The following example shows a configuration file for fine-tuning with torchtune.

File: 70B_lora.yaml
# Model Arguments
model:
  _component_: torchtune.models.llama3.lora_llama3_70b
  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
  apply_lora_to_mlp: False
  apply_lora_to_output: False
  lora_rank: 16
  lora_alpha: 32
 
tokenizer:
  _component_: torchtune.models.llama3.llama3_tokenizer
  path: /work/torchtune/Meta-Llama-3-70B/original/tokenizer.model
 
checkpointer:
  _component_: torchtune.utils.FullModelHFCheckpointer
  checkpoint_dir:  /work/torchtune/Meta-Llama-3-70B
  checkpoint_files: [
    model-00001-of-00030.safetensors,
    model-00002-of-00030.safetensors,
    model-00003-of-00030.safetensors,
    model-00004-of-00030.safetensors,
    model-00005-of-00030.safetensors,
    model-00006-of-00030.safetensors,
    model-00007-of-00030.safetensors,
    model-00008-of-00030.safetensors,
    model-00009-of-00030.safetensors,
    model-00010-of-00030.safetensors,
    model-00011-of-00030.safetensors,
    model-00012-of-00030.safetensors,
    model-00013-of-00030.safetensors,
    model-00014-of-00030.safetensors,
    model-00015-of-00030.safetensors,
    model-00016-of-00030.safetensors,
    model-00017-of-00030.safetensors,
    model-00018-of-00030.safetensors,
    model-00019-of-00030.safetensors,
    model-00020-of-00030.safetensors,
    model-00021-of-00030.safetensors,
    model-00022-of-00030.safetensors,
    model-00023-of-00030.safetensors,
    model-00024-of-00030.safetensors,
    model-00025-of-00030.safetensors,
    model-00026-of-00030.safetensors,
    model-00027-of-00030.safetensors,
    model-00028-of-00030.safetensors,
    model-00029-of-00030.safetensors,
    model-00030-of-00030.safetensors,
  ]
  recipe_checkpoint: null
  output_dir: /tmp/Meta-Llama-3-70B-Instruct
  model_type: LLAMA3
resume_from_checkpoint: False
 
# Dataset and Sampler
dataset:
  _component_: torchtune.datasets.databricks_dolly_15k
  train_on_input: True
seed: null
shuffle: True
batch_size: 2
 
# Optimizer and Scheduler
optimizer:
  _component_: torch.optim.AdamW
  weight_decay: 0.01
  lr: 3e-4
lr_scheduler:
  _component_: torchtune.modules.get_cosine_schedule_with_warmup
  num_warmup_steps: 100
 
loss:
  _component_: torch.nn.CrossEntropyLoss
 
# Training
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 1
 
# Logging
output_dir: /tmp/Meta-Llama-3-70B-Instruct
metric_logger:
  _component_: torchtune.utils.metric_logging.TensorBoardLogger
  log_dir: ${output_dir}
  organize_logs: true
log_every_n_steps: 1
log_peak_memory_stats: False
 
# Environment
device: cuda
dtype: bf16
enable_activation_checkpointing: True