7B_full.yaml

# Config for multi-device full finetuning in full_finetune_distributed.py
# using a Qwen2 7B model
#
# This config assumes that you've run the following command before launching
# this run:
#   tune download Qwen/Qwen2-7B-Instruct --output-dir /tmp/Qwen2-7B-Instruct
#
# To launch on 4 devices, run the following command from root:
#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen2/7B_full
#
# You can add specific overrides through the command line. For example
# to override the checkpointer directory while launching training
# you can run:
#   tune run --nnodes 1 --nproc_per_node 4 full_finetune_distributed --config qwen2/7B_full checkpointer.checkpoint_dir=<YOUR_CHECKPOINT_DIR>
#
# This config works best when the model is being fine-tuned on 2+ GPUs.
# Single device full finetuning requires more memory optimizations. It's
# best to use 7B_full.yaml for those cases

output_dir: /tmp/torchtune/qwen2_7B/full # /tmp may be deleted by your system. Change it to your preference.

# Tokenizer
tokenizer:
  _component_: torchtune.models.qwen2.qwen2_tokenizer
  path: /tmp/Qwen2-7B-Instruct/vocab.json
  merges_file: /tmp/Qwen2-7B-Instruct/merges.txt
  max_seq_len: null

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

# Model Arguments
model:
  _component_: torchtune.models.qwen2.qwen2_7b

checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: /tmp/Qwen2-7B-Instruct
  checkpoint_files: [
    model-00001-of-00004.safetensors,
    model-00002-of-00004.safetensors,
    model-00003-of-00004.safetensors,
    model-00004-of-00004.safetensors
  ]
  recipe_checkpoint: null
  output_dir: ${output_dir}
  model_type: QWEN2
resume_from_checkpoint: False

# Fine-tuning arguments
batch_size: 2
epochs: 1
optimizer:
  _component_: torch.optim.AdamW
  fused: True
  lr: 5e-6
loss:
  _component_: torchtune.modules.loss.CEWithChunkedOutputLoss
max_steps_per_epoch: null
gradient_accumulation_steps: 8  # Use to increase effective batch size
compile: False  # torch.compile the model + loss, True increases speed + decreases memory
optimizer_in_bwd: False  # True saves memory. Requires gradient_accumulation_steps=1

# 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
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