How do I know if stage-3 is a success by using deepspeed？

[hwhyyds]

I used accelerate to packag the model with GPUs, but the model are copied in GPUs, not shard in GPUs. Why?

deepspeed_plugin = DeepSpeedPlugin(hf_ds_config="./configs/rl_ds_zero_3.json", zero_stage=3)

base_model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
accelerator = Accelerator(deepspeed_plugin=deepspeed_plugin, mixed_precision="bf16")
base_model= accelerator.prepare(base_model)

rl_ds_zero_3.json

{
  "train_micro_batch_size_per_gpu": 1,
  "model_parallel_size": 3,
  "zero_allow_untested_optimizer": true,
  "bf16": {
    "enabled": "auto"
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": 1e-5,
      "betas": [0.9, 0.999],
      "eps": 1e-8,
      "weight_decay": 0.01
    }
  },
  "zero_optimization": {
    "stage": 3,
    "allgather_partitions": true,
    "allgather_bucket_size": 5e8,
    "reduce_scatter": true,
    "contiguous_gradients": true,
    "overlap_comm": true,
    "sub_group_size": 1e9,
    "reduce_bucket_size": 5e8,
    "stage3_prefetch_bucket_size": 5e7,
    "stage3_param_persistence_threshold": 5e6,
    "stage3_max_live_parameters": 1e9,
    "stage3_max_reuse_distance": 1e9,
    "stage3_gather_16bit_weights_on_model_save": true,
    "offload_param": {
      "device": "cpu",
      "pin_memory": true
  }
  },
  "activation_checkpointing": {
    "partition_activations": true,
    "cpu_checkpointing": true,
    "number_checkpoints": 4,
    "synchronize_checkpoint_boundary": false,
    "contiguous_memory_optimization": true
  }
}

[tjruwase]

@hwhyyds, can you please share more steps to reproduce, including logs, scripts, and command line?