llama_2B.yaml

checkpoints:
  checkpoint_interval: 2000
  checkpoints_path: /scratch/project_462000353/villekom/checkpoints/nanotron
  checkpoints_path_is_shared_file_system: true
  resume_checkpoint_path: null
  save_initial_state: false
data_stages:
  - data:
      dataset:
        dataset_folder: /scratch/project_462000353/data/nanosets/fineweb-edu/350BT
      num_loading_workers: 7
      seed: 42
    name: Stable Training Stage
    start_training_step: 1
general:
  benchmark_csv_path: null
  consumed_train_samples: null
  ignore_sanity_checks: true
  project: fineweb-ablations
  run: llama
  seed: 42
  step: null
lighteval: null
logging:
  iteration_step_info_interval: 10
  log_level: info
  log_level_replica: warning
model:
  ddp_bucket_cap_mb: 25
  dtype: bfloat16
  init_method:
    std: 0.02
  make_vocab_size_divisible_by: 1
  model_config:
    is_llama_config: true
    bos_token_id: 1
    eos_token_id: 2
    hidden_act: silu
    hidden_size: 2048
    initializer_range: 0.02
    intermediate_size: 8192
    max_position_embeddings: 2048
    num_attention_heads: 32
    num_hidden_layers: 24
    num_key_value_heads: 32
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_scaling: null
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 50272
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.0003
    lr_decay_starting_step: null
    lr_decay_steps: 10000
    lr_decay_style: cosine
    lr_warmup_steps: 500
    lr_warmup_style: linear
    min_decay_lr: 3.0e-05
  optimizer_factory:
    adam_beta1: 0.9
    adam_beta2: 0.95
    adam_eps: 1.0e-08
    name: adamW
    torch_adam_is_fused: true
  weight_decay: 0.1
  zero_stage: 0
parallelism:
  dp: 64 #This should be equal to amount of gpu's if tp and pp == 1
  expert_parallel_size: 1
  pp: 1
  pp_engine: 1f1b
  tp: 1
  tp_linear_async_communication: true
  tp_mode: REDUCE_SCATTER
profiler: null
tokenizer:
  tokenizer_max_length: null
  tokenizer_name_or_path: HuggingFaceFW/ablation-model-fineweb-edu
  tokenizer_revision: null
tokens:
  batch_accumulation_per_replica: 4
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 4
  sequence_length: 2048
  train_steps: 167000 #350B tokens if on 1M global batch size on fineweb-edu
  val_check_interval: 100