Paper | Model | Playground | VSCode | Chat

💫 StarCoder is a language model (LM) trained on source code and natural language text. Its training data incorporates more that 80 different programming languages as well as text extracted from GitHub issues and commits and from notebooks. This repository showcases how we get an overview of this LM's capabilities.

• May 9, 2023: We've fine-tuned StarCoder to act as a helpful coding assistant 💬! Check out the chat/ directory for the training code and play with the model here.

Before you can use the model go to hf.co/bigcode/starcoder and accept the agreement. And make sure you are logged into the Hugging Face hub with:

huggingface-cli login

1. Quickstart
   • Installation
   • Code generation with StarCoder
   • Text-generation-inference code
2. Fine-tuning
   • Step by step installation with conda
   • Datasets
     • Stack Exchange
   • Merging PEFT adapter layers
3. Evaluation
4. Inference hardware requirements

StarCoder was trained on GitHub code, thus it can be used to perform code generation. More precisely, the model can complete the implementation of a function or infer the following characters in a line of code. This can be done with the help of the 🤗's transformers library.

First, we have to install all the libraries listed in requirements.txt

pip install -r requirements.txt

The code generation pipeline is as follows

from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigcode/starcoder"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# to save memory consider using fp16 or bf16 by specifying torch_dtype=torch.float16 for example
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
outputs = model.generate(inputs)
# clean_up_tokenization_spaces=False prevents a tokenizer edge case which can result in spaces being removed around punctuation
print(tokenizer.decode(outputs[0], clean_up_tokenization_spaces=False))

or

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
checkpoint = "bigcode/starcoder"

model = AutoModelForCausalLM.from_pretrained(checkpoint)
tokenizer = AutoTokenizer.from_pretrained(checkpoint)

pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, device=0)
print( pipe("def hello():") )

For hardware requirements, check the section Inference hardware requirements.

docker run -p 8080:80 -v $PWD/data:/data -e HUGGING_FACE_HUB_TOKEN=<YOUR BIGCODE ENABLED TOKEN> -d  ghcr.io/huggingface/text-generation-inference:latest --model-id bigcode/starcoder --max-total-tokens 8192

For more details, see here.

Here, we showcase how we can fine-tune this LM on a specific downstream task.

Create a new conda environment and activate it

conda create -n env
conda activate env

Install the pytorch version compatible with your version of cuda here, for example the following command works with cuda 11.6

conda install pytorch==1.13.1 torchvision==0.14.1 torchaudio==0.13.1 pytorch-cuda=11.6 -c pytorch -c nvidia

Install transformers and peft

conda install -c huggingface transformers 
pip install git+https://github.com/huggingface/peft.git

Note that you can install the latest stable version of transformers by using

pip install git+https://github.com/huggingface/transformers

Install datasets, accelerate and huggingface_hub

conda install -c huggingface -c conda-forge datasets
conda install -c conda-forge accelerate
conda install -c conda-forge huggingface_hub

Finally, install bitsandbytes and wandb

pip install bitsandbytes
pip install wandb

To get the full list of arguments with descriptions you can run the following command on any script:

python scripts/some_script.py --help

Before you run any of the scripts make sure you are logged in and can push to the hub:

huggingface-cli login

Make sure you are logged in wandb:

wandb login

Now that everything is done, you can clone the repository and get into the corresponding directory.

💫 StarCoder can be fine-tuned to achieve multiple downstream tasks. Our interest here is to fine-tune StarCoder in order to make it follow instructions. Instruction fine-tuning has gained a lot of attention recently as it proposes a simple framework that teaches language models to align their outputs with human needs. That procedure requires the availability of quality instruction datasets, which contain multiple instruction - answer pairs. Unfortunately such datasets are not ubiquitous but thanks to Hugging Face 🤗's datasets library we can have access to some good proxies. To fine-tune cheaply and efficiently, we use Hugging Face 🤗's PEFT as well as Tim Dettmers' bitsandbytes.

Stack Exchange is a well-known network of Q&A websites on topics in diverse fields. It is a place where a user can ask a question and obtain answers from other users. Those answers are scored and ranked based on their quality. Stack exchange instruction is a dataset that was obtained by scrapping the site in order to build a collection of Q&A pairs. A language model can then be fine-tuned on that dataset to make it elicit strong and diverse question-answering skills.

To execute the fine-tuning script run the following command:

python finetune/finetune.py \
  --model_path="bigcode/starcoder"\
  --dataset_name="ArmelR/stack-exchange-instruction"\
  --subset="data/finetune"\
  --split="train"\
  --size_valid_set 10000\
  --streaming\
  --seq_length 2048\
  --max_steps 1000\
  --batch_size 1\
  --input_column_name="question"\
  --output_column_name="response"\ 
  --gradient_accumulation_steps 16\
  --learning_rate 1e-4\
  --lr_scheduler_type="cosine"\
  --num_warmup_steps 100\
  --weight_decay 0.05\
  --output_dir="./checkpoints" \

The size of the SE dataset is better manageable when using streaming. We also have to precise the split of the dataset that is used. For more details, check the dataset's page on 🤗. Similarly we can modify the command to account for the availability of GPUs

python -m torch.distributed.launch \
  --nproc_per_node number_of_gpus finetune/finetune.py \
  --model_path="bigcode/starcoder"\
  --dataset_name="ArmelR/stack-exchange-instruction"\
  --subset="data/finetune"\
  --split="train"\
  --size_valid_set 10000\
  --streaming \
  --seq_length 2048\
  --max_steps 1000\
  --batch_size 1\
  --input_column_name="question"\
  --output_column_name="response"\ 
  --gradient_accumulation_steps 16\
  --learning_rate 1e-4\
  --lr_scheduler_type="cosine"\
  --num_warmup_steps 100\
  --weight_decay 0.05\
  --output_dir="./checkpoints" \

If you train a model with PEFT, you'll need to merge the adapter layers with the base model if you want to run inference / evaluation. To do so, run:

python finetune/merge_peft_adapters.py --base_model_name_or_path model_to_merge --peft_model_path model_checkpoint

# Push merged model to the Hub
python finetune/merge_peft_adapters.py --base_model_name_or_path model_to_merge --peft_model_path model_checkpoint --push_to_hub

For example

python finetune/merge_peft_adapters.py --model_name_or_path bigcode/starcoder --peft_model_path checkpoints/checkpoint-1000 --push_to_hub

To evaluate StarCoder and its derivatives, you can use the BigCode-Evaluation-Harness for evaluating Code LLMs.

In FP32 the model requires more than 60GB of RAM, you can load it in FP16 or BF16 in ~30GB, or in 8bit under 20GB of RAM with

# make sure you have accelerate and bitsandbytes installed
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bigcode/starcoder")
# for fp16 replace with  `load_in_8bit=True` with   `torch_dtype=torch.float16`
model = AutoModelForCausalLM.from_pretrained("bigcode/starcoder", device_map="auto", load_in_8bit=True)
print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")

Memory footprint: 15939.61 MB

You can also try starcoder.cpp, a C++ implementation with ggml library.