refactor mm training

Former-commit-id: 179c0558699e287cbf38a2d73bff47e86d589c5a

examples/README.md

@@ -33,6 +33,7 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 
 ```bash
 llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
 ```
 
 #### Reward Modeling

examples/README_zh.md

@@ -33,6 +33,7 @@ llamafactory-cli train examples/train_lora/llama3_lora_sft.yaml
 
 ```bash
 llamafactory-cli train examples/train_lora/llava1_5_lora_sft.yaml
+llamafactory-cli train examples/train_lora/qwen2vl_lora_sft.yaml
 ```
 
 #### 奖励模型训练

examples/train_full/qwen2vl_full_sft.yaml

@@ -1,40 +0,0 @@
-### model
-model_name_or_path: qwen2-vl-hf/qwen2-vl-7b-hf
-visual_inputs: true
-
-### method
-stage: sft
-do_train: true
-finetuning_type: full
-deepspeed: examples/deepspeed/ds_z3_config.json
-
-### dataset
-dataset: qwen2vl_demo
-template: qwen2vl
-cutoff_len: 1024
-max_samples: 1000
-overwrite_cache: true
-preprocessing_num_workers: 16
-
-### output
-output_dir: saves/qwen2-vl-7b/full/sft
-logging_steps: 10
-save_steps: 500
-plot_loss: true
-overwrite_output_dir: true
-
-### train
-per_device_train_batch_size: 1
-gradient_accumulation_steps: 1
-learning_rate: 1.0e-5
-num_train_epochs: 3.0
-lr_scheduler_type: cosine
-warmup_ratio: 0.1
-bf16: true
-ddp_timeout: 180000000
-
-### eval
-val_size: 0.1
-per_device_eval_batch_size: 1
-eval_strategy: steps
-eval_steps: 500

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