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

data:
  hidden_states_path: ./cache/hidden_states/qwen3-8b-dflash-sharegpt
  max_length: 3072
  chat_template: qwen
  cache_dir: ./cache

training:
  strategy: dflash
  num_epochs: 6
  max_steps: 10000
  batch_size: 4
  learning_rate: 6.0e-4
  warmup_ratio: 0.04
  max_grad_norm: 1.0
  attention_backend: flex_attention
  num_anchors: 512
  loss_decay_gamma: 7.0
  save_interval: 1000
  log_interval: 50
  dist_timeout: 30
  seed: 42

run_id: qwen3-8b-dflash-offline
output_dir: ./outputs/qwen3-8b-dflash-offline

deployment:
  mode: local_colocated
  trainer:
    nnodes: 1
    nproc_per_node: 1