ROLL/examples/qwen2.5-0.5B-agentic/agentic_val_sokoban_lora.yaml at 4989ec480ce3db4b858b9f4af4ce38afc5a90c79 · alibaba/ROLL · 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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
defaults:
  - ../config/traj_envs@_here_
  - ../config/deepspeed_zero@_here_
  - ../config/deepspeed_zero2@_here_
  - ../config/deepspeed_zero3@_here_
  - ../config/deepspeed_zero3_cpuoffload@_here_

hydra:
  run:
    dir: .
  output_subdir: null

exp_name: "agentic_pipeline"
seed: 42
logging_dir: ./output/logs
output_dir: ./output
render_save_dir: ./output/render
system_envs:
  USE_MODELSCOPE: '1'

#track_with: wandb
#tracker_kwargs:
#  api_key:
#  project: roll-agentic
#  name: ${exp_name}_sokoban
#  notes: "agentic_pipeline"
#  tags:
#    - agentic
#    - roll
#    - baseline

track_with: stdout # Disable TensorBoard in smoke run to bypass SummaryWriter type constraints.
tracker_kwargs:
  log_dir: ./output/tensorboard/agentic_sokoban_lora_smoke # Use local path so smoke test does not depend on external mount.


checkpoint_config:
  type: file_system
  output_dir: /tmp/roll_output/agentic_sokoban_lora_smoke # Keep checkpoint path local for a portable smoke run.

num_gpus_per_node: 4 # Fit smoke test to a 4-GPU node.

max_steps: 3 # Minimal training smoke: one training step is enough to verify end-to-end path.
save_steps: 10000
logging_steps: 1
eval_steps: 0 # Disable eval loop for faster smoke validation.
resume_from_checkpoint: false
async_generation_ratio: 1 # Required by partial_gpu_mode validation in agentic_pipeline.

rollout_batch_size: 4 # Keep rollout tiny to reduce runtime/memory.
val_batch_size: 4
sequence_length: 2048 # Reduce memory pressure while preserving normal train path.

advantage_clip: 0.2
ppo_epochs: 1
adv_estimator: "grpo"
#pg_clip: 0.1
#dual_clip_loss: True
init_kl_coef: 0.0
whiten_advantages: true
entropy_loss_coef: 0
max_grad_norm: 1.0

pretrain: Qwen/Qwen2.5-0.5B-Instruct
reward_pretrain: Qwen/Qwen2.5-0.5B-Instruct

actor_train:
  model_args:
    attn_implementation: fa2
    disable_gradient_checkpointing: false
    dtype: bf16
    model_type: ~
    lora_target: all-linear
    lora_rank: 32
    lora_alpha: 32
  training_args:
    learning_rate: 2.0e-5
    weight_decay: 0
    per_device_train_batch_size: 1 # Minimal micro-batch for smoke stability.
    gradient_accumulation_steps: 2
    warmup_steps: 1
    lr_scheduler_type: cosine
  data_args:
    template: qwen2_5
  strategy_args:
#    strategy_name: deepspeed_train
#    strategy_config: ${deepspeed_zero3}
    strategy_name: megatron_train
    strategy_config:
      tensor_model_parallel_size: 1
      pipeline_model_parallel_size: 1
      expert_model_parallel_size: 1
      use_distributed_optimizer: true
      recompute_granularity: full
  device_mapping: list(range(0,2)) # Constrain actor_train to 4 GPUs for this smoke profile.
  infer_batch_size: 2

actor_infer:
  model_args:
    disable_gradient_checkpointing: true
    dtype: bf16
    lora_target: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
    lora_rank: 32
    lora_alpha: 32
  generating_args:
    max_new_tokens: 64 # Shorter generation keeps smoke test fast.
    top_p: 1
    top_k: 3
    num_beams: 1
    temperature: 0.0
    num_return_sequences: 1
  data_args:
    template: qwen2_5
  strategy_args:
    strategy_name: vllm
    strategy_config:
      gpu_memory_utilization: 0.8
      block_size: 16
      load_format: auto
  device_mapping: list(range(0,4)) # Constrain actor_infer to same 4-GPU pool.

reference:
  model_args:
    attn_implementation: fa2
    disable_gradient_checkpointing: true
    dtype: bf16
    model_type: ~
  data_args:
    template: qwen2_5
  strategy_args:
    strategy_name: hf_infer
    strategy_config: ~
  device_mapping: list(range(0,2)) # Keep reference mapping consistent with 4-GPU smoke topology.
  infer_batch_size: 2

reward_normalization:
  grouping: traj_group_id # 可以tags(env_type)/traj_group_id(group)/batch(rollout_batch)... group_by计算reward/adv
  method: mean_std # asym_clip / identity / mean_std

train_env_manager:
  format_penalty: -0.15 # sokoban env penalty_for_step=-0.1
  max_env_num_per_worker: 4 # Smaller env fanout for quick smoke startup.
  num_env_groups: 2
  # under the same group, the env config and env seed are ensured to be equal
  group_size: 2
  tags: [SimpleSokoban]
  num_groups_partition: [2] # Match reduced group count for smoke.

val_env_manager:
  max_env_num_per_worker: 4 # Keep validation manager light even though eval is disabled.
  num_env_groups: 4
  group_size: 1 # should be set to 1 because val temperature is set to 0 and same prompt leads to same output
  tags: [SimpleSokoban, LargerSokoban, SokobanDifferentGridVocab, FrozenLake]
  num_groups_partition: [1, 1, 1, 1] # Minimal partitioning for smoke.


# Here, you can override variables defined in the imported envs. max_tokens_per_step: 128 in custom_env.SimpleSokoban, here replaced by 64
max_tokens_per_step: 64

custom_envs:
  SimpleSokoban:
    ${custom_env.SimpleSokoban}
  LargerSokoban:
    ${custom_env.LargerSokoban}
  SokobanDifferentGridVocab:
    ${custom_env.SokobanDifferentGridVocab}
  FrozenLake:
    ${custom_env.FrozenLake}
  FrozenLakeThink:
    ${custom_env.FrozenLakeThink}