zero3_bf16_moe_offload.json

{
  "train_batch_size": 8,
  "bf16": { "enabled": true },

  "zero_optimization": {
    "stage": 3,
    "contiguous_gradients": true,
    "overlap_comm": true,
    "reduce_bucket_size": "auto",
    "sub_group_size": 1000000000,
    "offload_param":     { "device": "cpu", "pin_memory": true },
    "offload_optimizer": { "device": "cpu", "pin_memory": true },
    "stage3_param_persistence_threshold": 100000
  },

  "activation_checkpointing": {
    "partition_activations": true,
    "cpu_checkpointing": true,
    "contiguous_memory_optimization": true
  },

  "moe": { "ep_size": 8 }
}