guanghulab/factory/training/configs/deepspeed-zero2-1p7b.json

51 lines
1.1 KiB
JSON
Raw Normal View History

{
"_comment": "DeepSpeed ZeRO-2 配置 · MP 蒸馏 / 微调Qwen2.5-1.5B · 5-01 校准 / 4×A100 80G",
"_arch_ref": "HLDP-ARCH-002 · factory/training/README.md",
"_owner": "铸渊 ICE-GL-ZY001 · 2026-05-01",
"train_micro_batch_size_per_gpu": 4,
"train_batch_size": "auto",
"gradient_accumulation_steps": 4,
"gradient_clipping": 1.0,
"steps_per_print": 20,
"bf16": {
"enabled": true
},
"zero_optimization": {
"stage": 2,
"allgather_partitions": true,
"allgather_bucket_size": 500000000,
"overlap_comm": true,
"reduce_scatter": true,
"reduce_bucket_size": 500000000,
"contiguous_gradients": true
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": 5e-5,
"betas": [0.9, 0.95],
"eps": 1e-8,
"weight_decay": 0.05
}
},
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"warmup_min_lr": 0,
"warmup_max_lr": 5e-5,
"warmup_num_steps": "auto",
"total_num_steps": "auto"
}
},
"activation_checkpointing": {
"partition_activations": true,
"contiguous_memory_optimization": true
}
}