model:
  target_model_path: Qwen/Qwen3-8B
  draft_model_config: configs/qwen3-8b-domino.json
  target_backend: sglang
  embedding_key: model.embed_tokens.weight
  torch_dtype: bfloat16

data:
  train_data_path: ./cache/dataset/sharegpt_train.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: 6.0e-4
  warmup_ratio: 0.04
  max_grad_norm: 1.0
  attention_backend: flex_attention
  num_anchors: 256
  loss_decay_gamma: 7.0
  lambda_base_start: 1.0
  lambda_base_decay_ratio: 1.0
  save_interval: 2000
  log_interval: 50
  dist_timeout: 30
  seed: 42

tracking:
  report_to: wandb
  wandb_project: specforge-qwen3-8b-domino
  wandb_name: qwen3-8b-domino_sharegpt

run_id: qwen3-8b-domino-online
output_dir: ./outputs/qwen3-8b-domino-online

deployment:
  mode: disaggregated
  trainer:
    nnodes: 1
    nproc_per_node: 8
  disaggregated:
    control_dir: outputs/qwen3-8b-domino-online/control
    consumer_state_dir: outputs/qwen3-8b-domino-online/consumer-state
    backend: mooncake
    server_urls:
      - http://127.0.0.1:30000
    mooncake_metadata_server: http://127.0.0.1:35880/metadata
    mooncake_master_server_addr: 127.0.0.1:35551
    mooncake_protocol: tcp