vall-e/data/config.yaml

sample_rate: 24_000 # 44_000 / 44_100 for dac
audio_backend: "vocos" # or dac

# model definitions to train
models:
- name: "ar+nar" # vanity name
  size: "full" # model dimensionality
  resp_levels: 8 # RVQ levels this model targets
  tasks: 9 # tasks this model can attend to, only tts is guaranteed results at the moment
  langs: 4 # languages this model supports
  tones: 1 # tones this model supports, currently unused
  arch_type: llama # underlying LLM arch to use, currently focusing on llama
  training: True # signals this model is to be trained
  version: 5 # helps keep backwards compatibility for when I add new things to the model
  attention: auto # attention mechanism to use, "auto" for safety
  dropout: 0.1 # percentage of the model to disable during training

  # factors for split loss values, remove to have a unified loss calculation
  loss_factors:
    text: 0.1 # text phoneme portion of the sequence
    prom: 0.5 # input prompt portion of the sequence
    resp: 1.0 # output audio portin of the sequence

  capabilities: ["ar", "nar"] # macro-tasks this model can perform

  # experimental settings
  experimental:
    rvq_levels_p: "auto" # "equal" | "auto" | list[int], sets probabilities of which RVQ level to select during training, auto will have the next RVQ level half as likely as the previous one
    audio_embedding_sums: True # whether the input embeddings include all prior RVQ levels (sums) or only the current one (further experimentation is needed to see if this matters)
    unified_position_ids: False # specifies whether or not position IDs should be continuous across the whole sequence (if True, naive behavior), or restart them at the next segment of the sequence (if False)
    split_classifiers: True # use per-RVQ-level projection/output/classifiers for the model (further experimentation is needed to see if this matters)

# list of LoRA(s) to use
#loras:
#- name : "lora-shodan" # LoRA name to load from
#  rank: 128 # parameter size per Linear
#  alpha: 128 # "influence" value
#  training: True # 
#  rvq_levels: [] # RVQ levels to activate the LoRA on, leave empty for all

# hyperparameter settings (could be relegated to trainer settings)
hyperparameters:
  # deepspeed autotune
  autotune: False
  autotune_params:
    start_profile_step: 1
    end_profile_step: 50
    num_tuning_micro_batch_sizes: 8

  batch_size: 16 # samples per batch, governs maximum batch size if using batch sampling
  gradient_accumulation_steps: 4 # gradient accumulation: batches per update
  gradient_clipping: 1.0 # smooths out the gradient when updating
  warmup_steps: 100 # steps to warm up the optimizer but not update the model

  # optimizer settings
  optimizer: Prodigy
  learning_rate: 1.0 # prodigyopt can keep its LR to 1
  torch_optimizer: True # signals to deepspeed to not instantiate one
  
  # deepspeed scheduler, local does have it implemented because I don't use one
  scheduler: "" # ScheduleFree
  torch_scheduler: True # signals to deepspeed to not instantiate one

# evaluation settings (could be pushed under trainer)
evaluation:
  batch_size: 8 # batch size for evaluation / validation pass
  frequency: 5000 # how often to perform eval during training
  size: 8 # total samples to get for eval

  # arguments to pass for the AR/NAR (matches arguments passed through vall_e.inference)
  kwargs:
    max_steps: 500 # how many AR steps to perform
    ar_temp: 0.95 # temperature for AR sampling
    nar_temp: 0.25 # temperature for NAR sampling

trainer:
  iterations: 1_000_000 # how many total iterations to train before terminating, should just have this as 0 by default to not auto-terminiate
  
  save_tag: step # tag name to save checkpoints under
  save_on_oom: True # save if an OOM if caught
  save_on_quit: True # save when `quit` is entered in the trainer
  save_frequency: 250 # how often to save
  export_on_save: True # export the weights every time the trainer saves

  keep_last_checkpoints: 4 # how many previous checkpoints to keep

  gradient_checkpointing: True # gradient checkpointing to save VRAM at the cost of some performance throughput

  strict_loading: False # strict state dict loading (set to False if you're going to change some model settings)
  resize_modules: True # automatically resize core modules from the state dict to match

  #check_for_oom: False # wrap forward/backwards in a try/catch block and gracefully handles OOM conditions
  #load_state_dict: True # load the state dict from fp32.pth instead of a checkpoint, should automagically be done
  #load_tag: "9500" # specific tag to load from (instead of having to edit latest)
  #load_states: False # flag to load optimizer / scheduler states or not
  #restart_step_count: True # clear the trainer stats
  # gc_mode: None # "global_step" # flag to call GC at specific points, seems overkill now

  weight_dtype: float16 # float32 | float16 | bfloat16, dtype for the model to load under
  amp: True # mixed precision during training

  backend: deepspeed # deepspeed | local, training backend to use

  # deepspeed specific settings
  deepspeed:
    inferencing: True # use deepspeed inference wrapper for inferencing, should be relegated under inference
    amp: False # use deepspeed's AMP instead (requires nvidia/apex installed)
    zero_optimization_level: 0 # ZeRO optimization level to use
    use_compression_training: False # compression training (seems useless almost always)

  load_webui: False # initialize the web UI during training (the goal is to let you inference during training, but I never found a good way to go about it)

# inferencing settings
inference:
  backend: deepspeed # deepspeed | local, training backend to use
  normalize: False # normalize audio before encoding / after decoding, only enable if you know what you're doing

  weight_dtype: float32 # float32 | float16 | bfloat16, dtype for the model to load under
  amp: False # mixed precision during inferencing

# experimental optimization flags
optimizations:
  injects: False # replace the module in the torch package itself to achieve these
  replace: True # replace the module in the model itself to achieve these

  # bitsandbytes things
  linear: False # enable nn.Linear optimizations
  embedding: False # enable nn.Embedding optimizations
  optimizers: True  # enable torch.optim optimizations

  bitsandbytes: False # use bitsandbytes
  dadaptation: False # use dadaptation
  bitnet: False # use bitnet
  fp8: False # use nvidia/transformer-engine's fp8 AMP

# dataset settings
dataset:
  speaker_name_getter: "lambda p: f'{p.parts[-3]}_{p.parts[-2]}'" # function to eval when fetching the speaker from a name
  speaker_group_getter: "lambda p: f'{p.parts[-3]}'" # function to eval when fetching the group from a name
  # map to classify languages under when preparing a batch (in case the language is not provided in the dataset)
  speaker_languages:
    ja: []

  use_hdf5: True # use HDF5 file to load the dataset from 
  hdf5_flag: r # flag to load the HDF5 file under (should automatically set to `a` when generating the HDF5 dataset)
  use_metadata: True # use generated metadata to help prepare the dataset
  validate: True # cull samples if they are outside the duration threshold

  workers: 2 # worker processes to spawn for the dataloader
  cache: True # cache the dataloader to disk to speed things up
  
  duration_range: [3.0, 5.0] # allowed sample duration in the dataset

  prompt_max_samples: 1 # maximum prompts to sample for the input prompt during training
  prompt_duration_range: [3.0, 6.0] # duration range for the input prompt during training
  prompt_similar_p: 1.0 # odds to instead use a similar utterance instead of a random sample (1 to always do, 0 to never do)

  # not used
  resps_max_samples: 1 # maximum output utterances to sample for the output during training
  resps_append_p: 0.0 # odds to append another utterance to the output utterance sample

  sample_type: path # path | speaker | group, type to sample the paths from (by path, speaker, or group)
  sample_order: duration # duration | anything else, method of ordering the paths (duration is by duration, any other value will interleave reorder)
  sample_max_duration_batch: 0 # used when above = duration, 120 seconds per batch at 12GiB of VRAM works
  sample_shuffle: False # shuffle indices in the dataloader (avoid using with sample_order: duration and sample_max_duration_batch: 0)

  retokenize_text: False # do not rely on AOT'd tokens from the dataset, instead tokenize JIT (in case you botch your tokenizer during dataset preparation and don't want to recreate it)

  tasks_list: [ "tts", "stt" ] # , [ "tts", "tts-c", "ns", "sr", "tse", "cse", "nse", "stt" ], determines which tasks to randomly pick for a sample

  training: [] # paths for the training dataset
  validation: [] # paths for the validation dataset
  noise: [] # paths for the noise dataset (unused, but for the above tasks that call for injecting noise)