exp/final.yml

name: final

seed: 998244353
device_specific_seed: true
workder_specific_seed: true


data:
  data_path: train_data/chenyuan/final.yml
  image_size: 512
  apply_chat_template: true
  max_token_len: 4000
  img_scale_num: 16
  use_text_loss: false
  interleave: true
  use_input_image_for_vlm: true
  
  
model:
  # pretrained_model_path: ~
  pretrained_vae_model_name_or_path: /share_2/luoxin/modelscope/hub/models/FLUX.1-dev
  pretrained_text_encoder_model_name_or_path: /share/shitao/models/Qwen2.5-VL-3B-Instruct
  pretrained_diffusion_model_name_or_path: /share_2/luoxin/projects/Ominigenv2/experiments/0605_mix5/checkpoint-1000/pytorch_model_fsdp.bin
  use_vae_input: false
  use_all_condition: true
  use_image_tokens: false
  use_hw_tokens: false
  use_only_text_token_from_mllm: true
  use_only_prev_text_token_from_mllm: false

  arch_type: Lumina2Transformer_separate_ref_process
  arch_opt:
    patch_size: 2
    in_channels: 16
    hidden_size: 2520
    num_layers: 32
    num_refiner_layers: 2
    num_attention_heads: 21
    num_kv_heads: 7
    multiple_of: 256
    norm_eps: !!float 1e-05
    axes_dim_rope: [40, 40, 40]
    axes_lens: [12000, 12000, 12000]
    text_feat_dim: 2048
    timestep_scale: !!float 1000
    position_id_version: ominicontrol
    separate_ref_image_patch_embed: true
    separate_ref_image_refiner: true
    use_fused_rms_norm: true
    use_fused_swiglu: true
    proj_feat_depth: 8
    projector_dim: 2048
    proj_feat_dim: 1024

transport:
  snr_type: lognorm
  do_shift: true
  dynamic_time_shift: true

train:
  # Dataloader
  global_batch_size: 128
  batch_size: 2
  gradient_accumulation_steps: 1

  num_train_epochs: 5
  
  dataloader_num_workers: 6

  # Optimizer
  learning_rate: !!float 1e-6
  mllm_lr: !!float 3e-7
  mllm_projector_lr: !!float 1e-7
  mllm_vision_lr: !!float 1e-7
  mllm_image_token_lr: !!float 1e-4

  scale_lr: false
  lr_scheduler: timm_cosine
  t_initial: 200000
  lr_min: 0
  cycle_decay: 0.5
  warmup_t: 500
  warmup_lr_init: 0
  warmup_prefix: true
  t_in_epochs: false

  # resume_from_checkpoint: 

  use_8bit_adam: false
  adam_beta1: 0.9
  adam_beta2: 0.95
  adam_weight_decay: !!float 0.01
  adam_epsilon: !!float 1e-08
  max_grad_norm: 1

  gradient_checkpointing: true
  
  set_grads_to_none: true

  # Misc
  allow_tf32: false
  mixed_precision: 'bf16'

  ema_decay: 0.0

  repa_loss_weight: !!float 0.0
  repa_encoder: dinov2_vitl14_reg
  repa_batch_infer: true

  text_loss_weight: !!float 1.0

  train_diffusion: true
  train_mllm: true
  train_diffusion_context_refiner: true
  
val:
  validation_steps: 200
  train_visualization_steps: 200
  
logger:
  log_with: [wandb, tensorboard]
  # log_with: ~

  checkpointing_steps: 1000
  checkpoints_total_limit: ~

cache_dir: 
resume_from_checkpoint: latest
previous_stage_ckpt: