config/solver/magnet/magnet_32khz.yaml

# @package __global__

# This is the training loop solver
# for the base MusicGen model (text-to-music)
# on monophonic audio sampled at 32 kHz
defaults:
  - musicgen/default
  - /model: lm/musicgen_lm
  - override /dset: audio/default
  - _self_

lm_model: transformer_lm_magnet
solver: magnet

autocast: true
autocast_dtype: float16

# EnCodec large trained on mono-channel music audio sampled at 32khz
# with a total stride of 640 leading to 50 frames/s.
# rvq.n_q=4, rvq.bins=2048, no quantization dropout
# (transformer_lm card and n_q must be compatible)
compression_model_checkpoint: //pretrained/facebook/encodec_32khz

efficient_attention_backend: xformers # restricted attention implementation supports only xformers at the moment

channels: 1
sample_rate: 32000

deadlock:
  use: true  # deadlock detection

dataset:
  batch_size: 192  # 32 GPUs
  sample_on_weight: false  # Uniform sampling all the way
  sample_on_duration: false  # Uniform sampling all the way

optim:
  epochs: 500
  optimizer: dadam
  lr: 1
  ema:
    use: true
    updates: 10
    device: cuda

logging:
  log_tensorboard: true

schedule:
  lr_scheduler: cosine
  cosine:
    warmup: 4000
    lr_min_ratio: 0.0
    cycle_length: 1.0

codebooks_pattern:
  modeling: parallel
  parallel:
    empty_initial: -1
  
transformer_lm:
  card: 2048
  causal: false
  subcodes_context: 5
  compression_model_framerate: 50  # NOTE: Must match the actual frame rate of the used compression model 
  segment_duration: 0
  span_len: -1

masking:
  span_len: 3

generate:
  lm: 
    max_prompt_len: null
    max_gen_len: null
    remove_prompts: false
    use_sampling: true
    temp: 3.0
    top_k: 0
    top_p: 0.9
    max_cfg_coef: 10.0
    min_cfg_coef: 1.0
    decoding_steps: [60, 10, 10, 10]
    anneal_temp: true
    span_scoring: 'max'
    span_arrangement: 'nonoverlap'
    prompted_samples: false
    samples:
      prompted: false
      unprompted: true