# Phi-4-multimodal-instruct training config for SFT finetuning.
#
# Phi-4-multimodal-instruct is a multimodal model that combines text, visual, and audio # inputs. It uses a "Mixture of LoRAs" approach, allowing you to plug in adapters for
# each modality without needing to retrain the base model.
#
# Important Note: Oumi has currently integrated and fully tested Phi-4 for vision and
# text modalities only (!).
#
# 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/phi4/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: "microsoft/Phi-4-multimodal-instruct"
  torch_dtype_str: "bfloat16"
  model_max_length: 4096
  trust_remote_code: True
  attn_implementation: "flash_attention_2" # The model requires Flash Attention.

  # The model by default freezes the following audio/image-related modules:
  #   model.embed_tokens_extend.audio_embed
  #   model.embed_tokens_extend.image_embed

data:
  train:
    collator_name: "vision_language_sft"
    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: "microsoft/Phi-4-multimodal-instruct"
          return_tensors: True

      # Below are examples of other vision SFT datasets
      # - dataset_name: "HuggingFaceH4/llava-instruct-mix-vsft"
      #   split: "train"
      #   shuffle: True
      #   seed: 42
      #   trust_remote_code: True
      #   transform_num_workers: "auto"
      #   dataset_kwargs:
      #     processor_name: "microsoft/Phi-4-multimodal-instruct"
      #     return_tensors: True
      # - dataset_name: "coco_captions"
      #   split: "train"
      #   trust_remote_code: True
      #   dataset_kwargs:
      #     processor_name: "microsoft/Phi-4-multimodal-instruct"
      #     return_tensors: True
      # - dataset_name: vision_language_jsonl
      #   dataset_path: "training.jsonl"  # See notebook for example how to generate this file
      #   dataset_kwargs:
      #     data_column: "messages"
      #     processor_name: "microsoft/Phi-4-multimodal-instruct"

training:
  output_dir: "output/vlm_finetuned"
  trainer_type: "TRL_SFT"
  enable_gradient_checkpointing: True
  per_device_train_batch_size: 2
  gradient_accumulation_steps: 8
  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: True

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

  logging_steps: 10
  dataloader_main_process_only: True
  dataloader_num_workers: 4
  dataloader_prefetch_factor: 8
  include_performance_metrics: True
  enable_wandb: True