# Top-level output directory
output_dir: ./outputs/Llama-3.2-11B-Instruct-w2-full

# Model
model:
  _component_: torchtune.models.llama3_2_vision.llama3_2_vision_11b
  decoder_trainable: True
  encoder_trainable: True
  fusion_trainable: True
  image_size: 560 # Make sure this matches the image_size in tokenizer

# Tokenizer / vision transform
tokenizer:
  _component_: torchtune.models.llama3_2_vision.llama3_2_vision_transform
  path: ./Llama-3.2-11B-Vision-Instruct/original/tokenizer.model
  image_size: 560
  max_seq_len: 8192

# Checkpointing
checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: ./Llama-3.2-11B-Vision-Instruct
  checkpoint_files:
    filename_format: model-{}-of-{}.safetensors
    max_filename: "00005"
  recipe_checkpoint: null
  output_dir: ${output_dir}
  model_type: LLAMA3_VISION

resume_from_checkpoint: false
save_adapter_weights_only: False # PeFT formatting not available yet. This will save it in torchtune format only.

# Dataset
dataset:
  _component_: torchtune.datasets.multimodal.vqa_dataset
  source: arrow
  data_files:
    train: "fake_w2_us_tax_form_dataset_train30_test70/train/data-00000-of-00001.arrow"
  split: train
  column_map:
    input: input
    output: ground_truth
    image: image

# General data handling
seed: null
shuffle: true
collate_fn: torchtune.data.padded_collate_tiled_images_and_mask

# Training loop & hyperparams

epochs: 5
max_steps_per_epoch: null
batch_size: 4
gradient_accumulation_steps: 8 # Use to increase effective batch size
# explicit optimizer / scheduler / loss
optimizer:
  _component_: bitsandbytes.optim.PagedAdamW8bit
  lr: 2e-5
optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1

loss:
  _component_: torchtune.modules.loss.LinearCrossEntropyLoss

clip_grad_norm: 1.0
compile: false

# Device & memory
device: cuda
enable_activation_checkpointing: true
dtype: bf16

# Logging

metric_logger:
  _component_: torchtune.training.metric_logging.WandBLogger
  project: llama3_2_w2_extraction
  entity: <your_wandb_entity>
  job_type: full_finetune_single_device
  group: llama-cookbook
log_every_n_steps: 5
save_steps: 100
log_peak_memory_stats: true
log_level: INFO

# Profiler (off by default)
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: false
  output_dir: ${output_dir}/profiling_outputs
  cpu: true
  cuda: true
  profile_memory: false
  with_stack: false
  record_shapes: true
  with_flops: false
  wait_steps: 5
  warmup_steps: 3
  active_steps: 2
  num_cycles: 1