Model-Optimizer/modelopt_recipes/general/speculative_decoding/dflash.yaml at main · kvcache-ai/Model-Optimizer · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
# Base config for DFlash training. Override fields via OmegaConf dotlist on the CLI.

# maps to ModelArguments (main.py)
model:
  model_name_or_path:
  trust_remote_code: false
  use_fake_base_for_offline: false

# maps to DataArguments (main.py)
data:
  data_path:
  offline_data_path:
  # Jinja chat template with {% generation %} tags for answer_only_loss.
  # Required when answer_only_loss=true. Set in per-model launcher YAML.
  # Templates are in modelopt_recipes/general/speculative_decoding/chat_templates/
  chat_template:

# maps to TrainingArguments (main.py)
training:
  # --- commonly modified ---
  mode: dflash
  output_dir:
  num_train_epochs: 10
  per_device_train_batch_size: 1
  learning_rate: 6.0e-4
  warmup_steps: 100
  training_seq_len: 4096
  logging_steps: 100
  save_steps: 5000
  cp_size: 1
  dp_shard_size: 1
  disable_tqdm: true
  estimate_ar: false
  ar_validate_steps: 0
  answer_only_loss: true

  # --- rarely modified ---
  do_eval: false
  lr_scheduler_type: linear
  save_strategy: steps
  weight_decay: 0.0
  dataloader_drop_last: true
  bf16: true
  tf32: true
  remove_unused_columns: false
  ddp_find_unused_parameters: true
  ddp_timeout: 1800
  report_to: tensorboard

# maps to DFlashConfig (modelopt/torch/speculative/config.py).
dflash:
  dflash_block_size: 8
  dflash_num_anchors: 512
  dflash_use_torch_compile: false
  dflash_self_logit_distillation: true
  dflash_loss_decay_factor: 4.0
  dflash_architecture_config:
    num_hidden_layers: 5
    # mask_token_id: auto-detected from model vocab (override for specific models)
    # sliding_window and layer_types are inherited from base model config automatically