VeOmni/configs/multimodal/qwen3_vl/qwen3_vl_moe.yaml at main · ByteDance-Seed/VeOmni · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
model:
  model_path: Qwen3-VL-30B-A3B-Instruct
  ops_implementation:
    moe_implementation: fused
    attn_implementation: flash_attention_2

data:
  train_path: configs/multimodal/data/tulu_sharegpt4v_llavavideo.yaml
  data_type: conversation
  chat_template: qwen3vl
  max_seq_len: 4096
  train_size: 80000000
  mm_configs:
    image_max_pixels: 602112 # 28 * 28 * 768
    video_max_pixels: 602112 # 28 * 28 * 768
    max_frames: 16
    fps: 2.0
    use_audio_in_video: false

train:
  accelerator:
    ep_size: 1
    fsdp_config:
      fsdp_mode: fsdp2
  gradient_checkpointing:
    enable_reentrant: false
  wandb:
    enable: false
    project: qwen3_vl_moe
    name: qwen3_vl_moe
  freeze_vit: false
  optimizer:
    lr: 1.0e-5
    lr_decay_style: cosine
  num_train_epochs: 2
  micro_batch_size: 1
  global_batch_size: 32
  max_steps: 500
  init_device: meta
  profile:
    enable: true
    start_step: 20
    end_step: 21
    record_shapes: true
  checkpoint:
    output_dir: qwen3_vl_moe_sft
    manager: dcp