GPU Imbalanced Loading

[WillDreamer]

Reminder

• I have read the above rules and searched the existing issues.

System Info

When I submit a training jobs as follows:
llamafactory-cli train \ --stage sft \ --do_train True \ --model_name_or_path Qwen/Qwen2-VL-7B-Instruct \ --preprocessing_num_workers 16 \ --finetuning_type lora \ --template qwen2_vl \ --rope_scaling linear \ --flash_attn auto \ --dataset_dir /data/LLaMA-Factory/data \ --dataset EMMA-mini \ --cutoff_len 4096 \ --learning_rate 5e-05 \ --num_train_epochs 3.0 \ --max_samples 100000 \ --per_device_train_batch_size 1 \ --gradient_accumulation_steps 2 \ --lr_scheduler_type cosine \ --max_grad_norm 1.0 \ --logging_steps 5 \ --save_steps 100 \ --warmup_steps 0 \ --packing False \ --report_to \ --output_dir saves/Qwen2-VL-7B-Instruct/lora/train_2025-03-11-07-34-45 \ --pure_bf16 True \ --plot_loss True \ --trust_remote_code True \ --ddp_timeout 180000000 \ --include_num_input_tokens_seen True \ --optim adamw_torch \ --quantization_bit 4 \ --quantization_method bitsandbytes \ --double_quantization True \ --lora_rank 8 \ --lora_alpha 16 \ --lora_dropout 0 \ --lora_target all \ --deepspeed cache/ds_z2_config.json

The GPU states are not balanced. How could I deal with this?

Reproduction

GPU imbalance

Others

No response

[Kuangdd01]

This issue seems similar to #5991.
In your case, batchsize_per_device is set to 1. GPU utilization will be different due to different sequence length on per gpu.