checkpoints:
  checkpoint_interval: 10000
  checkpoints_path: checkpoints/360M_UT_original_withcos_early0.01
  checkpoints_path_is_shared_file_system: false
  load_lr_scheduler: false
  load_optimizer: false
  resume_checkpoint_path: /mnt/bn/ridger1/yxy/SmolLM2-nanotron-ckpt/360M/pre-decay
  save_final_state: false
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_folder:
      - /mnt/bn/ridger1/yxy/datasets/finewebedu-dedup
      - /mnt/bn/ridger1/datasets/cosmopedia-v2
      - /mnt/bn/ridger1/yxy/datasets/megamath/megamath-text-code-block
      - /mnt/bn/ridger1/yxy/datasets/megamath/megamath-qa
      - /mnt/bn/ridger1/datasets/megamath/megamath-translated-code
      - /mnt/bn/ridger1/yxy/datasets/megamath/megamath-web-pro
      dataset_weights:
      - 0.545
      - 0.08
      - 0.25
      - 0.035
      - 0.035
      - 0.075
      token_size_in_bytes: 2
      tokenizer_name: HuggingFaceTB/cosmo2-tokenizer
      vocab_size: 49152
    num_loading_workers: 1
    seed: 42
  name: stable phase
  start_training_step: 1
general:
  benchmark_csv_path: null
  consumed_train_samples: 51200000
  ignore_sanity_checks: true
  project: diffUT
  run: 360M_32_with_cos_constant_and_decay_early0.1
  seed: 8
  step: 100000
lighteval: null
logging:
  iteration_step_info_interval: 1
  log_level: info
  log_level_replica: info
model:
  ddp_bucket_cap_mb: 25
  dtype: bfloat16
  init_method:
    std: 0.041666666666666664
  make_vocab_size_divisible_by: 1
  model_config:
    _attn_implementation: flash_attention_2
    attention_bias: false
    bos_token_id: 0
    early_ratio: 0.01
    eos_token_id: 0
    hidden_act: silu
    hidden_size: 960
    initializer_range: 0.02
    intermediate_size: 2560
    is_llama_config: true
    max_position_embeddings: 2048
    num_attention_heads: 15
    num_hidden_layers: 32
    num_key_value_heads: 5
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_interleaved: false
    rope_scaling: null
    rope_theta: 10000.0
    tie_word_embeddings: true
    unroll: true
    unroll_end: 25
    unroll_start: 5
    unroll_type: 1
    use_cache: true
    vocab_size: 49152
    z_loss_coefficient: 0.0001
    z_loss_enabled: false
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.003
    lr_decay_starting_step: 10000
    lr_decay_steps: 100000
    lr_decay_style: linear
    lr_warmup_steps: 1024
    lr_warmup_style: linear
    min_decay_lr: 0.0003
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.01
  weight_decay_exclude_named_params: []
  zero_stage: 1
parallelism:
  context_parallel_size: 1
  dp: 64
  expert_parallel_size: 1
  moe_layer_recompute: false
  pp: 1
  pp_engine: 1f1b
  recompute_layer: false
  tp: 1
  tp_linear_async_communication: true
  tp_mode: REDUCE_SCATTER
  tp_recompute_allgather: true
profiler: null
s3_upload: null
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: HuggingFaceTB/cosmo2-tokenizer
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 1
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 8
  sequence_length: 2048
  train_steps: 102400
  val_check_interval: 10000