scripts/lora_finetune_phi4.yaml

# Config for single device LoRA finetuning in lora_finetune_single_device.py
# using a Phi4 (14B) model
#
# This config assumes that you've run the following command before launching
# this run:
#   tune download microsoft/phi-4 --output-dir /tmp/phi-4 --hf-token <HF_TOKEN>
#
# To launch on a single device, run the following command from root:
#   tune run lora_finetune_single_device --config phi4/lora_single_device
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training
# you can run:
#   tune run lora_finetune_single_device --config phi4/lora_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works only for training on single device.

# Model arguments
model:
  _component_: torchtune.models.phi4.lora_phi4_14b
  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

# Tokenizer
tokenizer:
  _component_: torchtune.models.phi4.phi4_14b_tokenizer
  vocab_path:  {{model_dir}}/vocab.json
  merges_path:  {{model_dir}}/merges.txt
  max_seq_len: null


# Checkpointer
checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: {{model_dir}}
  checkpoint_files: [
    model-00001-of-00006.safetensors,
    model-00002-of-00006.safetensors,
    model-00003-of-00006.safetensors,
    model-00004-of-00006.safetensors,
    model-00005-of-00006.safetensors,
    model-00006-of-00006.safetensors,
  ]
  recipe_checkpoint: null
  output_dir: {{model_output_dir}}
  model_type: PHI3_MINI
resume_from_checkpoint: False
save_adapter_weights_only: False


# Dataset
dataset:
    _component_: torchtune.datasets.instruct_dataset
    source: json
    data_files: {{train_path}}
    column_map:
        input: instruction
        output: output
    train_on_input: False
    packed: False
    split: train
seed: null
shuffle: True

# Fine-tuning arguments
epochs: 1
max_steps_per_epoch: null
batch_size: 2
gradient_accumulation_steps: 8
optimizer:
  _component_: torch.optim.AdamW
  fused: True
  weight_decay: 0.01
  lr: 3e-4
lr_scheduler:
  _component_: torchtune.training.get_cosine_schedule_with_warmup
  num_warmup_steps: 100
loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
clip_grad_norm: null
compile: False


# Training env
device: cuda

# Memory management
enable_activation_checkpointing: True  # True reduces memory
enable_activation_offloading: False  # True reduces memory

# Reduced precision
dtype: bf16

# Logging
output_dir: {{log_dir}}/lora_finetune_output
metric_logger:
  _component_: torchtune.training.metric_logging.{{metric_logger}}
  log_dir: {{log_dir}}/training_logs
log_every_n_steps: 1
log_peak_memory_stats: False

# Showcase the usage of PyTorch profiler
# Set enabled to False as it's only needed for debugging training
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: False

  #Output directory of trace artifacts
  output_dir: {{log_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: 5
  active_steps: 2
  num_cycles: 1