hailin
/
jd_train
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

.

This commit is contained in:
hailin 2025-09-09 23:02:04 +08:00
parent 515565b3cd
commit 47c5cbe68c
2 changed files with 8 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
ds_config_zero3_lora.json
View File

@ -1,21 +1,22 @@
{
"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,
"reduce_bucket_size": 500000000,
"stage3_prefetch_bucket_size": 200000000,
"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
}

2
train_mm_zero3_lora.sh
View File

@ -16,5 +16,5 @@ FORCE_COLOR=1 deepspeed --hostfile hostfile \
--log_interval 10 \
--eval_steps 50 \
--gradient_checkpointing \
--deepspeed /home/test/jd_train/ds_config_zero3_lora_gpu.json \
--deepspeed /home/test/jd_train/ds_config_zero3_lora.json \
--report_to wandb --wandb_project ds-qwen3-lora