OpenVLThinker/EasyR1/examples/config_g2rpo.yaml at main · uclanlp/OpenVLThinker · 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
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
data:
  train_files: hiyouga/math12k@train
  val_files: ""
  test_files: []
  prompt_key: problem
  answer_key: answer
  image_key: images
  video_key: videos
  image_dir: null
  val_image_dir: null
  video_fps: 2.0
  max_prompt_length: 16384
  max_response_length: 4096
  rollout_batch_size: 512  # equivalent to verl's data.train_batch_size
  mini_rollout_batch_size: null  # equivalent to verl's data.gen_batch_size
  val_batch_size: 1024
  format_prompt: examples/format_prompt/math_qwen3.jinja
  override_chat_template: null
  shuffle: true
  seed: 1
  min_pixels: 3136
  max_pixels: 1048576
  filter_overlong_prompts: false

algorithm:
  adv_estimator: gs_grpo
  disable_kl: true
  use_kl_loss: true
  use_entropy_loss: false
  kl_penalty: low_var_kl
  kl_coef: 1.0e-2
  online_filtering: true  # dapo filter groups
  filter_key: accuracy
  filter_low: 0.01
  filter_high: 0.99

worker:
  actor:
    global_batch_size: 128  # equivalent to verl's actor.ppo_mini_batch_size
    micro_batch_size_per_device_for_update: 1  # equivalent to verl's actor.ppo_micro_batch_size_per_gpu
    micro_batch_size_per_device_for_experience: 2  # equivalent to verl's rollout.log_prob_micro_batch_size_per_gpu
    max_grad_norm: 1.0
    padding_free: true
    dynamic_batching: true
    ulysses_size: 1
    model:
      model_path: Qwen/Qwen2.5-7B-Instruct
      enable_gradient_checkpointing: true
      trust_remote_code: false
      freeze_vision_tower: false
    optim:
      lr: 5.0e-6
      weight_decay: 1.0e-2
      strategy: adamw  # {adamw, adamw_bf16}
      lr_warmup_ratio: 0.0
    fsdp:
      enable_full_shard: true
      enable_cpu_offload: false
      enable_rank0_init: true
    offload:
      offload_params: false  # true: more CPU memory; false: more GPU memory
      offload_optimizer: false  # true: more CPU memory; false: more GPU memory

  rollout:
    n: 8
    temperature: 1.0
    top_p: 1.0
    limit_images: 0
    gpu_memory_utilization: 0.5
    enforce_eager: false
    enable_chunked_prefill: false
    tensor_parallel_size: 2
    disable_tqdm: true
    max_num_batched_tokens: 20480
    val_override_config:
      temperature: 0.7
      top_p: 0.95
      n: 1

  ref:
    fsdp:
      enable_full_shard: true
      enable_cpu_offload: false  # true: more CPU memory; false: more GPU memory
      enable_rank0_init: true
    offload:
      offload_params: false

  reward:
    reward_type: batch
    reward_function: verl/reward_function/multitask_reward.py:compute_score

trainer:
  total_epochs: 1
  max_steps: null
  project_name: easy_r1
  experiment_name: qwen2_5_7b_math_grpo
  logger: ["file", "wandb"]
  nnodes: 1
  n_gpus_per_node: 8
  max_try_make_batch: 20  # -1 means no limit
  val_freq: 1  # -1 to disable
  val_before_train: false
  val_only: false
  val_generations_to_log: 3
  save_freq: 200  # -1 to disable
  save_limit: 3  # -1 to disable
  save_model_only: false
  save_checkpoint_path: checkpoints/qwen3-8b-rl
  load_checkpoint_path: null
  find_last_checkpoint: true