fish-speech/fish_speech/configs/text2semantic_pretrain.yaml

defaults:
  - base
  - _self_

project: text2semantic_400m_pretrain
max_length: 1024

# Lightning Trainer
trainer:
  accumulate_grad_batches: 2
  gradient_clip_val: 1.0
  gradient_clip_algorithm: 'norm'
  max_steps: 1_000_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}

val_dataset:
  _target_: fish_speech.datasets.text.AutoAugTextDataset
  tokenizer: ${tokenizer}
  max_length: ${max_length}

data:
  _target_: fish_speech.datasets.text.TextDataModule
  train_dataset: ${train_dataset}
  val_dataset: ${val_dataset}
  num_workers: 4
  batch_size: 16
  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: 4096
      vocab_size: 36408
      n_layer: 24
      n_head: 16
      dim: 1024
      rope_base: 10000
      norm_eps: 1e-5
      num_codebooks: 4  # single codebook
      codebook_size: 168 # codebook size 160 + 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