examples/sft/train_config.yaml

# 1.Data setting
data_txt_root: '/dataset/example/train_data_info.txt'  # data csv_filepath
resolution: 1024 
aspect_ratio_type: 'mar_1024'  # data bucketing strategy, mar_256、mar_512、mar_1024
null_text_ratio: 0.1
dataloader_num_workers: 8
train_batch_size: 4
repeats: 1

prompt_template_encode_prefix: '<|im_start|>system\nAs an image captioning expert, generate a descriptive text prompt based on an image content, suitable for input to a text-to-image model.<|im_end|>\n<|im_start|>user\n'
prompt_template_encode_suffix: '<|im_end|>\n<|im_start|>assistant\n'
prompt_template_encode_start_idx: 36
prompt_template_encode_end_idx: 5

# 2. Model setting
text_tokenizer_max_length: 512  # tokenizer max len 
pretrained_model_name_or_path: "/xxx/weights/Longcat-Image-Dev" # root directory of the model，with vae、transformer、scheduler eta;
diffusion_pretrain_weight:  null # if a specified diffusion weight path is provided, load the model parameters from the current directory.
use_dynamic_shifting: true #  scheduler dynamic shifting
resume_from_checkpoint: latest
  # - "latest"  # Loads most recent step checkpoint
  # - "/path/to/checkpoint"  # Resumes from specified directory

# 3. Training setting
use_ema: False
ema_rate: 0.999
mixed_precision: 'bf16'
max_train_steps: 100000
gradient_accumulation_steps: 1
gradient_checkpointing: true
gradient_clip: 1.0
learning_rate: 1.0e-5
adam_weight_decay: 1.0e-2
adam_epsilon: 1.0e-8
adam_beta1: 0.9
adam_beta2: 0.999
lr_num_cycles: 1
lr_power: 1.0
lr_scheduler: 'constant'
lr_warmup_steps: 1000

#4. Log setting
log_interval: 20
save_model_steps: 1000
work_dir: 'output/sft_model'
seed: 43