jd_train/ds_config_zero3_lora.json

23 lines
682 B
JSON

{
"train_micro_batch_size_per_gpu": 1,
"gradient_accumulation_steps": 4,
"bf16": { "enabled": true },
"fp16": { "enabled": false },
"zero_optimization": {
"stage": 3,
"overlap_comm": true,
"contiguous_gradients": true,
"allgather_partitions": true,
"reduce_scatter": true,
"round_robin_gradients": true,
"reduce_bucket_size": 150000000,
"stage3_prefetch_bucket_size": 100000000,
"stage3_param_persistence_threshold": 1000000,
"offload_optimizer": { "device": "none" },
"offload_param": { "device": "none" }
},
"stage3_gather_16bit_weights_on_model_save": false,
"gradient_clipping": 1.0,
"wall_clock_breakdown": false
}