model_name_or_path: Qwen/Qwen2-VL-7B-Instruct
image_max_pixels: 262144
video_max_pixels: 16384

do_train: true
stage: sft
finetuning_type: full # only support full for now
dataset: llava_1k_en
preprocessing_num_workers: 8
cutoff_len: 4096
template: qwen2_vl

output_dir: saves/mca/qwen2_vl_full
per_device_train_batch_size: 1
gradient_accumulation_steps: 2
num_train_epochs: 2
learning_rate: 2e-5
logging_steps: 1
save_steps: 100
lr_scheduler_type: cosine
bf16: true

# mcore speed up
tensor_model_parallel_size: 4
sequence_parallel: true
pipeline_model_parallel_size: 2
bias_activation_fusion: true
apply_rope_fusion: true
use_distributed_optimizer: true