| { | |
| "_NOTE": "UNUSED — This file is kept for reference only. The training template now uses LoRA on a single H200 GPU instead of DeepSpeed ZeRO-3 on dual A100s. No scripts reference this file.", | |
| "bf16": { | |
| "enabled": true | |
| }, | |
| "zero_optimization": { | |
| "stage": 3, | |
| "offload_optimizer": { | |
| "device": "cpu", | |
| "pin_memory": true, | |
| "fast_init": false | |
| }, | |
| "overlap_comm": true, | |
| "contiguous_gradients": true, | |
| "reduce_bucket_size": 5e7, | |
| "stage3_prefetch_bucket_size": 5e7, | |
| "stage3_param_persistence_threshold": 1e5, | |
| "stage3_gather_16bit_weights_on_model_save": true | |
| }, | |
| "optimizer": { | |
| "type": "AdamW", | |
| "params": { | |
| "lr": "auto", | |
| "betas": "auto", | |
| "eps": "auto", | |
| "weight_decay": "auto" | |
| } | |
| }, | |
| "gradient_accumulation_steps": "auto", | |
| "gradient_clipping": "auto", | |
| "train_batch_size": "auto", | |
| "train_micro_batch_size_per_gpu": "auto", | |
| "wall_clock_breakdown": false | |
| } | |