11B_full.yaml

# Config for single device full finetuning in full_finetune_single_device.py
# using a Llama3.2 11B Vision Instruct model
#
# This config assumes that you've run the following command before launching:
#   tune download meta-llama/Llama-3.2-11B-Vision-Instruct --output-dir /tmp/Llama-3.2-11B-Vision-Instruct --ignore-patterns "original/consolidated*"
#
# To launch on a single device, run the following command from root:
#   tune run --nproc_per_node 4 full_finetune_distributed --config llama3_2_vision/11B_full
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training:
#    tune run --nproc_per_node 4 full_finetune_distributed --config llama3_2_vision/11B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works best when the model is being fine-tuned on 2+ GPUs.
# Single device full finetuning requires more memory optimizations. It's
# best to use 11B_full_single_device.yaml for those cases.

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

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

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

# Dataset
dataset:
  packed: False # Set to true for great speed ups
  _component_: torchtune.datasets.multimodal.the_cauldron_dataset
  subset: ocrvqa
seed: null
shuffle: True
collate_fn: torchtune.data.padded_collate_tiled_images_and_mask

# Fine-tuning arguments
epochs: 1
max_steps_per_epoch: null
batch_size: 2
gradient_accumulation_steps: 4
optimizer:
  _component_: torch.optim.AdamW
  lr: 2e-5
  fused: True
optimizer_in_bwd: False # Set to True to use less memory. Requires gradient_accumulation_steps=1.

loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
clip_grad_norm: 1.0
compile: False # pytorch compile, set to true for perf/memory improvement

# Training env
device: cuda

# Memory management
enable_activation_checkpointing: True
custom_sharded_layers: ['decoder.tok_embeddings']
dtype: bf16

# Logging
output_dir: /tmp/full-llama3.2-vision--finetune
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: /tmp/Llama-3.2-11B-Vision-Instruct/logs
log_every_n_steps: 1
log_peak_memory_stats: True