fix: remove fire ported from Hari's PR #303

[HarikrishnanBalagopal]

Description of the change

Remove the 2nd CLI parser (fire) because it's not necessary and throws errors when training with LoRA.

Related issue number

Fixes #47

How to verify the PR

make test

Was the PR tested

• I have added >=1 unit test(s) for every new method I have added.
• I have ensured all unit tests pass

[HarikrishnanBalagopal]

The errors are completely unrelated to the changes in the PR.

https://github.com/foundation-model-stack/fms-hf-tuning/actions/runs/10637170607/job/29490507719?pr=324#step:4:1329

E               ValueError: You passed `packing=False` to the SFTTrainer/SFTConfig, but you didn't pass a `dataset_text_field` or `formatting_func` argument.

Seems to be coming from

fms-hf-tuning/tests/test_sft_trainer.py

Lines 509 to 520 in 6cfb2ff

	# below args not needed for pretokenized data
	data_formatting_args.data_formatter_template = None
	data_formatting_args.dataset_text_field = None
	data_formatting_args.response_template = None
	# update the training data path to tokenized data
	data_formatting_args.training_data_path = dataset_path
	train_args = copy.deepcopy(TRAIN_ARGS)
	train_args.output_dir = tempdir
	sft_trainer.train(MODEL_ARGS, data_formatting_args, train_args)

https://github.com/huggingface/trl/blob/d57e4b726561e5ae58fdc335f34029052944a4a3/trl/trainer/sft_trainer.py#L351

Same error in other PR builds as well https://github.com/foundation-model-stack/fms-hf-tuning/actions/runs/10637170607/job/29490507719?pr=324#step:4:1329

[willmj]

LGTM!

[ashokponkumar]

@anhuong can we merge this when you are comfortable with it?

[anhuong]

@HarikrishnanBalagopal what testing did you do for this change? We would want to ensure that with this change users can still run training from CLI, with JSON config, and from importing python module. Unit tests cover the python module and JSON config use cases. Thus if you test the CLI piece, then this should be sufficient. Including an example of running with CLI and with target_modules and not getting the error would be great.

[willmj]

Implementing these changes on my machine:

Running LoRA:

% python tuning/sft_trainer.py \
--model_name_or_path Maykeye/TinyLLama-v0 \
--training_data_path tests/data/twitter_complaints_small.json \
--output_dir outputs/lora-tuning \
--num_train_epochs 5 \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--learning_rate 1e-5 \
--response_template "\n### Label:" \
--dataset_text_field "output" \
--use_flash_attn false \
--torch_dtype "float32" \
--peft_method "lora" \
--r 8 \
--lora_dropout 0.05 \
--lora_alpha 16  \
--target_modules "v_proj" "c_proj"

Log:

{'loss': 7.9407, 'grad_norm': 4.002565860748291, 'learning_rate': 8.000000000000001e-06, 'epoch': 1.0}
 20%|█████████                                    | 1/5 [00:00<00:02,  1.74it/s]
{'loss': 3.9542, 'grad_norm': 2.8579704761505127, 'learning_rate': 4.000000000000001e-06, 'epoch': 2.0}
 60%|███████████████████████████                  | 3/5 [00:01<00:00,  3.46it/s]
 80%|████████████████████████████████████         | 4/5 [00:01<00:00,  2.96it/s]
{'loss': 3.8709, 'grad_norm': 1.6323113441467285, 'learning_rate': 0.0, 'epoch': 3.0}
100%|█████████████████████████████████████████████| 5/5 [00:01<00:00,  2.96it/s]
{'train_runtime': 1.7348, 'train_samples_per_second': 28.822, 'train_steps_per_second': 2.882, 'train_loss': 4.7181743621826175, 'epoch': 3.0}
100%|█████████████████████████████████████████████| 5/5 [00:01<00:00,  2.88it/s]

Fine Tuning

% python tuning/sft_trainer.py \
--model_name_or_path Maykeye/TinyLLama-v0 \
--training_data_path tests/data/twitter_complaints_small.jsonl \
--output_dir outputs/full-tuning \
--num_train_epochs 5 \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--learning_rate 1e-5 \
--response_template "\n### Label:" \
--dataset_text_field "output" \
--use_flash_attn false \
--torch_dtype "float32"

Log:

{'loss': 7.9407, 'grad_norm': 23.814661026000977, 'learning_rate': 8.000000000000001e-06, 'epoch': 1.0}
{'loss': 3.8985, 'grad_norm': 17.079011917114258, 'learning_rate': 4.000000000000001e-06, 'epoch': 2.0}
{'loss': 3.7799, 'grad_norm': 9.279784202575684, 'learning_rate': 0.0, 'epoch': 3.0}
{'train_runtime': 1.6872, 'train_samples_per_second': 29.635, 'train_steps_per_second': 2.963, 'train_loss': 4.659521102905273, 'epoch': 3.0}
100%|█████████████████████████████████████████████| 5/5 [00:01<00:00,  2.96it/s]

So it seems to run with no problem from CLI

[anhuong]

Thanks for running the test Will!