qwen3-8b-eagle3.yaml

model:
  target_model_path: Qwen/Qwen3-8B
  draft_model_config: configs/qwen3-8b-eagle3.json
  target_backend: sglang
  embedding_key: model.embed_tokens.weight
  vocab_mapping_path: ./cache/vocab_mapping/qwen3-8b-eagle3.pt
  torch_dtype: bfloat16

data:
  hidden_states_path: ./cache/hidden_states/qwen3-8b-sharegpt
  max_length: 4096
  chat_template: qwen
  cache_dir: ./cache

training:
  strategy: eagle3
  num_epochs: 10
  batch_size: 1
  learning_rate: 1.0e-4
  max_grad_norm: 0.5
  ttt_length: 7
  attention_backend: flex_attention
  save_interval: 1000
  log_interval: 50
  dist_timeout: 20
  seed: 0

run_id: qwen3-8b-eagle3-offline
output_dir: ./outputs/qwen3-8b-eagle3-offline

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