_train_batch_size ${PER_DEVICE_TRAIN_BATCH_SIZE} \
        --per_device_eval_batch_size 1 \
        --gradient_accumulation_steps ${GRADIENT_ACCUMULATION_STEPS} \
        --gradient_checkpointing False \
        --num_train_epochs 1 \
        --predict_with_generate \
        --logging_steps 10 \
        --save_strategy "steps" \
        --save_total_limit 3 \
        --learning_rate $LR \
        --dataloader_num_workers 60 \
        --preprocessing_num_workers 60 \
        --do_train \
        --overwrite_output_dir \
        --max_steps 100 \
        --fp16

LR、PER_DEVICE_TRAIN_BATCH_SIZE、GRADIENT_ACCUMULATION_STEPS分别代表学习率、单个设备训练批次大小、梯度累计步数，作为超参数可以调优获得较好模型。同样，${HOME} 需要根据数据集模型等路径做对应替换，这里脚本适配的数据集是Firefly，其中deepspeed使用了zero 1显存优化方式，配置方式如下：

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "bf16": {
        "enabled": "auto"
    },
    "optimizer": {
        "type": "AdamW",
        "params": {
            "lr": "auto",
            "betas": "auto",
            "eps": "auto",
            "weight_decay": "auto"
        }
    },
    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },
    "zero_optimization": {
        "stage": 1,
        "offload_optimizer": {
            "device": "cpu",
            "pin_memory": true
        },
        "contiguous_gradients": true,
        "overlap_comm": true,
        "allgather_partitions": true,
        "reduce_scatter": true,
        "allgather_bucket_size": 2e8,
        "reduce_bucket_size": 4e8
    },
    "flops_profiler":{
        "enabled": true,
        "profile_step": 1,
        "module_path": -1,
        "top_modules": 1,
        "detailed": false,
        "output_file": null
    },
    "zero_allow_untested_optimizer": "true",
    "gradient_clipping": "auto",
    "gradient_accumulation_steps": "auto",
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

上一篇：AI开发平台MODELARTS-迁移适配:自动迁移适配

下一篇：AI开发平台MODELARTS-迁移适配:单卡方式训练