# @package _global_
# Qwen3-4B + DAPO, single H800 node (8 GPUs).
# Launch: bash examples/start.sh qwen3_4b/dapo_h800_1node
# Fill in the two ??? below, or override them on the command line:
#   hf_model_path                         HF weight dir of Qwen3-4B
#   data_source.dataset.prompt_data_path  dapo-math-17k.parquet
defaults:
  - /default
  - _self_

checkpoint_path: ./checkpoints
hf_model_path: ???
colocate: true
total_steps: 100

data_source:
  dataset:
    prompt_data_path: ???
    input_key: prompt
    label_key: reward_model
  reward:
    name: dapo_math
    overlong_penalty_length: 4096
  num_prompts_per_step: 64
  num_trajectories_per_prompt: 8
  max_response_len_per_trajectory: 20480

rollout:
  num_gpus_per_node: 8
  sglang_args:
    mem_fraction_static: 0.7
    disable_cuda_graph: false

agentflow:
  dump_trajectory_path: "log/trajectories"

trainer:
  mini_batch_size: 128
  max_tokens_per_gpu: 20480
  use_dynamic_batch_size: true

megatron:
  model:
    tensor_model_parallel_size: 4
    pipeline_model_parallel_size: 2
    virtual_pipeline_model_parallel_size: null
    context_parallel_size: 1
    expert_model_parallel_size: 1
    expert_tensor_parallel_size: 1

algorithm:
  advantage_estimator: grpo
  advantage_norm_mode: group_zscore
  policy_loss: grpo
  loss_agg_mode: token-mean
  is_correction:
    enable: true
    action: mask
    level: token
    lower_bound: 0.5
    upper_bound: 2