# SFT config for CALM 70B.

model:
  model_name: "meta-llama/Llama-3.3-70B-Instruct"
  model_max_length: 4096
  torch_dtype_str: "bfloat16"
  attn_implementation: "sdpa"
  load_pretrained_weights: True
  trust_remote_code: True
  tokenizer_pad_token: "<|finetune_right_pad_id|>"
  enable_liger_kernel: True

data:
  train:
    datasets:
      - dataset_name: "text_sft_jsonl"
        dataset_path: "/path/to/training/dataset.jsonl"
        shuffle: True
        seed: 42
    collator_name: "text_completions_only_with_padding"
    target_col: "messages"
    use_async_dataset: True
    seed: 42
  validation:
    datasets:
      - dataset_name: "text_sft_jsonl"
        dataset_path: "/path/to/validation/dataset.jsonl"
    collator_name: "text_completions_only_with_padding"
    target_col: "messages"
    use_async_dataset: True
    seed: 42

training:
  trainer_type: "TRL_SFT"
  save_steps: 688
  num_train_epochs: 1
  per_device_train_batch_size: 7
  gradient_accumulation_steps: 1
  eval_strategy: "steps"
  eval_steps: 344
  per_device_eval_batch_size: 1

  enable_gradient_checkpointing: True
  gradient_checkpointing_kwargs:
    use_reentrant: False
  ddp_find_unused_parameters: False
  optimizer: "adamw_torch_fused"
  learning_rate: 4.0e-05
  warmup_steps: 24
  weight_decay: 0.01
  max_grad_norm: 10
  compile: False

  dataloader_num_workers: "auto"
  dataloader_prefetch_factor: 32

  logging_steps: 8
  log_model_summary: False
  empty_device_cache_steps: 1
  include_performance_metrics: True
  output_dir: "output/llama70b.sft"
  enable_wandb: True

fsdp:
  enable_fsdp: True
  forward_prefetch: True
  use_orig_params: True
  cpu_offload: True
  auto_wrap_policy: "TRANSFORMER_BASED_WRAP"
  transformer_layer_cls: "LlamaDecoderLayer"