【大模型系列】分布式训练相关技术记录

一，显存管理：

激活显存：nvidia-smi
失活显存：fuser -v /dev/nvidia*

二，分布式管理：

1，当前分布式使用deepspeed进行tensor并行

2，使用deepspeed时，不可以使用os.environ[“CUDA_VISIBLE_DEVICES”] = '0,1,2,3’来指定使用的GPU

使用os.environ[“CUDA_VISIBLE_DEVICES”]可以控制数量，但不能控制编号，编号始终从0开始

3，使用deepspeed时，要通过参数 deepspeed --include localhost:1 train.py 来指定

4，使用deepspeed时，当显存成为瓶颈，需要配置–deepspeed ds_config.json 参数，填写配置信息

ZeRO Stage 1: 划分optimizer states。优化器参数被划分到多个memory上，每个momoey上的进程只负责更新它自己那部分参数。
ZeRO Stage 2: 划分gradient。每个memory，只保留它分配到的optimizer state所对应的梯度。这很合理，因为梯度和optimizer是紧密联系在一起的。只知道梯度，不知道optimizer state，是没有办法优化模型参数的。
ZeRO Stage 3: 划分模型参数，或者说，不同的layer. ZeRO-3会在forward和backward的时候，自动将模型参数分配到多个memory。
由于ZeRO-1只分配optimizer states(参数量很小)，实际使用的时候，我们一般只会考虑ZeRO-2和ZeRO-3。
ZeRO-2配置示例：

{
    "bfloat16": {
        "enabled": "auto"
    },
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "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": "cpu",
            "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",
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "steps_per_print": 1e5
}

ZeRO-3配置示例：

{
    "bfloat16": {
        "enabled": false
    },
    "fp16": {
        "enabled": "auto",
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "initial_scale_power": 16,
        "hysteresis": 2,
        "min_loss_scale": 1
    },
    "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": 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_fp16_weights_on_model_save": true
    },
    "gradient_accumulation_steps": "auto",
    "gradient_clipping": "auto",
    "steps_per_print": 1e5,
    "train_batch_size": "auto",
    "train_micro_batch_size_per_gpu": "auto",
    "wall_clock_breakdown": false
}