torchtune_config.yaml

output_dir: /tmp/torchtune/llama4_17Bx16E/lora
model:
  _component_: torchtune.models.llama4.lora_llama4_scout_17b_16e
  decoder_trainable: lora
  encoder_trainable: frozen
  fusion_trainable: lora
  lora_attn_modules:
  - q_proj
  - v_proj
  - output_proj
  apply_lora_to_mlp: true
  apply_lora_to_output: false
  lora_rank: 16
  lora_alpha: 32
  lora_dropout: 0.0
tokenizer:
  _component_: torchtune.models.llama4.llama4_transform
  path: /dev/shm/model_cache/Llama-4-Scout-17B-16E-Instruct/tokenizer.model
  max_seq_len: 2048
  max_num_tiles: 16
checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /dev/shm/model_cache/Llama-4-Scout-17B-16E-Instruct
  checkpoint_files:
    filename_format: model-{}-of-{}.safetensors
    max_filename: '00050'
  recipe_checkpoint: null
  output_dir: ${output_dir}
  model_type: LLAMA4
save_adapter_weights_only: true
resume_from_checkpoint: false
dataset:
  _component_: torchtune.datasets.alpaca_dataset
  packed: true
seed: null
shuffle: true
epochs: 1
max_steps_per_epoch: null
batch_size: 4
gradient_accumulation_steps: 1
optimizer:
  _component_: torch.optim.AdamW
  lr: 2.0e-05
  fused: false
optimizer_in_bwd: false
lr_scheduler:
  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
  num_warmup_steps: 100
loss:
  _component_: torchtune.modules.loss.LinearCrossEntropyLoss
clip_grad_norm: null
device: cuda
enable_activation_checkpointing: true
enable_activation_offloading: false
custom_sharded_layers:
- tok_embeddings
fsdp_cpu_offload: false
compile: false
dtype: bf16
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: ${output_dir}/logs
log_every_n_steps: 1
log_peak_memory_stats: true
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: false