Integrate with torchtune¶

torchtune is a PyTorch-native library for easily authoring, fine-tuning and experimenting with LLMs.

Instrument your runs with Comet to start managing experiments, create dataset versions and track hyperparameters for faster and easier reproducibility and collaboration.

Comet SDK	Minimum SDK version	Minimum torchtune version
Python-SDK	3.45.0	master

Start logging¶

Enable the Comet Logger in your recipe's config:

metric_logger:
  _component_: torchtune.training.metric_logging.CometLogger
  project: <your-project-name>
  experiment_name: <my-experiment-name>

Tip

Find a full list of torchtune recipe configs here.

Log automatically¶

When using the CometMLLogger, Comet automatically logs the following items, by default, with no additional configuration:

Training metrics like loss, and tokens_per_second_per_gpu.
All hyperparameters like lora_rank, lora_alpha and anything else included in the config file.

End-to-end example¶

The following is a basic example of using Comet with torchtune using Mistral7B.

If you can't wait, check out the results of this example torchtune experiment for a preview of what's to come.

Clone the repo¶

git clone https://github.com/pytorch/torchtune/

Install dependencies¶

python -m pip install ".torchtune[dev]" comet_ml

Log-in to Comet¶

comet_ml login

Download the model¶

tune download mistralai/Mistral-7B-v0.1 --output-dir ./Mistral-7B-v0.1/

Tip

Set your environment variable HF_TOKEN or pass in --hf-token to the command in order to validate your access. You can find your token at https://huggingface.co/settings/tokens. For more information on how to set your Hugging Face token, see here.

Write the torchtune config with Comet¶

Write the following config file to mistral_comet_lora.yaml:

tokenizer:
  _component_: torchtune.models.mistral.mistral_tokenizer
  path: ./Mistral-7B-v0.1/tokenizer.model

# Dataset
dataset:
  _component_: torchtune.datasets.alpaca_dataset
  train_on_input: True
seed: null
shuffle: True

# Model Arguments
model:
  _component_: torchtune.models.mistral.lora_mistral_7b
  lora_attn_modules: ["q_proj", "k_proj", "v_proj"]
  apply_lora_to_mlp: True
  apply_lora_to_output: True
  lora_rank: 64
  lora_alpha: 16

checkpointer:
  _component_: torchtune.training.FullModelHFCheckpointer
  checkpoint_dir: ./Mistral-7B-v0.1
  checkpoint_files:
    [pytorch_model-00001-of-00002.bin, pytorch_model-00002-of-00002.bin]
  recipe_checkpoint: null
  output_dir: ./Mistral-7B-v0.1
  model_type: MISTRAL
resume_from_checkpoint: False

optimizer:
  _component_: torch.optim.AdamW
  lr: 2e-5

lr_scheduler:
  _component_: torchtune.modules.get_cosine_schedule_with_warmup
  num_warmup_steps: 100

loss:
  _component_: torch.nn.CrossEntropyLoss

# Fine-tuning arguments
batch_size: 4
epochs: 1
max_steps_per_epoch: 100
gradient_accumulation_steps: 2
compile: False

# Training env
device: cuda

# Memory management
enable_activation_checkpointing: True

# Reduced precision
dtype: bf16
############################### Enable Comet ###################################
################################################################################
# Logging
# enable logging to the built-in CometLogger
metric_logger:
  _component_: torchtune.training.metric_logging.CometLogger
  # the Comet project to log to
  project: comet-examples-torchtune-mistral7b
  experiment_name: mistral7b-alpaca-cleaned
################################################################################
################################################################################
output_dir: ./Mistral-7B-v0.1
log_peak_memory_stats: True

# Profiler (disabled)
profiler:
  _component_: torchtune.training.setup_torch_profiler
  enabled: False

Run the example with a single GPU¶

tune run lora_finetune_single_device --config mistral_comet_lora.yaml

Try it out!¶

Try out an example of using Comet with torchtune and Mistral7B in this Colab notebook.

Sep. 17, 2024