StarCoder 2

[🤗 Models & Datasets] | [Paper]

StarCoder2 is a family of code generation models (3B, 7B, and 15B), trained on 600+ programming languages from The Stack v2 and some natural language text such as Wikipedia, Arxiv, and GitHub issues. The models use Grouped Query Attention, a context window of 16,384 tokens, with sliding window attention of 4,096 tokens. The 3B & 7B models were trained on 3+ trillion tokens, while the 15B was trained on 4+ trillion tokens. For more details check out the paper.

Table of Contents

1. Quickstart
   • Installation
   • Model usage and memory footprint
   • Text-generation-inference code
2. Fine-tuning
   • Setup
   • Training
3. Evaluation

Quickstart

StarCoder2 models are intended for code completion, they are not instruction models and commands like "Write a function that computes the square root." do not work well.

Installation

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

pip install -r requirements.txt
# export your HF token, found here: https://huggingface.co/settings/account
export HF_TOKEN=xxx

Model usage and memory footprint

Here are some examples to load the model and generate code, with the memory footprint of the largest model, StarCoder2-15B. Ensure you've installed transformers from source (it should be the case if you used requirements.txt)

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

Running the model on CPU/GPU/multi GPU

• Using full precision

# pip install git+https://github.com/huggingface/transformers.git # TODO: merge PR to main
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigcode/starcoder2-15b"
device = "cuda" # for GPU usage or "cpu" for CPU usage

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
# to use Multiple GPUs do `model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto")`
model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

• Using torch.bfloat16

# pip install accelerate
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

checkpoint = "bigcode/starcoder2-15b"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)

# for fp16 use `torch_dtype=torch.float16` instead
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", torch_dtype=torch.bfloat16)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

>>> print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")
Memory footprint: 32251.33 MB

Quantized Versions through bitsandbytes

• Using 8-bit precision (int8)

# pip install bitsandbytes accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

# to use 4bit use `load_in_4bit=True` instead
quantization_config = BitsAndBytesConfig(load_in_8bit=True)

checkpoint = "bigcode/starcoder2-15b_16k"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained("bigcode/starcoder2-15b_16k", quantization_config=quantization_config)

inputs = tokenizer.encode("def print_hello_world():", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

>>> print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")
# load_in_8bit
Memory footprint: 16900.18 MB
# load_in_4bit
>>> print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")
Memory footprint: 9224.60 MB

You can also use pipeline for the generation:

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
checkpoint = "bigcode/starcoder2-15b"

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

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

Text-generation-inference:

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/starcoder2-15b --max-total-tokens 8192

For more details, see here.

Fine-tuning

Here, we showcase how you can fine-tune StarCoder2 models. For more fine-tuning resources you can check StarCoder's GitHub repository and SantaCoder-Finetuning.

Setup

Install pytorch see documentation, for example the following command works with cuda 12.1:

conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Install the requirements (this installs transformers from source to support the StarCoder2 architecture):

pip install -r requirements.txt

Before you run any of the scripts make sure you are logged in wandb and HuggingFace Hub to push the checkpoints:

wandb login
huggingface-cli login

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

Training

To fine-tune efficiently with a low cost, we use PEFT library for Low-Rank Adaptation (LoRA) training and bitsandbytes for 4bit quantization. We also use the SFTTrainer from TRL.

For this example, we will fine-tune StarCoder2-3b on the Rust subset of the-stack-smol. This is just for illustration purposes; for a larger and cleaner dataset of Rust code, you can use The Stack dedup.

To launch the training:

accelerate launch finetune.py \
        --model_id "bigcode/starcoder2-3b" \
        --dataset_name "bigcode/the-stack-smol" \
        --subset "data/rust" \
        --dataset_text_field "content" \
        --split "train" \
        --max_seq_length 1024 \
        --max_steps 10000 \
        --micro_batch_size 1 \
        --gradient_accumulation_steps 8 \
        --learning_rate 2e-5 \
        --warmup_steps 20 \
        --num_proc "$(nproc)"

If you want to fine-tune on other text datasets, you need to change dataset_text_field argument to the name of the column containing the code/text you want to train on.

Evaluation

To evaluate StarCoder2 and its derivatives, you can use the BigCode-Evaluation-Harness for evaluating Code LLMs. You can also check the BigCode Leaderboard.