guanghulab/factory/training/configs/deepspeed-zero3-8b.json

68 lines
1.6 KiB
JSON
Raw Permalink Normal View History

{
"_comment": "DeepSpeed ZeRO-3 配置 · M0 全参训练Qwen2.5-7B · 5-01 校准 / 8×A100 80G or 8×H100 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": 1,
"train_batch_size": "auto",
"gradient_accumulation_steps": 16,
"gradient_clipping": 1.0,
"steps_per_print": 10,
"bf16": {
"enabled": true
},
"zero_optimization": {
"stage": 3,
"offload_optimizer": {
"device": "cpu",
"pin_memory": true
},
"offload_param": {
"device": "cpu",
"pin_memory": true
},
"overlap_comm": true,
"contiguous_gradients": true,
"sub_group_size": 1000000000,
"reduce_bucket_size": "auto",
"stage3_prefetch_bucket_size": "auto",
"stage3_param_persistence_threshold": "auto",
"stage3_max_live_parameters": 1000000000,
"stage3_max_reuse_distance": 1000000000,
"stage3_gather_16bit_weights_on_model_save": true
},
"optimizer": {
"type": "AdamW",
"params": {
"lr": 2e-5,
"betas": [0.9, 0.95],
"eps": 1e-8,
"weight_decay": 0.1
}
},
"scheduler": {
"type": "WarmupDecayLR",
"params": {
"warmup_min_lr": 0,
"warmup_max_lr": 2e-5,
"warmup_num_steps": "auto",
"total_num_steps": "auto"
}
},
"activation_checkpointing": {
"partition_activations": true,
"cpu_checkpointing": false,
"contiguous_memory_optimization": true,
"number_checkpoints": null,
"synchronize_checkpoint_boundary": false,
"profile": false
},
"wall_clock_breakdown": false
}