model:
  _component_: models.lora_mmllama3_8b
  lora_attn_modules:
  - q_proj
  - v_proj
  apply_lora_to_mlp: false
  apply_lora_to_output: false
  lora_rank: 8
  lora_alpha: 16
  perception_tokens: 2
  use_clip: false
tokenizer:
  _component_: models.a2a_tokenizer
  path: models/tokenizer.model
checkpointer:
  _component_: torchtune.utils.FullModelMetaCheckpointer
  checkpoint_dir: checkpoints/pruning
  checkpoint_files:
  - meta_model_3.pt
  adapter_checkpoint: null
  recipe_checkpoint: null
  output_dir: output_checkpoints/experiment_1
  model_type: LLAMA3
resume_from_checkpoint: false
interim_checkpoint_steps: 5000
interim_gen_steps: null
max_new_tokens: 111
temperature: 0.6
top_k: 333
dataset:
  _component_: ds.EvenBatcher
  buffer_size: 36
  dataset:
    _component_: ds.RoundRobinDataset
    datasets:
    - _component_: ds.OmegaVideoCaptionDataset
      length: 500000
    - _component_: ds.LlavaInstructDataset
      dataset_path: ds/coco_llava_instruct/output.parquet
      train_on_input: false
    - _component_: ds.LlavaInstructDataset
      dataset_path: ds/vision_flan/output.parquet
      train_on_input: false
    - _component_: ds.CaptionInstructDataset
      dataset_path: ds/sam_llava/output.parquet
      train_on_input: false
seed: null
shuffle: true
batch_size: 4
optimizer:
  _component_: torch.optim.AdamW
  weight_decay: 0.01
  lr: 3.0e-07
lr_scheduler:
  _component_: torchtune.modules.get_cosine_schedule_with_warmup
  num_warmup_steps: 100
loss:
  _component_: torch.nn.CrossEntropyLoss
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 16
compile: false
output_dir: /tmp/lora_finetune_output
metric_logger:
  _component_: torchtune.utils.metric_logging.DiskLogger
  log_dir: ${output_dir}
log_every_n_steps: null
device: cuda
dtype: bf16
enable_activation_checkpointing: false
profiler:
  _component_: torchtune.utils.profiler
  enabled: false
inference:
  prompt_template: 'Video:

    {video}

    Caption the previous video.'
  max_new_tokens: 333
  temperature: 0.6
  top_k: 333
  quantizer: null
gradient-accumulation-steps: 16