H200全量训练DeepSeek-R1-Distill-Llama-70B，采用zero3(batch_size=1)时溢出。采用zero3+offload（优化器、参数）时，显存占用少（141G显存只占用27G)，cpu占用高

[github-eliviate]

Reminder

• I have read the above rules and searched the existing issues.

System Info

llama-factory版本：0.8.3.dev0
ubuntu:22.04
python: python3.10
torch: 2.5.1

Reproduction

Put your message here.

Others

No response

[github-eliviate]

H200(8*141G)全量训练DeepSeek-R1-Distill-Llama-70B

1. 采用zero3(batch_size=1)时溢出。
2. 采用zero3+offload优化器和参数（batch_size=1），显存占用少（141G显存只占用27G)，cpu占用高
3. 增加batch_size=64，仍采用zero3+offload优化器和参数，显存占用不变（141G显存只占用27G)，cpu占用增加

配置文件examples/train_full/llama3_full_sft_ds3.yaml如下：

### model
model_name_or_path: /disk2/liweichao/DeepSeek-R1-Distill-Llama-70B

### method
stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_offload_config.json

### dataset
dataset: identity
template: llama3
cutoff_len: 1024
# max_samples: 1000
overwrite_cache: true
preprocessing_num_workers: 16

### output
output_dir: saves/llama3-70B/full/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

### train
per_device_train_batch_size: 1
gradient_accumulation_steps: 1
learning_rate: 1.0e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000


### eval
val_size: 0.1
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

deepspeed参数（examples/deepspeed/ds_z3_offload_config.json ）：

{
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "bf16": {
    "enabled": "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_16bit_weights_on_model_save": true
  }
}

训练命令：

FORCE_TORCHRUN=1 torchrun --nnodes 1 --node_rank 0 --nproc_per_node 8 --master_addr 127.0.0.1 --master_port 20004 src/train.py  examples/train_full/llama3_full_sft_ds3.yaml