Switches dependencies from txt file to toml file

[jbusche]

Description of the change

This PR creates a pyproject.toml file and refers to the two requirements.txt and flashattn_requirements.txt files for dependencies.

• It creates a pyproject.toml that points to the two requirements.txt and flashattn_requirements.txt files
• Updates the README.md to reflect the new method to install
• Updates the CONTRIBUTING.md to reflect the new method to install
• Updates the build/Dockerfile to reflect the new method to install
• Tightens up the requirements.txt to as small as possible
• In the pyproject.toml, there's opportunity to specify more information about the project, like the name, description, authors, maintainers, public URLS, etc. I've added some people and urls into it, but it's open to discussion to whom else might need to be added or if more changes are needed.

Related issue number

Closes issue #56

How to verify the PR

I've been doing this:

On a Red Hat system (or a Mac system, skipping step 2 below)

1. Download the branch

git clone https://github.com/jbusche/fms-hf-tuning -b jb-dependency-issue56
cd fms-hf-tuning/

2. Install python39 if you don't already have it:

yum install https://dl.fedoraproject.org/pub/epel/epel-release-latest-8.noarch.rpm -y
yum install python39 -y

3. Startup a python virtual environment:

python3.9 -m venv .env
source .env/bin/activate

4. Install with this command:

pip install -e .

5. You can test installing the optional FlashAttention dependencies with this command, though they'll complain you don't have CUDA if you don't have CUDA installed.

pip install -e ".[dev]"
pip install -e ".[flash-attn]"

6. Then you can try running with this command:
   Note: you'll need a twitter_complaints.json file that I can send you.

export CUDA_VISIBLE_DEVICES=0
export WORLD_SIZE=1
export MODEL_PATH="bigscience/bloomz-560m"
export DATA_PATH="twitter_complaints.json"
export OUTPUT_PATH="/tmp/fms-hf-tuning/output"
python -m tuning.sft_trainer  \
--model_name_or_path $MODEL_PATH  \
--training_data_path $DATA_PATH  \
--output_dir $OUTPUT_PATH  \
--num_train_epochs 1  \
--per_device_train_batch_size 4  \
--per_device_eval_batch_size 4  \
--gradient_accumulation_steps 4  \
--evaluation_strategy "no"  \
--save_strategy "epoch"  \
--learning_rate 1e-5  \
--weight_decay 0.0  \
--lr_scheduler_type "cosine"  \
--logging_steps 1  \
--packing false  \
--include_tokens_per_second true \
--response_template "\n### Response:"  \
--dataset_text_field "output" \
--tokenizer_name_or_path "bigscience/tokenizer" \
--use_flash_attn false \
--torch_dtype "float32" \
--peft_method "pt"

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
• Manually tested in the above steps

[jbusche]

Thank you for the help @lchu-ibm and @FilM for the help, I've switched back to the txt files and having the toml file pull up the dependencies from the txt files.

[jbusche]

@gkumbhat and @Ssukriti I think I've accounted for all the suggestions and changes... hoping that this is good to go.

[jbusche]

Does someone have the ability to test the flash-attn pip install -e .[flash-attn]? I don't have a GPU location to install CUDA and test with it.

[tedhtchang]

I think @z103cb should review this.

[jbusche]

Thanks @tedhtchang, yes we need this so we can start with creating the wheel. @gkumbhat and @Ssukriti can you review too?

[gkumbhat]

I noticed that the project.name defined here is fms-hf-tuning, but the package being installed is tuning, probably because of the name of the main folder being tuning. This makes the installation a bit confusing IMO.

Also I tried the pip install -e ".[flash-attn], but couldn't get flash-att get installed as separate dependency

[jbusche]

Oh my, well my rebase went crazy, was trying to fix the requirements file but now I'm getting all the last 3-4 weeks of changes, causing conflicts. Let me see if I can recover. Sorry

[z103cb]

This is looking good to me.

[jbusche]

I'm still testing and making sure I understand/address the changes requested.

[jbusche]

I've made the name change to fms-hf-tuning from just tuning, thank you @gkumbhat for the suggestion.

[gkumbhat]

Looks good to me! Thanks for contributing and making all the changes!

One suggestion for future iteration: We can look into getting rid of setup.py by defining py-modules or packages in tool.setuptools in pyproject.toml