90B_qlora.yaml

# Config for multi-device QLoRA finetuning in lora_finetune_distributed.py
# using a Llama3.2 90B Vision Instruct model
#
# This config assumes that you've run the following command before launching:
#   tune download meta-llama/Llama-3.2-90B-Vision-Instruct --output-dir /tmp/Llama-3.2-90B-Vision-Instruct --ignore-patterns "original/consolidated*"
#
# To launch on 4 devices, run the following command from root:
#   tune run --nproc_per_node 4 lora_finetune_distributed --config llama3_2_vision/90B_qlora
#
# 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 lora_finetune_distributed --config llama3_2_vision/90B_qlora checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works best when the model is being fine-tuned on 4+ GPUs.

# Model arguments
model:
  _component_: torchtune.models.llama3_2_vision.qlora_llama3_2_vision_90b
  decoder_trainable: "frozen"
  encoder_trainable: "lora"
  fusion_trainable: "lora"
  lora_attn_modules: ['q_proj', 'v_proj', 'output_proj']
  apply_lora_to_mlp: True
  apply_lora_to_output: False
  lora_rank: 8  # higher increases accuracy and memory
  lora_alpha: 16  # usually alpha=2*rank
  lora_dropout: 0.0
  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-3.2-90B-Vision-Instruct/original/tokenizer.model
  image_size: 560
  max_seq_len: 8192

# Checkpointer
checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /tmp/Llama-3.2-90B-Vision-Instruct/
  checkpoint_files:
    filename_format: model-{}-of-{}.safetensors
    max_filename: "00037"
  recipe_checkpoint: null
  output_dir: /tmp/Llama-3.2-90B-Vision-Instruct/
  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.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: 8  # Use to increase virtual batch size
optimizer:
  _component_: torch.optim.AdamW
  fused: True
  weight_decay: 0.01
  lr: 1e-4
lr_scheduler:
  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
  num_warmup_steps: 100
loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
clip_grad_norm: 1.0
compile: False  # pytorch compile, set to true for better perf/memory

# Training env
device: cuda

# Memory management
enable_activation_checkpointing: True  # True reduces memory
dtype: bf16

# Logging
output_dir: /tmp/qlora-llama3.2-vision-finetune
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: /tmp/Llama-3.2-90B-Vision-Instruct/logs
log_every_n_steps: 1
log_peak_memory_stats: False

# Profiler (disabled)
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: False

  #Output directory of trace artifacts
  output_dir: ${output_dir}/profiling_outputs

  #`torch.profiler.ProfilerActivity` types to trace
  cpu: True
  cuda: True

  #trace options passed to `torch.profiler.profile`
  profile_memory: False
  with_stack: False
  record_shapes: True
  with_flops: False

  # `torch.profiler.schedule` options:
  # wait_steps -> wait, warmup_steps -> warmup, active_steps -> active, num_cycles -> repeat
  wait_steps: 5
  warmup_steps: 3
  active_steps: 2
  num_cycles: 1