fix: Remove deprecated push_to_hub_token to resolve warning

[Luka-D]

Description of the change

Running tuning locally produces this warning message:

FutureWarning: `--push_to_hub_token` is deprecated and will be removed in version 5 of 🤗 Transformers. Use `--hub_token` instead.

When I print out the transformer_kwargs dictionary, I can see that both hub_token and push_to_hub_token are in the dictionary. Here is a snippet of it.

.....'hub_token': '<HUB_TOKEN>', 'hub_private_repo': False, 'hub_always_push': False, 'gradient_checkpointing': False, 'gradient_checkpointing_kwargs': None, 'include_inputs_for_metrics': False, 'eval_do_concat_batches': True, 'fp16_backend': 'auto', 'evaluation_strategy': None, 'push_to_hub_model_id': None, 'push_to_hub_organization': None, 'push_to_hub_token': '<PUSH_TO_HUB_TOKEN>',......

I made a temporary fix by simply popping push_to_hub_token out of the dictionary and now the warning is gone.

# Remove deprecated push_to_hub_token
transformer_kwargs.pop("push_to_hub_token", None)

I am wondering if anyone knows where push_to_hub_token is originating from or best way to remove it.

Related issue number

Issue #1205

How to verify the PR

Run tuning locally:

python3 tuning/sft_trainer.py \
--model_name_or_path Maykeye/TinyLLama-v0 \
--training_data_path tests/artifacts/testdata/jsonl/twitter_complaints_small.jsonl \
--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  \
--log_level "error"

[github-actions]

Thanks for making a pull request! 😃
One of the maintainers will review and advise on the next steps.

[Luka-D]

Summary of changes made with 9e567e2

It seems that using .to_dict() method on train_args was returning the hub_token and push_to_hub_token fields, which had a default value like <PUSH_TO_HUB_TOKEN>. Changing it to train_args.__dict__.items() has removed some of those default values.

Because of the change to train_args.__dict__.items(), I needed to change SFTConfig fields function to only return parameters that are accepted by SFTConfig. Previously it was also returning fields like _n_gpu, which is inherited from TrainingArguments class but is not accepted in SFTConfig and will raise an error. Using SFTConfig.__dict__["__match_args__"] will return the same parameters as those listed in the variables window on HF: https://huggingface.co/docs/trl/en/sft_trainer#trl.SFTConfig.

In summary, the changes to transformer_kwargs are:

• hub_token and push_to_hub_token are now both None, instead of <hub_token> and <push_to_hub_token>
• Some of the parameters have changed from strings into Object types. eg. 'lr_scheduler_type': 'linear' has become 'lr_scheduler_type': <SchedulerType.LINEAR: 'linear'>. 'save_strategy': 'epoch' has become 'save_strategy': <IntervalStrategy.EPOCH: 'epoch'> etc.
• I do not know if this change to Object types is the desired result, would like some feedback on this.

All code has passed tox -e py tests.

Please let me know your thoughts, thank you.

[kmehant]

Use of __match_args__ requires Python>=3.10 isn't it? Correct me.

[Luka-D]

Yes, I believe it is Python 3.10 and greater. Although I'm a little confused on whether __match_args__ are auto generated by Python when creating a class or manually set by those writing the class. As it is now, it is a key in the dictionary, so I wonder if the key would still exist in Python 3.9 and lower.

[kmehant]

I guess we should avoid usage of __match_args__ in that case, WDYT?

[Luka-D]

I guess so, which version of Python is the lowest supported by fms-hf-tuning repo?

Also, if we remove __match_args__ do you know another way to get all the input variables for SFTConfig as listed here? Using __dict__.keys() only returns the 12 parameters, but I need all of the variables that SFTConfig inherits from TrainingArguments class. For example, output_dir, do_train, do_eval etc.

[Luka-D]

I made small change with c7d540e. I changed __dict__ to vars(). The functionality is the same for both, vars() accesses the __dict__ method, however using vars() is more readable and more pythonic in my opinion.