2025-11-03 21:36:57 +08:00

14 lines
322 B
JSON

{
"bf16": {
"enabled": "auto"
},
"train_micro_batch_size_per_gpu": "auto",
"train_batch_size": "auto",
"gradient_accumulation_steps": "auto",
"zero_optimization": {
"stage": 2,
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1e9
}
}