# @package _global_
#
# DAPO math on DeepSeek-V4-Flash-BF16, 6-node H20.
#
# Set before launching: checkpoint_path, hf_model_path, data_source.dataset.prompt_data_path.
# The first two must live on storage reachable by every node. See docs/en/run-guide.md for
# how to fetch the dataset (Task 1) and the FP8 -> BF16 model conversion (Task 4).
#
# Launch:
#   bash examples/start_ray_cluster.sh <master-ip>   # on every node first
#   bash examples/start.sh dsv4_flash_bf16/dapo_h20_6node

defaults:
  - /default
  - _self_

checkpoint_path: ???
hf_model_path: ???
colocate: true    # must be false in fully-async mode
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: 32
  num_trajectories_per_prompt: 8
  max_response_len_per_trajectory: 20480

rollout:
  retry_limit: 3
  num_gpus_per_node: 8
  sglang_args:
    mem_fraction_static: 0.75
    disable_cuda_graph: false
    tp_size: 8
    ep_size: 8
    load_format: dummy
    moe_runner_backend: triton
  sglang_replicas:
    regular:
      num_nodes: 6
      num_gpus_per_replica: 8
  sampler:
    num_oversample: 0
    timeout: 144000
  use_fault_tolerance: true

agentflow:
  dump_trajectory_path: "log/trajectories"
  router:
    proxy_timeout_seconds: 1800
  tokenizer:
    generation_prompt_kwargs:
      thinking: true

trainer:
  num_gpus_per_node: 8
  num_nodes: 6
  mini_batch_size: 128
  max_tokens_per_gpu: 20480
  use_dynamic_batch_size: true
  use_rollout_routing_replay: true
  use_fp32_lm_head: true
  save_freq: 5
  nccl_timeout_minutes: 60
  gloo_timeout_minutes: 60

megatron:
  model:
    tensor_model_parallel_size: 1
    pipeline_model_parallel_size: 6
    context_parallel_size: 8
    sequence_packing_scheduler: "default_dynamic_cp"
    cp_partition_mode: "contiguous"
    expert_model_parallel_size: 8
    expert_tensor_parallel_size: 1
    dsa_indexer_loss_coeff: 0.0
    dsa_indexer_use_sparse_loss: true
    apply_dsa_kernel_fusion: true
    moe_shared_expert_overlap: false
    recompute_granularity: full # selective|full
    recompute_method: uniform
    recompute_num_layers: 1
  optimizer:
    lr: 1.0e-6
    optimizer_cpu_offload: true
    optimizer_offload_fraction: 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