model:
  target_model_path: Qwen/Qwen3-8B
  draft_model_config: configs/qwen3-8b-eagle3.json
  vocab_mapping_path: cache/vocab_mapping/qwen3-8b.pt
data:
  hidden_states_path: cache/hidden_states/qwen3-8b
  max_length: 4096
training:
  strategy: eagle3
  num_epochs: 10
  batch_size: 1
  learning_rate: 0.0001
  ttt_length: 7
  save_interval: 5000
  dist_timeout: 20
  seed: 0
run_id: qwen3-8b-eagle3-offline-disaggregated
output_dir: outputs/qwen3-8b-eagle3-offline-disaggregated

deployment:
  mode: disaggregated
  trainer:
    nnodes: 1
    nproc_per_node: 1
  disaggregated:
    control_dir: outputs/qwen3-8b-eagle3-offline-disaggregated/control
    backend: shared_dir
    store_root: outputs/qwen3-8b-eagle3-offline-disaggregated/features