defaults:
  - .@model.encoder: megatron_model_base_config
  - .@model.decoder: megatron_model_base_config

name: megatron_ul2
restore_from_path: null # used when starting from a .nemo file

trainer:
  devices: 1
  num_nodes: 1
  accelerator: gpu
  precision: 16
  logger: False # logger provided by exp_manager
  enable_checkpointing: False
  replace_sampler_ddp: False
  max_epochs: -1 # PTL default. In practice, max_steps will be reached first. 
  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: ${name}
  create_wandb_logger: False
  wandb_logger_kwargs:
    project: null
    name: null
  resume_if_exists: True
  resume_ignore_no_checkpoint: True
  create_checkpoint_callback: True
  checkpoint_callback_params:
    monitor: val_loss
    save_top_k: 10
    mode: min
    always_save_nemo: False # saves nemo file during validation, not implemented for model parallel
    filename: '${name}--{val_loss:.2f}-{step}-{consumed_samples}'
    model_parallel_size: ${multiply:${model.tensor_model_parallel_size}, ${model.pipeline_model_parallel_size}}

model:
  # model parallelism 
  micro_batch_size: 4
  global_batch_size: 8 # will use more micro batches to reach global batch size
  tensor_model_parallel_size: 1
  pipeline_model_parallel_size: 1
  resume_from_checkpoint: null # manually set the checkpoint file to load from
  pipeline_model_parallel_split_rank: 0 # rank at which decoder starts.

  # model architecture
  make_vocab_size_divisible_by: 128 # Pad the vocab size to be divisible by this value for computation efficiency.

  megatron_amp_O2: False # use AMP with O2 style mixed precision instead of native amp on-the-fly weight autocasting.
  grad_allreduce_chunk_size_mb: 125
  grad_div_ar_fusion: True # Fuse grad division into torch.distributed.all_reduce
  gradient_as_bucket_view: True # Allocate gradients in a contiguous bucket to save memory (less fragmentation and buffer memory)

  seq_length: 512
  max_position_embeddings: ${.seq_length}

  tokenizer:
    library: 'megatron'
    type: 'BertWordPieceCase'
    model: null
    vocab_file: null
    merge_file: null
    num_sentinel_tokens: 100
    sentencepiece_legacy: True # Legacy=True allows you to add special tokens to sentencepiece tokenizers.

  # weight init
  embedding_init_method_std: 0.02 # Standard deviation of the zero mean normal distribution used for weight initialization.')

  # embedding dropout
  embedding_dropout: 0.1

  # embedding sharing
  share_token_embeddings: True # If True share encoder/decoder embeddings
  share_decoder_tokens_head_embeddings: True # If True share decoder embeddings and decoder projection to logits

  # token head
  tokens_head_bias: True

  # precision
  native_amp_init_scale: 4294967296 # 2 ** 32
  native_amp_growth_interval: 1000
  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)
  apex_transformer_log_level: 30 # Python logging level displays logs with severity greater than or equal to this

  data:
    # Path to data must be specified by the user.
    # can override from the CLI: "model.data.data_prefix=[.5,/raid/data/pile/my-t5_00_text_document,.5,/raid/data/pile/my-t5_01_text_document]",
    # Or see example below:
    # data_prefix: 
    #   - .5
    #   - /raid/data/pile/my-t5_00_text_document
    #   - .5
    #   - /raid/data/pile/my-t5_01_text_document
    data_prefix: ???
    index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix
    data_impl: mmap
    # data_impl_kwargs: # currently used only for text_mmap, csv_mmap (should be data_impl dependant)
    #     # defaults for text_memmap
    #     newline_int: 10 # byte-value of newline (Use ord('\n') to get value)
    #     header_lines: 0 # skip first N header lines
    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
    #     sort_dataset_paths: False # if True datasets will be sorted by name
    #     # defaults for csv_memmap
    #     newline_int: 10 # byte-value of newline
    #     header_lines: 1 # skip first N header lines
    #     workers: null # number of workers when creating missing index files (null defaults to cpu_num // 2)
    #     sort_dataset_paths: False # if True datasets will be sorted by name
    #     data_col: 1 # column to use for data
    #     data_sep: ',' # string to split text into columns
    splits_string: 949,45,5
    seq_length: ${model.seq_length}
    seq_length_dec: ${model.seq_length}
    skip_warmup: True
    num_workers: 0
    dataloader_type: single # cyclic
    masked_lm_prob: 0.15
    extreme_masked_lm_prob: 0.5
    dataset_type: 'ul2'
    short_seq_prob: 0.0
    max_ngram_size: 10
    extreme_max_ngram_size: 128
    extreme_min_ngram_size: 32
    extreme_mean_ngram_size: 64
    ngram_span_length_distribution: 'geometric'
    extreme_ngram_span_length_distribution: 'truncated_normal'
    prefix_lm_pivot_mean: 0.25
    mean_ngram_size: 3
    permutation: False
    whole_word_masking: True
    favor_longer_ngrams: False
    respect_document_boundaries: True # If true, a single training exampl cannot cross document boundaries, increasing the fraction of <pad> tokens within a batch.

  optim:
    name: fused_adam
    lr: 0.0001
    betas:
      - 0.9
      - 0.999
    eps: 1e-8
    weight_decay: 0.01
    sched:
      name: WarmupAnnealing
      min_lr: 0.00001
      last_epoch: -1
      warmup_ratio: 0.01