# Full fine-tune config for SmolVLM Instruct, with the TRL trainer.
#
# Requirements:
#   - Log into WandB (`wandb login`) or disable `enable_wandb`
#   - Run `pip install -U flash-attn --no-build-isolation`
#
# Usage:
#   oumi train -c configs/recipes/vision/smolvlm/sft/full/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: "HuggingFaceTB/SmolVLM-Instruct"
  torch_dtype_str: "bfloat16"
  model_max_length: 3072
  trust_remote_code: True
  attn_implementation: "flash_attention_2" # The model requires Flash Attention!
  chat_template: "llava"
  freeze_layers:
    - "vision_model"

data:
  train:
    collator_name: "vision_language_with_padding"
    use_torchdata: True
    datasets:
      - dataset_name: "merve/vqav2-small"
        split: "validation"
        shuffle: True
        seed: 42
        trust_remote_code: True
        transform_num_workers: "auto"
        dataset_kwargs:
          processor_name: "HuggingFaceTB/SmolVLM-Instruct"
          # limit: 4096 # Uncomment to limit dataset size!
          return_tensors: True

training:
  output_dir: "output/vlm_finetuned"
  trainer_type: "TRL_SFT" # Or "OUMI".
  enable_gradient_checkpointing: True
  per_device_train_batch_size: 1 # Must be 1: the model generates variable-sized image features.
  gradient_accumulation_steps: 16
  max_steps: 20

  gradient_checkpointing_kwargs:
    # Reentrant docs: https://pytorch.org/docs/stable/checkpoint.html#torch.utils.checkpoint.checkpoint
    use_reentrant: False
  ddp_find_unused_parameters: False
  empty_device_cache_steps: 2
  compile: False

  optimizer: "adamw_torch_fused"
  learning_rate: 2e-5
  warmup_ratio: 0.03
  weight_decay: 0.0
  lr_scheduler_type: "cosine"

  logging_steps: 5
  save_steps: 0
  dataloader_num_workers: 2
  dataloader_prefetch_factor: 8
  include_performance_metrics: True
  log_model_summary: False
  enable_wandb: True