fish-speech/fish_speech/configs/text2semantic_pretrain_small.yaml

defaults:
  - base
  - _self_

project: text2semantic_pretrain_small_dual_ar
max_length: 2048

# Lightning Trainer
trainer:
  accumulate_grad_batches: 1
  gradient_clip_val: 1.0
  gradient_clip_algorithm: 'norm'
  max_steps: 100_000
  precision: bf16-true
  limit_val_batches: 10

# Dataset Configuration
tokenizer:
  _target_: transformers.AutoTokenizer.from_pretrained
  pretrained_model_name_or_path: fishaudio/speech-lm-v1

# Dataset Configuration
train_dataset:
  _target_: fish_speech.datasets.text.AutoAugTextDataset
  tokenizer: ${tokenizer}
  max_length: ${max_length}
  num_codebooks: ${model.model.config.num_codebooks}
  use_speaker: false
  phones_prob: 0.5
  interactive_prob: 0.5

val_dataset:
  _target_: fish_speech.datasets.text.AutoAugTextDataset
  tokenizer: ${tokenizer}
  max_length: ${max_length}
  num_codebooks: ${model.model.config.num_codebooks}
  use_speaker: false
  phones_prob: 0.5
  interactive_prob: 0.5

data:
  _target_: fish_speech.datasets.text.TextDataModule
  train_dataset: ${train_dataset}
  val_dataset: ${val_dataset}
  num_workers: 4
  batch_size: 32
  tokenizer: ${tokenizer}
  max_length: ${max_length}

# Model Configuration
model:
  _target_: fish_speech.models.text2semantic.TextToSemantic

  model:
    # ~ 130M parameters, for debug purpose
    _target_: fish_speech.models.text2semantic.llama.Transformer
    config:
      _target_: fish_speech.models.text2semantic.llama.ModelArgs
      max_seq_len: ${max_length}
      vocab_size: 36408
      n_slow_layer: 12
      n_fast_layer: 4
      n_head: 12
      dim: 768
      rope_base: 10000
      norm_eps: 1e-5
      num_codebooks: 8  # input/output codebook size
      codebook_size: 264 # codebook size 256 + 2 special tokens

  optimizer:
    _target_: torch.optim.AdamW
    _partial_: true
    lr: 3e-4
    weight_decay: 0.1
    betas: [0.9, 0.95]
    eps: 1e-5

  lr_scheduler:
    _target_: torch.optim.lr_scheduler.LambdaLR
    _partial_: true
    lr_lambda:
      _target_: fish_speech.scheduler.get_cosine_schedule_with_warmup_lr_lambda
      _partial_: true
      num_warmup_steps: 2000
      num_training_steps: ${trainer.max_steps}
      final_lr_ratio: 0.1