knowledge_distillation_single_device.yaml

# Config for single device knowledge distillation in kd_single_device.py
# using a teacher and student model
#
# This config assumes that you've ran the following commands before launching KD:
# First download the student and teacher models
#   tune download Qwen/Qwen2-0.5B-Instruct --output-dir /tmp/Qwen2-0.5B-Instruct
#   tune download Qwen/Qwen2-1.5B-Instruct --output-dir /tmp/Qwen2-1.5B-Instruct
#
# You get better results using KD if the teacher model has already been fine-tuned on the target dataset:
#   tune run lora_finetune_single_device --config qwen2/1.5B_lora_single_device
#
# To launch on a single device, run the following command from root:
#   tune run knowledge_distillation_single_device --config qwen2/knowledge_distillation_single_device
#
# This config works only for distilling on a single device.


# Model Arguments
model:
  _component_: torchtune.models.qwen2.lora_qwen2_0_5b
  lora_attn_modules: ['q_proj', 'k_proj', 'v_proj']
  apply_lora_to_mlp: False
  lora_rank: 32
  lora_alpha: 64

teacher_model:
  _component_: torchtune.models.qwen2.qwen2_1_5b

tokenizer:
  _component_: torchtune.models.qwen2.qwen2_tokenizer
  path: /tmp/Qwen2-0.5B-Instruct/vocab.json
  merges_file: /tmp/Qwen2-0.5B-Instruct/merges.txt
  max_seq_len: null

checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /tmp/Qwen2-0.5B-Instruct
  checkpoint_files: [
    model.safetensors
  ]
  recipe_checkpoint: null
  output_dir: /tmp/Qwen2-0.5B-Instruct
  model_type: QWEN2

teacher_checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /tmp/Qwen2-1.5B-Instruct
  checkpoint_files: [
    model.safetensors
  ]
  recipe_checkpoint: null
  output_dir: /tmp/Qwen2-1.5B-Instruct
  model_type: QWEN2

resume_from_checkpoint: False

# Dataset and Sampler
dataset:
  _component_: torchtune.datasets.alpaca_cleaned_dataset
  packed: False  # True increases speed
seed: null
shuffle: True
batch_size: 8

# Optimizer and Scheduler
optimizer:
  _component_: torch.optim.AdamW
  weight_decay: 0.01
  lr: 3e-4
lr_scheduler:
  _component_: torchtune.training.lr_schedulers.get_cosine_schedule_with_warmup
  num_warmup_steps: 100

loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss

kd_loss:
  _component_: torchtune.modules.loss.ForwardKLWithChunkedOutputLoss
kd_ratio: 0.5

# Training
epochs: 1
max_steps_per_epoch: null
gradient_accumulation_steps: 8  # Use to increase virtual batch size
compile: False  # pytorch compile, set to true for better perf/memory

# Logging
output_dir: /tmp/qwen_kd
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: ${output_dir}
log_every_n_steps: 1
log_peak_memory_stats: True

# Environment
device: cuda
dtype: bf16
enable_activation_checkpointing: False  # True reduces memory

# 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