refactor mm training

Former-commit-id: 179c0558699e287cbf38a2d73bff47e86d589c5a

examples/train_lora/qwen2vl_lora_sft.yaml

@@ -1,5 +1,5 @@
 ### model
-model_name_or_path: qwen2-vl-hf/qwen2-vl-7b-hf
+model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
 visual_inputs: true
 
 ### method
@@ -9,23 +9,23 @@ finetuning_type: lora
 lora_target: all
 
 ### dataset
-dataset: qwen2vl_demo
-template: qwen2vl
+dataset: mllm_demo
+template: qwen2_vl
 cutoff_len: 1024
 max_samples: 1000
 overwrite_cache: true
 preprocessing_num_workers: 16
 
 ### output
-output_dir: saves/qwen2-vl-7b/lora/sft
+output_dir: saves/qwen2_vl-7b/lora/sft
 logging_steps: 10
 save_steps: 500
 plot_loss: true
 overwrite_output_dir: true
 
 ### train
-per_device_train_batch_size: 2
-gradient_accumulation_steps: 1
+per_device_train_batch_size: 1
+gradient_accumulation_steps: 8
 learning_rate: 1.0e-4
 num_train_epochs: 3.0
 lr_scheduler_type: cosine