model:
  target_model_path: "Qwen/Qwen3-8B"
  draft_model_config: "configs/qwen3-8b-domino.json"
  target_backend: "sglang"
  trust_remote_code: true
  embedding_key: "model.embed_tokens.weight"
  torch_dtype: "bfloat16"
  mask_token_id: 151669
data:
  train_data_path: "./cache/dataset/perfectblend_qwen3-8b_regen.jsonl"
  max_length: 3072
  chat_template: "qwen"
  build_dataset_num_proc: 32
  cache_dir: "./cache"
training:
  strategy: "domino"
  num_epochs: 6
  batch_size: 2
  learning_rate: 0.0006
  warmup_ratio: 0.04
  max_grad_norm: 1
  attention_backend: "flex_attention"
  num_anchors: 256
  loss_decay_gamma: 7
  lambda_base_start: 1
  lambda_base_decay_ratio: 1
  save_interval: 1000
  log_interval: 10
  dist_timeout: 30
  seed: 42
tracking:
  report_to: "none"
  wandb_project: "qwen3-8b-domino-disagg"
  wandb_name: "qwen3-8b-domino-2srv-dp2"
run_id: "qwen3-8b-domino-multiserver-disaggregated"
output_dir: "./outputs/qwen3-8b-domino-multiserver-disaggregated"

deployment:
  mode: disaggregated
  trainer:
    nnodes: 1
    nproc_per_node: 2
  disaggregated:
    control_dir: outputs/qwen3-8b-domino-multiserver-disaggregated/control
    backend: mooncake
    managed_local:
      trainer_cuda_visible_devices:
        - "2"
        - "3"
      mooncake:
        protocol: tcp
        global_segment_size_bytes: 34359738368
        local_buffer_size_bytes: 1073741824
      capture_servers:
        - port: 30000
          cuda_visible_devices:
            - "0"
          tp_size: 1
          mem_fraction_static: 0.85
        - port: 30001
          cuda_visible_devices:
            - "1"
          tp_size: 1
          mem_fraction_static: 0.85