projects/llava_sam2/deepspeed_zero2_sam2.json

{
  "gradient_accumulation_steps": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "gradient_clipping": "auto",
  "zero_allow_untested_optimizer": true,
  "zero_force_ds_cpu_optimizer": false,
  "zero_optimization": {
    "stage": 2,
    "overlap_comm": true,
    "allgather_bucket_size": 5368709120,
    "reduce_bucket_size": 5368709120,
    "reduce_scatter": true,
    "sub_group_size": 1e9,
    "contiguous_gradients": true,
    "allgather_partitions": true
  },
  "fp16": {
    "enabled": false,
    "initial_scale_power": 16
  },
  "bf16": {
    "enabled": true
  }
}