# @package _global_
#
# OpenCode / R2E-Gym on Qwen3-Coder-30B-A3B-Instruct, 4-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.
#   hf download Qwen/Qwen3-Coder-30B-A3B-Instruct --local-dir <hf_model_path>
#   hf download loongsage/R2E-Gym --include "opencode/data/*.parquet" --repo-type=dataset --local-dir <prompt_data_path>
#
# Launch:
#   bash examples/start_ray_cluster.sh <head-node-ip>   # on every node first
#   bash examples/start.sh qwen3_coder_30b_a3b/opencode_h20_4node

defaults:
  - /default
  - _self_

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

data_source:
  sandbox:
    type: k8s
    command_exec_timeout_seconds: 3600
    sandbox_creation_timeout_seconds: 600
    # R2E-Gym repositories live under /testbed.
    working_dir: /testbed
    kubeconfig: k8s/kubeconfig.yaml
    pod_manifest_path: k8s/pod_manifest.yaml
  dataset:
    prompt_data_path: ???
    data_pre_processor: r2e_gym
    input_key: prompt
    label_key: label
    metadata_key: metadata
  agent:
    name: opencode
    # Qwen3-Coder-30B-A3B-Instruct has a native 262144 (256K) context. 64K is
    # restored because 32K pushed OpenCode to auto-compact early and fork a
    # storm of subagents. Activation memory is bounded by CP=2 instead: with
    # use_dynamic_batch_size=false and micro_batch_size=1 one full trajectory
    # is one microbatch, and CP splits that single sequence across 2 GPUs.
    context_length: 65536
    max_output_tokens: 32768
  reward:
    name: r2e_gym
  num_trajectories_per_prompt: 8
  num_prompts_per_step: 32
  max_response_len_per_trajectory: 63488

rollout:
  retry_limit: 5
  num_gpus_per_node: 8
  partial: true
  mask_offpolicy_in_partial_rollout: true
  sampler:
    num_oversample: 4
  sglang_replicas:
    regular:
      num_nodes: 4
      num_gpus_per_replica: 8
  sglang_args:
    # Qwen3-Coder-30B-A3B is MoE: tensor + expert parallelism.
    tp_size: 8
    ep_size: 8
    mem_fraction_static: 0.8
    disable_cuda_graph: false
    # skip_server_warmup: false
    # enable_fp32_lm_head: false
    # max_prefill_tokens: 16384
  use_fault_tolerance: true

agentflow:
  dump_trajectory_path: "log/trajectories"
  router:
    proxy_timeout_seconds: 1800
  tokenizer:
    generation_prompt_kwargs: {}

trainer:
  num_gpus_per_node: 8
  num_nodes: 4
  mini_batch_size: 64
  micro_batch_size: 2
  max_tokens_per_gpu: 8192
  use_dynamic_batch_size: true
  save_freq: 5
  # Routing replay (R3) is MoE-only; Qwen3-Coder-30B-A3B is MoE.
  use_rollout_routing_replay: true
  use_fp32_lm_head: true

megatron:
  model:
    # Qwen3-Coder-30B-A3B is MoE: tensor + expert parallelism (mirrors the
    # proven opencode-swe-rebench setup for this same model).
    tensor_model_parallel_size: 1
    pipeline_model_parallel_size: 2
    virtual_pipeline_model_parallel_size: null
    # CP=2 splits each sequence along the token dimension across 2 GPUs, so a
    # single long trajectory's activation is halved per GPU (the only way to
    # shard ONE sequence — dynamic batching can't). Requires 16 GPUs across 2
    # nodes: TP1 x PP1 x EP8 x CP2 x DP1.
    context_parallel_size: 2
    expert_model_parallel_size: 8
    expert_tensor_parallel_size: 1
    dsa_indexer_loss_coeff: 0.0
    apply_dsa_kernel_fusion: true
    recompute_granularity: full # selective|full
    recompute_method: uniform
    recompute_num_layers: 1
  optimizer:
    lr: 5.0e-7
    optimizer_cpu_offload: true
    optimizer_offload_fraction: 1

algorithm:
  advantage_estimator: grpo
  advantage_norm_mode: group_zscore
  policy_loss: grpo
  loss_agg_mode: seq-mean-token-mean
  #clip_ratio_low: 3e-4
  #clip_ratio_high: 4e-4
  is_correction:
    enable: false
    action: mask
    level: token
    lower_bound: 0.5
    upper_bound: 2