| { | |
| "train_batch_size": 16, | |
| "gradient_accumulation_steps": 1, | |
| "optimizer": { | |
| "type": "AdamW", | |
| "params": { | |
| "lr": 0.00004 | |
| } | |
| }, | |
| "fp16": { | |
| "enabled": true, | |
| "auto_cast": false | |
| }, | |
| "zero_optimization": { | |
| "stage": 1, | |
| "allgather_partitions": true, | |
| "allgather_bucket_size": 1e9, | |
| "overlap_comm": true, | |
| "reduce_scatter": false, | |
| "reduce_bucket_size": 1e9, | |
| "contiguous_gradients": true | |
| }, | |
| "log": { | |
| "steps_per_print": 50 | |
| }, | |
| "wall_clock_breakdown": false | |
| } | |