project: AgentScope-ReAct
name: Learn_to_Ask-Qwen2.5-7B-fixed
# directory to save checkpoints, default to ./checkpoints if TRINITY_CHECKPOINT_ROOT_DIR not set
checkpoint_root_dir: ${oc.env:TRINITY_CHECKPOINT_ROOT_DIR,./checkpoints}
algorithm:
  algorithm_type: grpo  # a GRPO-based algorithm for multi-step reasoning
model:
  # path to the pre-trained model, default to Qwen/Qwen2.5-7B-Instruct if TRINITY_MODEL_PATH not set
  model_path: ${oc.env:TRINITY_MODEL_PATH,Qwen/Qwen2.5-7B-Instruct}
  tinker:  # tinker config, you can set tinker parameters here
    enable: false  # if true, tinker will be enabled
cluster:
  node_num: 1  # cluster with 1 node
  gpu_per_node: 8  # each node has 8 GPUs
buffer:
  total_epochs: 4  # run taskset for 4 epoch
explorer:
  runner_per_model: 32  # each model has 32 runners for parallel rollout
  max_timeout: 600  # max timeout for each rollout is 600 seconds
synchronizer:
  sync_style: fixed
  sync_method: 'nccl'
  sync_interval: 10
  sync_timeout: 7200  # wait for 120 minutes
trainer:
  save_interval: 90  # save checkpoint every 90 steps
  use_dynamic_bsz: true
  ulysses_sequence_parallel_size: 1  # use sequence parallelism to reduce memory usage
monitor:
  monitor_type: wandb  # here we use wandb; you can also use tensorboard, mlflow or swanlab