# Llama 3.1 8B pretraining config.
#
# Requirements:
#   - Log into WandB (`wandb login`) or disable `enable_wandb`
#   - Log into HF: `hf auth login`
#   - Request access to Llama 3.1: https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct
#
# Usage:
#   oumi train -c configs/recipes/llama3_1/pretraining/8b/train.yaml
#
# See Also:
#   - Documentation: https://oumi.ai/docs/en/latest/user_guides/train/train.html
#   - Config class: oumi.core.configs.TrainingConfig
#   - Config source: https://github.com/oumi-ai/oumi/blob/main/src/oumi/core/configs/training_config.py
#   - Other training configs: configs/**/*train.yaml

model:
  model_name: "meta-llama/Llama-3.1-8B"
  chat_template: "chat_ml"
  model_max_length: 2048
  torch_dtype_str: "bfloat16"
  attn_implementation: "sdpa"
  load_pretrained_weights: False
  trust_remote_code: True
  enable_liger_kernel: True # Helps reduce required VRAM

data:
  train:
    datasets:
      - dataset_name: "HuggingFaceFW/fineweb-edu"
        subset: "sample-10BT"
        split: "train"
        dataset_kwargs:
          seq_length: 2048
      # Polaris copy of the dataset:
      # - dataset_name: "/eagle/community_ai/datasets/fineweb-edu/sample-10BT"
      #   subset: "default"
      #   split: "train"
    stream: True
    pack: True

training:
  trainer_type: "TRL_SFT"
  max_steps: 1000
  save_steps: 500
  save_final_model: True
  per_device_train_batch_size: 16
  gradient_accumulation_steps: 1
  enable_gradient_checkpointing: True
  gradient_checkpointing_kwargs:
    use_reentrant: False
  ddp_find_unused_parameters: False
  optimizer: "adamw_torch_fused"
  empty_device_cache_steps: 1
  compile: True
  dataloader_main_process_only: False
  dataloader_num_workers: "auto"
  dataloader_prefetch_factor: 32
  seed: 123
  logging_steps: 10
  log_model_summary: False
  output_dir: "output/fineweb.pt"
  include_performance_metrics: True
  enable_wandb: True

fsdp:
  enable_fsdp: True
  sharding_strategy: "HYBRID_SHARD"
  forward_prefetch: True
  auto_wrap_policy: "TRANSFORMER_BASED_WRAP"
  transformer_layer_cls: "LlamaDecoderLayer"