# @package _global_
#
# mini-swe / R2E-Gym on DeepSeek-V4-Flash-BF16, 8-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 and convert the FP8 model to BF16 (Task 4).
#
# Launch:
#   bash examples/start_ray_cluster.sh <master-ip>   # on every node first
#   bash examples/start.sh dsv4_flash_bf16/swe_h20_8node

defaults:
  - /default
  - _self_

checkpoint_path: ???
hf_model_path: ???
colocate: true
total_steps: 300

data_source:
  sandbox:
    type: k8s
    command_exec_timeout_seconds: 600
    sandbox_creation_timeout_seconds: 600
    working_dir: /rl-sandbox
    kubeconfig: k8s/kubeconfig.yaml
    pod_manifest_path: k8s/pod_manifest.yaml
  dataset:
    prompt_data_path: ???
    input_key: prompt
    label_key: label
    metadata_key: metadata
    data_pre_processor: r2e_gym
    max_prompt_len: 16384  # 16K
  agent:
    name: mini-swe
    step_limit: 100
    request_timeout: 1200
    max_consecutive_format_errors: 10
    # R2E-Gym images keep the testbed Python in a venv/conda env that is not on the
    # default PATH; prepend it to every agent command.
    shell_prefix: "source /root/.bashrc >/dev/null 2>&1 || true; export PATH=/testbed/.venv/bin:/opt/miniconda3/envs/testbed/bin:/root/.local/bin:$PATH; "
  reward:
    name: r2e_gym
  num_trajectories_per_prompt: 16
  num_prompts_per_step: 32
  max_response_len_per_trajectory: 147456  # 160K - 16K

rollout:
  retry_limit: 5
  num_gpus_per_node: 8
  sglang_args:
    mem_fraction_static: 0.75
    disable_cuda_graph: false
    tp_size: 8
    ep_size: 8
    enable_hierarchical_cache: true
    hicache_ratio: 1.0
    load_format: dummy
    moe_runner_backend: triton
  sglang_replicas:
    regular:
      num_nodes: 8
      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: 8
  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: 8
    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
    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