defaults:
  - _self_
  - /callbacks: [checkpoint_every_n_steps, checkpoint_monitor, learning_rate_monitor]
  - /data: Korean_dataset
  - /model: tiny-ar
  - /strategy: ddp
  - /noise: loglinear
  - /lr_scheduler: constant_warmup

mode: sample_eval  # train / ppl_eval / sample_eval
diffusion: absorbing_state
backbone: ar  # dit / dimamba / ar
parameterization: ar  # subs / d3pm / sedd
time_conditioning: False
T: 0  # 0 (continuous time) / 1000 
subs_masking: False

seed: 1

loader:
  global_batch_size: 32
  eval_global_batch_size: ${.global_batch_size}
  # Note: batch_size and eval_batch_size are **per machine**
  batch_size: ${div_up:${.global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes}}}
  eval_batch_size: 1  
  #${div_up:${.eval_global_batch_size}, ${eval:${trainer.devices} * ${trainer.num_nodes}}}
  num_workers: ${eval:"len(__import__('os').sched_getaffinity(0))"}
  pin_memory: True

sampling:
  predictor: ddpm_cache  # analytic, ddpm, ddpm_cache
  steps: 128
  noise_removal: True
  # TODO(yair): @subham, why aren't these params under `eval`?
  num_sample_batches: 1  # Total samples: `num_gpus` * `loader.eval_batch_size` * num_sample_batches
  num_sample_log: 1
  semi_ar: False
  stride_length: 1
  num_strides: 1

training:
  ema: 0.9999
  antithetic_sampling: True
  importance_sampling: False
  sampling_eps: 1e-3
  change_of_variables: False

eval:
  checkpoint_path: /home/elicer/lhb01/mdlm/outputs/parkseongjun/psjkodata/2025.04.05/051927/checkpoints/best.ckpt  # Used to evaluate a checkpoint after training.
  disable_ema: False
  compute_generative_perplexity: True
  perplexity_batch_size: 8
  compute_perplexity_on_sanity: False
  gen_ppl_eval_model_name_or_path: gpt2-large  # gpt2-large, meta-llama/Llama-2-7b-hf
  generate_samples: True

optim:
  weight_decay: 0.01
  lr: 5e-5
  beta1: 0.9
  beta2: 0.999
  eps: 1e-8

trainer:
  _target_: lightning.Trainer
  accelerator: cuda
  num_nodes: 1
  devices: ${device_count:}
  accumulate_grad_batches: ${div_up:${loader.global_batch_size}, ${eval:${trainer.devices} * ${loader.batch_size} * ${trainer.num_nodes}}}
  gradient_clip_val: 1.0
  precision: 'bf16'
  num_sanity_val_steps: 0
  max_steps: 50000
  log_every_n_steps: 10
  limit_train_batches: 1.0  # train on full dataset, can be used to toggle quick run
  limit_val_batches: 1.0  # validate on full dataset, can be used to toggle quick run
  val_check_interval: 0.5

wandb:
  project: test-ar
  mode: online
  notes: Mulan for text
  resume: must
  group: null
  job_type: null
  name: ar
  id: f12b7c5e-07c9-48ae-96fa-4798823b8492
  tags:
    - ${noise.type}
    - ${data.train}
    - ${data.valid}

hydra:
  run:
    dir: ./outputs/${data.train}/${now:%Y.%m.%d}/${now:%H%M%S}
  job:
    chdir: true

checkpointing:
  # Use custom `save_dir` if, e.g., saving to S3 bucket, otherwise leave this parameter as is
  save_dir: ${cwd:}
  # Note: `checkpoints` path should correspond to `checkpoint_every_n_steps.dirpath`
  resume_from_ckpt: true
  resume_ckpt_path: /home/elicer/lhb01/mdlm/outputs/parkseongjun/psjkodata/2025.04.05/045928/checkpoints/last.ckpt