### model
model_name_or_path: google/gemma-3n-E2B-it
image_max_pixels: 1048576
trust_remote_code: true

### method
stage: sft
do_train: true
finetuning_type: lora
freeze_vision_tower: false
freeze_multi_modal_projector: false
freeze_language_model: false
lora_rank: 32
lora_target: "q_proj, k_proj, v_proj, o_proj"
deepspeed: examples/deepspeed/ds_z2_config.json

### dataset
dataset:  and_ctrl_train # video: mllm_video_demo
template: gemma3n
cutoff_len: 1024               # 2048 -> 1024 (토큰 짧으면 속도/메모리 이득)
# max_samples: 1000            # 전체 학습이면 제거/주석 처리
overwrite_cache: true
preprocessing_num_workers: 8
dataloader_num_workers: 8     # 8~16 권장, 중간값으로 상향

### output
output_dir: saves/gemma-e2b/lora/sft/train
logging_steps: 1
save_steps: 300
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: wandb  # choices: [none, wandb, tensorboard, swanlab, mlflow]
run_name: and_ctrl-skt-gemma-e2b-lora-sft-train

### train
per_device_train_batch_size: 4
gradient_accumulation_steps: 48
learning_rate: 2e-5
num_train_epochs: 5.0
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
resume_from_checkpoint: null
seed: 42

# 추가 적용(요청 사항)
gradient_checkpointing: true
optim: adamw_torch_fused
weight_decay: 0.01

### eval
val_size: 0.05
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500