1.5B_full_single_device.yaml

# Config for single device full finetuning in full_finetune_single_device.py
# using a Qwen2.5 1.5B
#
# This config assumes that you've run the following command before launching:
#   tune download Qwen/Qwen2.5-1.5B-Instruct
#
# The default config uses an optimizer from bitsandbytes. If you do not have it installed,
# you can install it with:
#   pip install bitsandbytes
#
# To launch on a single device, run the following command from root:
#   tune run full_finetune_single_device --config qwen2_5/1.5B_full_single_device
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training:
#   tune run full_finetune_single_device --config qwen2_5/1.5B_full_single_device checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works only for training on single device.

# Model arguments
model:
  _component_: torchtune.models.qwen2_5.qwen2_5_1_5b_instruct

# Tokenizer
tokenizer:
  _component_: torchtune.models.qwen2_5.qwen2_5_tokenizer
  path: /tmp/Qwen2.5-1.5B-Instruct/vocab.json
  merges_file: /tmp/Qwen2.5-1.5B-Instruct/merges.txt
  max_seq_len: null

# Checkpointer
checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /tmp/Qwen2.5-1.5B-Instruct
  checkpoint_files: [model.safetensors]
  recipe_checkpoint: null
  output_dir: /tmp/Qwen2.5-1.5B-Instruct-finetune
  model_type: QWEN2
resume_from_checkpoint: False

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

# Fine-tuning arguments
epochs: 1
max_steps_per_epoch: null
batch_size: 4
gradient_accumulation_steps: 1  # Use to increase virtual batch size
optimizer:
  _component_: bitsandbytes.optim.PagedAdamW
  lr: 2e-5
optimizer_in_bwd: True  # True saves memory. Requires gradient_accumulation_steps=1
loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss

# Training env
device: cuda

# Memory management / performance
enable_activation_checkpointing: True  # True reduces memory
enable_activation_offloading: False  # True reduces memory
dtype: bf16
compile: False  # torch.compile the model + loss, True increases speed + decreases memory

# Logging
output_dir: /tmp/Qwen2.5-1.5B-Instruct-finetune
metric_logger:
  _component_: torchtune.training.metric_logging.DiskLogger
  log_dir: ${output_dir}/logs
log_every_n_steps: 1
log_peak_memory_stats: True

# 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