qwen3_vl_dense.yaml

model:
  model_path: Qwen/Qwen3-VL-8B-Instruct
  attn_implementation: flash_attention_2

data:
  train_path: sharegpt4v_pretrain
  data_type: conversation
  chat_template: qwen3vl
  max_seq_len: 4096
  train_size: 80000000
  mm_config:
    fps: 2.0
    use_audio_in_video: false

train:
  output_dir: qwen3_vl_dense_sft
  data_parallel_mode: fsdp2
  enable_reentrant: false
  use_wandb: false
  wandb_project: qwen3_vl_dense
  wandb_name: qwen3_vl_dense
  rmpad: false
  rmpad_with_pos_ids: true
  ulysses_parallel_size: 1
  freeze_vit: false
  lr: 1.0e-5
  lr_decay_style: cosine
  num_train_epochs: 2
  micro_batch_size: 1
  global_batch_size: 16
  max_steps: 500
  init_device: meta
  enable_profiling: true
  profile_start_step: 20
  profile_end_step: 21
  profile_record_shapes: true
  ckpt_manager: dcp
  save_hf_weights: false