config.yaml

# ================ Train Config ================ #
lyric_processor:
max_dur: 150
min_dur: 30
prompt_len: 10
pad_to_max: true


# ================ Audio tokenzier ================ #
audio_tokenizer_checkpoint: Flow1dVAE1rvq_./ckpt/model_1rvq/model_2_fixed_fp16.npz
audio_tokenizer_frame_rate: 25
audio_tokenizer_code_depth: 1
sample_rate: 48000

audio_tokenizer_checkpoint_sep: Flow1dVAESeparate_./ckpt/model_septoken/model_2_fp16.npz
audio_tokenizer_frame_rate_sep: 25
audio_tokenizer_code_depth_sep: 2
sample_rate_sep: 48000

# ================ VAE ================ #
vae_config: ./ckpt/vae/stable_audio_1920_vae.json
vae_model: ./ckpt/vae/autoencoder_music_1320k.npz

# ================== LM =========================== #
lm:
  lm_type: Llama # [Llama]
  dim: 1536
  intermediate_size: 8960
  num_heads: 12
  num_layers: 28
  num_layers_sub: 12
  code_depth: 3
  code_size: 16384
  max_position_embeddings: 8196
  max_position_embeddings_sub: 10000
  rope_theta: 100000.0
  rope_theta_sub: 500000.0
  dropout: 0.0
  use_flash_attn_2: true
  activation: gelu
  norm_first: true
  bias_ff: false
  bias_attn: false
  causal: true
  custom: false
  memory_efficient: true
  attention_as_float32: false
  layer_scale: null
  positional_embedding: sin
  xpos: false
  checkpointing: none
  weight_init: gaussian
  depthwise_init: current
  zero_bias_init: true
  norm: layer_norm
  cross_attention: false
  qk_layer_norm: false
  qk_layer_norm_cross: false
  attention_dropout: null
  kv_repeat: 1

codebooks_pattern:
  modeling: delay
  delay:
    delays: [ 0, 250, 250 ]
    flatten_first: 0
    empty_initial: 0

# ================ Conditioners ===================== #
classifier_free_guidance:
  # drop all conditions simultaneously
  training_dropout: 0.15
  inference_coef: 1.5

attribute_dropout:
  # drop each condition separately
  args:
    active_on_eval: false
  text:
    description: 0.0
    type_info: 0.5
  audio:
    prompt_audio: 0.0


use_text_training: True
fuser:
  sum: []
  prepend: [ description, prompt_audio, type_info ] # this order is the SAME with the input concatenation order

conditioners:
  prompt_audio:
    model: qt_embedding
    qt_embedding:
      code_size: 16384
      code_depth: 3
      max_len: ${eval:${prompt_len}*${audio_tokenizer_frame_rate}+2} # 25*10+2+1
  description:
    model: QwTokenizer
    QwTokenizer:
      token_path: third_party/Qwen2-7B
      max_len: 300
      add_token_list: ${load_yaml:conf/vocab.yaml}
  type_info:
    model: QwTextTokenizer
    QwTextTokenizer:
      token_path: third_party/Qwen2-7B
      max_len: 50

offload:
  audiolm:
    offload_module: self
    cpu_mem_gb: 0
    pre_copy_step: 1
    clean_cache_after_forward: false
    dtype: float16
    offload_layer_dict:
      transformer: 4
      transformer2: 4
    ignore_layer_list: []
    clean_cache_wrapper:
      module: self
      method_name: _sample_next_token
      diff_mem_gb_thre: 2
    debug: false

  wav_tokenizer_diffusion:
    offload_module: self.model.model
    pre_copy_step: 1
    clean_cache_after_forward: false
    cpu_mem_gb: -1
    dtype: null
    offload_layer_dict:
      cfm_wrapper: 5
      hubert: 4
    ignore_layer_list: []
    clean_cache_wrapper:
      module: self.model.model.cfm_wrapper.estimator
      method_name: forward
      diff_mem_gb_thre: 1
    debug: false