| # 4 * 36GB | |
| # Multimodal packing currently only supports qwen2_vl, qwen2_5_vl, qwen2_5_omni, internvl2_5/3 | |
| # Efficiency: With packing: 10 minutes; Without packing: >=1 hour | |
| # For local datasets, it is recommended to use streaming: `--streaming true` (save memory) | |
| NPROC_PER_NODE=4 \ | |
| MAX_PIXELS=1003520 \ | |
| CUDA_VISIBLE_DEVICES=0,1,2,3 \ | |
| swift sft \ | |
| --model Qwen/Qwen2.5-VL-7B-Instruct \ | |
| --train_type lora \ | |
| --dataset 'AI-ModelScope/LaTeX_OCR#20000' \ | |
| --torch_dtype bfloat16 \ | |
| --attn_impl flash_attn \ | |
| --packing true \ | |
| --num_train_epochs 3 \ | |
| --per_device_train_batch_size 1 \ | |
| --per_device_eval_batch_size 1 \ | |
| --learning_rate 1e-4 \ | |
| --lora_rank 8 \ | |
| --lora_alpha 32 \ | |
| --target_modules all-linear \ | |
| --gradient_accumulation_steps 1 \ | |
| --eval_steps 100 \ | |
| --save_steps 100 \ | |
| --save_total_limit 2 \ | |
| --logging_steps 5 \ | |
| --max_length 8192 \ | |
| --output_dir output \ | |
| --warmup_ratio 0.05 \ | |
| --dataloader_num_workers 4 \ | |
| --dataset_num_proc 8 \ | |
| --deepspeed zero2 | |