NeMo/examples/nlp/language_modeling/conf/megatron_gpt_config.yaml

name: megatron_gpt
restore_from_path: null # used when starting from a .nemo file

trainer:
  gpus: 1
  num_nodes: 1
  accelerator: ddp
  precision: 16
  logger: False # logger provided by exp_manager
  checkpoint_callback: False
  replace_sampler_ddp: False
  max_epochs: null
  max_steps: 100000 # consumed_samples = global_step * micro_batch_size * data_parallel_size * accumulate_grad_batches
  log_every_n_steps: 10
  val_check_interval: 100
  limit_val_batches: 50
  limit_test_batches: 500
  accumulate_grad_batches: 1
  gradient_clip_val: 1.0

exp_manager:
  explicit_log_dir: null
  exp_dir: null
  name: megatron_gpt
  create_wandb_logger: True
  wandb_logger_kwargs:
    project: null
    name: null
  resume_if_exists: True
  resume_ignore_no_checkpoint: True
  create_checkpoint_callback: True
  checkpoint_callback_params:
    monitor: val_loss
    save_top_k: 10
    mode: min
    always_save_nemo: False # saves nemo file during validation, not implemented for model parallel
    filename: 'megatron_gpt--{val_loss:.2f}-{step}-{consumed_samples}'


model:
  # model parallelism 
  micro_batch_size: 4
  tensor_model_parallel_size: 1

  # model architecture
  encoder_seq_length: 512
  max_position_embeddings: 512
  num_layers: 12
  hidden_size: 768
  ffn_hidden_size: 3072 # Transformer FFN hidden size. Usually 4 * hidden_size.
  num_attention_heads: 12
  init_method_std: 0.02 # Standard deviation of the zero mean normal distribution used for weight initialization.')
  hidden_dropout: 0.1 # Dropout probability for hidden state transformer.
  kv_channels: null # Projection weights dimension in multi-head attention. Set to hidden_size // num_attention_heads if null
  apply_query_key_layer_scaling: True # scale Q * K^T by 1 / layer-number.
  layernorm_epsilon: 1e-5
  make_vocab_size_divisible_by: 128 # Pad the vocab size to be divisible by this value for computation efficiency.
  pre_process: True # add embedding
  post_process: True # add pooler

  tokenizer:
    library: 'megatron'
    type: 'GPT2BPETokenizer'
    model: null
    vocab_file: null
    merge_file: null 

  # precision
  native_amp_init_scale: 4294967296 # 2 ** 32
  native_amp_growth_interval: 1000
  fused_fp16: True # False if using fp32 or bf16
  fused_bf16: False # True if using bf16
  fp32_residual_connection: False # Move residual connections to fp32
  fp16_lm_cross_entropy: False # Move the cross entropy unreduced loss calculation for lm head to fp16


  # miscellaneous
  seed: 1234
  use_cpu_initialization: False # Init weights on the CPU (slow for large models)
  onnx_safe: False # Use work-arounds for known problems with Torch ONNX exporter.

  # not implemented in NeMo yet
  activations_checkpoint_method: null # 'uniform', 'block'
  activations_checkpoint_num_layers: 1 

  data:
    # Path to data must be specified by the user.
    # can override from the CLI: "model.data.data_prefix=[.5,/raid/data/pile/my-gpt3_00_text_document,.5,/raid/data/pile/my-gpt3_01_text_document]",
    # Or see example below: 
    # data_prefix: 
    #   - .5
    #   - /raid/data/pile/my-gpt3_00_text_document
    #   - .5
    #   - /raid/data/pile/my-gpt3_01_text_document
    data_prefix: ???
    data_impl: mmap
    splits_string: 900,50,50
    seq_length: 1024
    skip_warmup: True
    num_workers: 0
    dataloader_type: single # cyclic
    reset_position_ids: False # Reset position ids after end-of-document token
    reset_attention_mask: False # Reset attention mask after end-of-document token
    eod_mask_loss: False # Mask loss for the end of document tokens
  
  optim:
    name: adam
    lr: 2e-4
    weight_decay: 0.01 
    betas: 
    - 0.9
    - 0.98
    sched:
      name: CosineAnnealing
      warmup_steps: 500
      constant_steps: 50000
      min_lr: 2e-5