attention/mha.yaml

checkpoints:
  checkpoint_interval: 1000
  checkpoints_path: /scratch/loubna/checkpoints
  checkpoints_path_is_shared_file_system: false
  load_lr_scheduler: true
  load_optimizer: true
  resume_checkpoint_path: null
  save_final_state: true
  save_initial_state: false
data_stages:
- data:
    dataset:
      dataset_read_path:
      - /fsx/loubna/projects_v2/data/blog/fineweb-edu
      - /fsx/loubna/projects_v2/data/blog/finemath-3plus
      - /fsx/loubna/projects_v2/data/blog/stack-edu-python
      dataset_folder:
      - s3://smollm3/datasets/llama_tokenized-global-chunks/fineweb-edu/fineweb-edu/
      - s3://smollm3/datasets/llama_tokenized-individual-chunks/finemath/
      - s3://smollm3/datasets/llama_tokenized-individual-chunks/stack-edu-Python/
      dataset_weights:
      - 0.7
      - 0.1
      - 0.2
      pad_samples_to_global_batch_size: false
      return_positions: true
      token_size_in_bytes: 4
      tokenizer_name: HuggingFaceTB/SmolLM3-3B
      use_old_brrr_dataloader: false
      vocab_size: 128256
    num_loading_workers: 0
    seed: 6
  name: stable
  start_training_step: 1
general:
  benchmark_csv_path: null
  consumed_train_samples: null
  ignore_sanity_checks: true
  project: smollm3-blog
  run: baseline-mha-14layers
  seed: 6
  step: null
logging:
  iteration_step_info_interval: 1
  log_level: info
  log_level_replica: info
model:
  ddp_bucket_cap_mb: 50
  dtype: bfloat16
  init_method:
    std: 0.02
  make_vocab_size_divisible_by: 1
  model_config:
    _attn_implementation: flash_attention_2
    _fused_rms_norm: true
    _fused_rotary_emb: true
    _use_doc_masking: false
    _use_qkv_packed: true
    attention_bias: false
    bos_token_id: 128000
    eos_token_id: 128001
    flex_attention_mask: null
    hidden_act: silu
    hidden_size: 2048
    initializer_range: 0.02
    intermediate_size: 8192
    is_qwen2_config: true
    max_position_embeddings: 4096
    moe_config: null
    num_attention_heads: 32
    num_hidden_layers: 14
    num_key_value_heads: 32
    pad_token_id: null
    pretraining_tp: 1
    rms_norm_eps: 1.0e-05
    rope_interleaved: false
    rope_scaling: null
    rope_theta: 50000.0
    sliding_window_size: null
    tie_word_embeddings: true
    use_cache: true
    vocab_size: 128256
    no_rope_layer: null
optimizer:
  accumulate_grad_in_fp32: true
  clip_grad: 1.0
  learning_rate_scheduler:
    learning_rate: 0.0005
    lr_decay_starting_step: 2000
    lr_decay_steps: 28000
    lr_decay_style: cosine
    lr_warmup_steps: 2000
    lr_warmup_style: linear
    min_decay_lr: 5.0e-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
  weight_decay_exclude_named_params: null
  zero_stage: 0
parallelism:
  context_parallel_size: 1
  dp: 8
  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:
  remove_after_upload: true
  s5cmd_concurrency: 5
  s5cmd_numworkers: 16
  s5cmd_path: /fsx/loubna/.venv-2-6-cu124/bin/s5cmd
  upload_s3_path: s3://smollm3/blogpost-ablations/baseline-mha-14layers
tokenizer:
  tokenizer_max_length: 4096
  tokenizer_name_or_path: HuggingFaceTB/SmolLM3-3B
  tokenizer_revision: null
metrics_logging:
  log_level: 1
  log_detail_interval: 100
tokens:
  batch_accumulation_per_replica: 16
  limit_test_batches: 0
  limit_val_batches: 0
  micro_batch_size: 3 # 1.5M tokens
  sequence_length: 4096
  train_steps: 30000 # 30B tokens
  val_check_interval: 100
lighteval:
  slurm_script_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/launch-config"  # Default path for launch scripts
  logs_path: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/logs"  # Default path for evaluation logs
  local_checkpoint_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/evals-ckpt"  # Default path for temporary checkpoint storage. Will store under {local_checkpoint_dir}/{run_name}/{step}
  nanotron_path: "/fsx/loubna/projects_v2/smollm3/nanotron"
  output_dir: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_results/results"
  s3_save_path: "s3://smollm3/blogpost-evals"
  upload_to_wandb: true
  wandb_project: smollm3-blog-evals
  wandb_entity: huggingface
  parallelism:
    dp: 8
    pp: 1
    tp: 1
    tp_linear_async_communication: true
  batch_size: 1  # Optional batch size for evaluation
  eval_config_override: "/fsx/loubna/projects_v2/smollm3/nanotron/ablations/eval_configs/ablations_blog.yaml"  # The evaluation config file to use
  eval_interval: 1000
  eval_interval_file: null
  slurm:
    gpus_per_node: 8
    partition: "hopper-prod"
    hf_cache: "/fsx/loubna/.cache/huggingface"
    cpus_per_task: 88
    qos: "normal"
    time: "01:59:00"
    reservation: null