| { | |
| "train_micro_batch_size_per_gpu": "auto", | |
| "gradient_accumulation_steps": "auto", | |
| "bf16": { | |
| "enabled": true | |
| }, | |
| "zero_optimization": { | |
| "stage": 2, | |
| "offload_optimizer": { | |
| "device": "cpu", | |
| "pin_memory": false | |
| }, | |
| "offload_param": { | |
| "device": "none" | |
| }, | |
| "overlap_comm": true, | |
| "contiguous_gradients": true, | |
| "reduce_scatter": true, | |
| "stage3_prefetch_bucket_size": 536870912, | |
| "reduce_bucket_size": 536870912 | |
| }, | |
| "wall_clock_breakdown": false, | |
| "flops_profiler": { | |
| "enabled": false | |
| } | |
| } |