# 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](https://huggingface.co/datasets/bigcode/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](https://drive.google.com/file/d/17iGn3c-sYNiLyRSY-A85QOzgzGnGiVI3/view). # Table of Contents 1. [Quickstart](#quickstart) - [Installation](#installation) - [Model usage and memory footprint](#model-usage-and-memory-footprint) - [Text-generation-inference code](#text-generation-inference) 2. [Fine-tuning](#fine-tuning) - [Setup](#setup) - [Training](#training) 3. [Evaluation](#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` ```bash 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`) ```bash pip install git+https://github.com/huggingface/transformers.git ``` ### Running the model on CPU/GPU/multi GPU * _Using full precision_ ```python # 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`_ ```python # 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])) ``` ```bash >>> 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)_ ```python # 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])) ``` ```bash >>> 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: ```python 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: ```bash docker run -p 8080:80 -v $PWD/data:/data -e HUGGING_FACE_HUB_TOKEN= -d ghcr.io/huggingface/text-generation-inference:latest --model-id bigcode/starcoder2-15b --max-total-tokens 8192 ``` For more details, see [here](https://github.com/huggingface/text-generation-inference). # Fine-tuning Here, we showcase how you can fine-tune StarCoder2 models. For more fine-tuning resources you can check [StarCoder's GitHub repository](https://github.com/bigcode-project/starcoder) and [SantaCoder-Finetuning](https://github.com/loubnabnl/santacoder-finetuning). ## Setup Install `pytorch` [see documentation](https://pytorch.org/), for example the following command works with cuda 12.1: ```bash 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): ```bash 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: ```bash 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](https://github.com/huggingface/peft) library for Low-Rank Adaptation (LoRA) training and [bitsandbytes](https://github.com/TimDettmers/bitsandbytes) for 4bit quantization. We also use the `SFTTrainer` from [TRL](https://github.com/huggingface/trl). For this example, we will fine-tune StarCoder2-3b on the `Rust` subset of [the-stack-smol](https://huggingface.co/datasets/bigcode/the-stack-smol). This is just for illustration purposes; for a larger and cleaner dataset of Rust code, you can use [The Stack dedup](https://huggingface.co/datasets/bigcode/the-stack-dedup). To launch the training: ```bash 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](https://github.com/bigcode-project/bigcode-evaluation-harness) for evaluating Code LLMs. You can also check the [BigCode Leaderboard](https://huggingface.co/spaces/bigcode/bigcode-models-leaderboard).