algo:
  T: 0
  backbone: dit_legacy
  causal_attention: false
  loss_type: elbo
  name: duo_base
  parameterization: mean
  subs_masking: false
  time_conditioning: true
  adaLN: true
data:
  mode: phase2
  tokenizer_name_or_path: tokenizer/tiktokenizer.json
  train: nvidia
  valid: nvidia
eval:
  checkpoint_path: model.ckpt
  compute_generative_perplexity: false
  compute_perplexity_on_sanity: false
  disable_ema: true
  gen_ppl_eval_model_name_or_path: gpt2-large
  generate_samples: false
  generated_samples_path: samples.json
  perplexity_batch_size: 8
loader:
  batch_size: 2
  eval_batch_size: 1
  eval_global_batch_size: 1024
  global_batch_size: 1024
  n_chunks: 1
  num_workers: 224
  pin_memory: true
model:
  cond_dim: 128
  dropout: 0
  hidden_size: 2176
  length: 2048
  n_blocks: 28
  n_heads: 17
  name: 2121M
  scale_by_sigma: true
  tie_word_embeddings: false
  type: ddit
noise:
  denoiser_latent_conditioning: -1
  eps: 0
  freeze_decoder: false
  freeze_encoder: false
  parameterization: log-linear
  type: log-linear
optim:
  beta1: 0.9
  beta2: 0.95
  eps: 1.0e-08
  lr: 0.0004
  min_lr: 0.0001
  weight_decay: 0.1
prior:
  latent_height: 0
  latent_width: 0
  type: none
sampling:
  kv_cache: false
  noise_removal: ancestral
  num_log_samples: 2
  num_samples: 1024
  p_nucleus: 1
  predictor: ancestral
  steps: 1000
  use_float64: true
seed: 1
training:
  antithetic_sampling: true
  ema: -1
  finetune_path: ''
  flops_1e18: 1
  sampling_eps: 0.001
  ssl_ratio: 0.01