Daimon / training-template /ds_config_zero2.json
HumboldtJoker's picture
Add training template: ds_config_zero2.json
ea74ace verified
Raw
History Blame Contribute Delete
770 Bytes
{
"_NOTE": "ZeRO Stage 2 for full-parameter SFT on single H200 SXM. Stage 2 shards gradients (not params) and offloads optimizer states to CPU RAM. No optimizer block here — Adafactor is configured in the training script directly, because DeepSpeed's optimizer config conflicts with custom optimizers passed to the Trainer.",
"bf16": {
"enabled": true
},
"zero_optimization": {
"stage": 2,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "none"
},
"overlap_comm": true,
"contiguous_gradients": true,
"reduce_scatter": true
},
"gradient_clipping": 1.0,
"gradient_accumulation_steps": 8,
"train_micro_batch_size_per_gpu": 1,
"steps_per_print": 10
}