Deepspeed的zero2和zero3的配置文件Demo

小怪兽会微笑

已于 2025-02-13 00:34:30 修改

阅读量491

点赞数 6

分类专栏：学习笔记 Deep Learning 文章标签：人工智能机器学习深度学习

于 2025-02-09 16:44:54 首次发布

本文链接：https://blog.csdn.net/a1920993165/article/details/145533734

版权

学习笔记同时被 2 个专栏收录

140 篇文章

订阅专栏

Deep Learning

39 篇文章

订阅专栏

Deepspeed的zero2和zero3的配置文件Demo

Zero2

使用下面这个文件，在8*H100的机器上训练了Qwen2.5-7B模型。上下文窗口8K，参数大小bf16，batch_size = 8（num_gpu）*2（per_gpu_batch_size）*4(gradient_accumulation_steps) =64，deepspeed采用zero2，显存占用约78G，数据量1.4B tokens，训练时长8 hour。

{
    "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": 2,
        "offload_optimizer": {
            "device": "none",
            "pin_memory": true
        },
        "allgather_partitions": true,
        "allgather_bucket_size": 2e8,
        "overlap_comm": true,
        "reduce_scatter": true,
        "reduce_bucket_size": 2e8,
        "contiguous_gradients": true
    },

    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 100,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}

Zero3

使用下面这个文件，在8*H100的机器上训练了Qwen2.5-32B模型。上下文窗口8K，参数大小bf16，batch_size = 8（num_gpu）*2（per_gpu_batch_size）8(gradient_accumulation_steps) =128，deepspeed采用zero3，显存占用约65G，内存占用约653G（还有数据占的少量显存，数据量为0.06B tokens），训练时长2 hour。
使用下面这个文件，在8H100的机器上训练了Qwen2.5-3B模型。上下文窗口32K，参数大小bf16，batch_size = 8（num_gpu）*1（per_gpu_batch_size）*1(gradient_accumulation_steps) =8，deepspeed采用zero3，显存占用约79G，内存占用约91G（还有数据占的少量显存，数据量为0.06B tokens），训练时长1 hour。

{
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },

    "bf16": {
        "enabled": "auto"
    },

    "scheduler": {
        "type": "WarmupLR",
        "params": {
            "warmup_min_lr": "auto",
            "warmup_max_lr": "auto",
            "warmup_num_steps": "auto"
        }
    },

    "zero_optimization": {
        "stage": 3,
        "offload_optimizer": {
            "device": "cpu"
        },
        "offload_param": {
            "device": "cpu"
        },
        "overlap_comm": true,
        "stage3_gather_16bit_weights_on_model_save": true
    },

    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 100,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}