🤗 Hugging Face | 🤖 ModelScope | 📑 Blog ｜ 📖 Documentation
🖥️ Demo | 💬 WeChat (微信) | 🫨 Discord
Visit our Hugging Face or ModelScope organization (click links above), search checkpoints with names starting with CodeQwen1.5-, and you will find all you need! Enjoy!
CodeQwen1.5 is the Code-Specific version of Qwen1.5. It is a transformer-based decoder-only language model pretrained on a large amount of data of codes.
- ✨ Strong code generation capabilities and competitve performance across a series of benchmarks;
- ✨ Supporting long context understanding and generation with the context length of 64K tokens;
- ✨ Supporting 92 coding languages;
['ada', 'agda', 'alloy', 'antlr', 'applescript', 'assembly', 'augeas', 'awk', 'batchfile', 'bluespec', 'c', 'c#', 'c++', 'clojure', 'cmake', 'coffeescript', 'common-lisp', 'css', 'cuda', 'dart', 'dockerfile', 'elixir', 'elm', 'emacs-lisp', 'erlang', 'f#', 'fortran', 'glsl', 'go', 'groovy', 'haskell', 'html', 'idris', 'isabelle', 'java', 'java-server-pages', 'javascript', 'json', 'julia', 'jupyter-notebook', 'kotlin', 'lean', 'literate-agda', 'literate-coffeescript', 'literate-haskell', 'lua', 'makefile', 'maple', 'markdown', 'mathematica', 'matlab', 'objectc++', 'ocaml', 'pascal', 'perl', 'php', 'powershell', 'prolog', 'protocol-buffer', 'python', 'r', 'racket', 'restructuredtext', 'rmarkdown', 'ruby', 'rust', 'sas', 'scala', 'scheme', 'shell', 'smalltalk', 'solidity', 'sparql', 'sql', 'stan', 'standard-ml', 'stata', 'swift', 'systemverilog', 'tcl', 'tcsh', 'tex', 'thrift', 'typescript', 'verilog', 'vhdl', 'visual-basic', 'vue', 'xslt', 'yacc', 'yaml', 'zig']
- ✨ Excellent performance in text-to-SQL, bug fix, etc.
Detailed performance and introduction are shown in this 📑 blog.
python>=3.9transformers>=4.37.0for Qwen1.5 dense models.
Warning
You can install the required packages with the following command:
pip install -r requirements.txtImportant
CodeQwen1.5-7B-Chat is a instruction model for chatting;
CodeQwen1.5-7B is a base model typically used for completion, serving as a better starting point for fine-tuning.
You can just write several lines of code with transformers to chat with CodeQwen1.5-7B-Chat. Essentially, we build the tokenizer and the model with from_pretrained method, and we use generate method to perform chatting with the help of chat template provided by the tokenizer. Below is an example of how to chat with CodeQwen1.5-7B-Chat:
from transformers import AutoTokenizer, AutoModelForCausalLM
device = "cuda" # the device to load the model onto
# Now you do not need to add "trust_remote_code=True"
tokenizer = AutoTokenizer.from_pretrained("Qwen/CodeQwen1.5-7B-Chat")
model = AutoModelForCausalLM.from_pretrained("Qwen/CodeQwen1.5-7B-Chat", device_map="auto").eval()
# tokenize the input into tokens
# Instead of using model.chat(), we directly use model.generate()
# But you need to use tokenizer.apply_chat_template() to format your inputs as shown below
prompt = "write a quick sort algorithm."
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
# Directly use generate() and tokenizer.decode() to get the output.
# Use `max_new_tokens` to control the maximum output length.
generated_ids = model.generate(
model_inputs.input_ids,
max_new_tokens=2048
)
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]The apply_chat_template() function is used to convert the messages into a format that the model can understand.
The add_generation_prompt argument is used to add a generation prompt, which refers to <|im_start|>assistant\n to the input. Notably, we apply ChatML template for chat models following our previous practice.
The max_new_tokens argument is used to set the maximum length of the response. The tokenizer.batch_decode() function is used to decode the response. In terms of the input, the above messages is an example to show how to format your dialog history and system prompt.
The model completes the code snipplets according to the given prompts, without any additional formatting, which is usually termed as code completion in the code generation tasks.
Essentially, we build the tokenizer and the model with from_pretrained method, and we use generate method to perform code completion. Below is an example on how to chat with CodeQwen1.5-base:
from transformers import AutoTokenizer, AutoModelForCausalLM
device = "cuda" # the device to load the model onto
# Now you do not need to add "trust_remote_code=True"
TOKENIZER = AutoTokenizer.from_pretrained("Qwen/CodeQwen1.5-7B")
MODEL = AutoModelForCausalLM.from_pretrained("Qwen/CodeQwen1.5-7B", device_map="auto").eval()
# tokenize the input into tokens
input_text = "#write a quick sort algorithm"
model_inputs = TOKENIZER([input_text], return_tensors="pt").to(device)
# Use `max_new_tokens` to control the maximum output length.
generated_ids = MODEL.generate(model_inputs.input_ids, max_new_tokens=512, do_sample=False)[0]
# The generated_ids include prompt_ids, so we only need to decode the tokens after prompt_ids.
output_text = TOKENIZER.decode(generated_ids[len(model_inputs.input_ids[0]):], skip_special_tokens=True)
print(f"Prompt: {input_text}\n\nGenerated text: {output_text}")The max_new_tokens argument is used to set the maximum length of the response.
The input_text could be any text that you would like model to continue with.
The code insertion task, also referred to as the "fill-in-the-middle" challenge, requires the insertion of code segments in a manner that bridges the gaps within a given code context.
For an approach aligned with best practices, we recommend adhering to the formatting guidelines outlined in the paper "Efficient Training of Language Models to Fill in the Middle"[arxiv]. This involves the use of three specialized tokens<fim_prefix>, <fim_suffix>, and <fim_middle> to denote the respective segments of the code structure.
The prompt should be structured as follows:
prompt = '<fim_prefix>' + prefix_code + '<fim_suffix>' + suffix_code + '<fim_middle>'Following the approach mentioned, an example would be structured in this manner:
from transformers import AutoTokenizer, AutoModelForCausalLM
# load model
device = "cuda" # the device to load the model onto
TOKENIZER = AutoTokenizer.from_pretrained("Qwen/CodeQwen1.5-7B")
MODEL = AutoModelForCausalLM.from_pretrained("Qwen/CodeQwen1.5-7B", device_map="auto").eval()
input_text = """<fim_prefix>def quicksort(arr):
if len(arr) <= 1:
return arr
pivot = arr[len(arr) // 2]
<fim_suffix>
middle = [x for x in arr if x == pivot]
right = [x for x in arr if x > pivot]
return quicksort(left) + middle + quicksort(right)<fim_middle>"""
model_inputs = TOKENIZER([input_text], return_tensors="pt").to(device)
# Use `max_new_tokens` to control the maximum output length.
generated_ids = MODEL.generate(model_inputs.input_ids, max_new_tokens=512, do_sample=False)[0]
# The generated_ids include prompt_ids, we only need to decode the tokens after prompt_ids.
output_text = TOKENIZER.decode(generated_ids[len(model_inputs.input_ids[0]):], skip_special_tokens=True)
print(f"Prompt: {input_text}\n\nGenerated text: {output_text}")The repository level code completion task involves feeding the model the content of multiple files from the same repository. This enables the model to understand the interrelationships between different calls within these files, thereby facilitating the completion of code content.
We recommend using the two special tokens <reponame> and <file_sep> to indicate the repository structure.
For example, assuming the repository name is stored in repo_name, and it contains files with their respective paths and contents listed as [(file_path1, file_content1), (file_path2, file_content2)], the format of the final input prompt would be as follows:
input_text = f'''<reponame>{repo_name}
<file_sep>{file_path1}
{file_content1}
<file_sep>{file_path2}
{file_content2}'''👇🏻 Below is a complete example of a repository level code completion task: :: click to expand ::
from transformers import AutoTokenizer, AutoModelForCausalLM
device = "cuda" # the device to load the model onto
# Now you do not need to add "trust_remote_code=True"
TOKENIZER = AutoTokenizer.from_pretrained("Qwen/CodeQwen1.5-7B")
MODEL = AutoModelForCausalLM.from_pretrained("Qwen/CodeQwen1.5-7B", device_map="auto").eval()
# tokenize the input into tokens
input_text = """<reponame>library-system
<file_sep>library.py
class Book:
def __init__(self, title, author, isbn, copies):
self.title = title
self.author = author
self.isbn = isbn
self.copies = copies
def __str__(self):
return f"Title: {self.title}, Author: {self.author}, ISBN: {self.isbn}, Copies: {self.copies}"
class Library:
def __init__(self):
self.books = []
def add_book(self, title, author, isbn, copies):
book = Book(title, author, isbn, copies)
self.books.append(book)
def find_book(self, isbn):
for book in self.books:
if book.isbn == isbn:
return book
return None
def list_books(self):
return self.books
<file_sep>student.py
class Student:
def __init__(self, name, id):
self.name = name
self.id = id
self.borrowed_books = []
def borrow_book(self, book, library):
if book and book.copies > 0:
self.borrowed_books.append(book)
book.copies -= 1
return True
return False
def return_book(self, book, library):
if book in self.borrowed_books:
self.borrowed_books.remove(book)
book.copies += 1
return True
return False
<file_sep>main.py
from library import Library
from student import Student
def main():
# Set up the library with some books
library = Library()
library.add_book("The Great Gatsby", "F. Scott Fitzgerald", "1234567890", 3)
library.add_book("To Kill a Mockingbird", "Harper Lee", "1234567891", 2)
# Set up a student
student = Student("Alice", "S1")
# Student borrows a book
"""
model_inputs = TOKENIZER([input_text], return_tensors="pt").to(device)
# Use `max_new_tokens` to control the maximum output length.
generated_ids = MODEL.generate(model_inputs.input_ids, max_new_tokens=1024, do_sample=False)[0]
# The generated_ids include prompt_ids, so we only need to decode the tokens after prompt_ids.
output_text = TOKENIZER.decode(generated_ids[len(model_inputs.input_ids[0]):], skip_special_tokens=True)
print(f"Prompt: \n{input_text}\n\nGenerated text: \n{output_text}")The expected output as following:
Generated text:
book = library.find_book("1234567890")
if student.borrow_book(book, library):
print(f"{student.name} borrowed {book.title}")
else:
print(f"{student.name} could not borrow {book.title}")
# Student returns a book
if student.return_book(book, library):
print(f"{student.name} returned {book.title}")
else:
print(f"{student.name} could not return {book.title}")
# List all books in the library
print("All books in the library:")
for book in library.list_books():
print(book)
if __name__ == "__main__":
main()
As a family member of Qwen1.5, CodeQwen1.5 are supported by vLLM. The detail tutorial could be found in Qwen tutorial. Here, we give you an simple example of offline batched inference in vLLM.
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams
# Initialize the tokenizer
tokenizer = AutoTokenizer.from_pretrained("Qwen/CodeQwen1.5-7B")
# Pass the default decoding hyperparameters of Qwen1.5-7B-Chat
# max_tokens is for the maximum length for generation.
sampling_params = SamplingParams(temperature=0.7, top_p=0.8, repetition_penalty=1.05, max_tokens=1024)
# Input the model name or path. Can be GPTQ or AWQ models.
llm = LLM(model="Qwen/CodeQwen1.5-7B")
# Prepare your prompts
prompt = "#write a quick sort algorithm.\ndef quick_sort("
# generate outputs
outputs = llm.generate([prompt], sampling_params)
# Print the outputs.
for output in outputs:
prompt = output.prompt
generated_text = output.outputs[0].text
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")To scale up your serving throughputs, distributed serving helps you by leveraging more GPU devices.
When using ultra-long sequences for inference, it might cause insufficient GPU memory. Here, we demonstrate how to run CodeQwen1.5-7B with tensor parallelism just by passing in the argument tensor_parallel_size.
llm = LLM(model="Qwen/CodeQwen1.5-7B", tensor_parallel_size=4)We recommend using EvalPlus to evaluate the effectiveness of HumaneEval and MBPP. Here is our evaluation script.
| Model | Size |
HumanEval
0-shot
|
HumanEval+
0-shot
|
MBPP
0-shot
|
MBPP+
0-shot
|
MBPP
3-shot
|
| Base Model | ||||||
| CodeLlama-Base | 7B | 33.5 | 25.6 | 52.1 | 41.6 | 38.6 |
| StarCoder2 | 7B | 35.4 | 29.9 | 54.4 | 45.6 | 51.0 |
| DeepSeek-Coder-Base | 6.7B | 47.6 | 39.6 | 70.2 | 56.6 | 60.6 |
| CodeQwen1.5 | 7B | 51.8 | 45.7 | 72.2 | 60.2 | 61.8 |
| Chat Model | ||||||
| GPT-3.5-Turbo | - | 76.8 | 70.7 | 82.5 | 69.7 | 70.8 |
| GPT-4-Turbo (Nov 2023) | - | 85.4 | 81.7 | 83.5 | 70.7 | 80.0 |
| DeepSeek-Coder-Instruct | 6.7B | 73.8 | 70.1 | 73.2 | 63.4 | 65.4 |
| CodeQwen1.5-Chat | 7B | 83.5 | 78.7 | 77.7 | 67.2 | 70.6 |
LiveCodeBench provides holistic and contamination-free evaluation of coding capabilities of LLMs. Particularly, LiveCodeBench continuously collects new problems over time from contests across three competition platforms -- LeetCode, AtCoder, and CodeForces. Here is our evaluation script.
| Model | Size |
Code Generation
All Time
Pass@1
|
Code Generation
2023/9/1 ~ 2024/4/1
Pass@1
|
| Base Model | |||
| CodeLlama-Base | 7B | 6.5 | 7.6 |
| StarCoder2 | 7B | 11.3 | 12.7 |
| DeepSeek-Coder-Base | 6.7B | 19.1 | 13.7 |
| CodeQwen1.5 | 7B | 21.8 | 19.3 |
| Chat Model | |||
| CodeLlama-Instruct | 7B | 10.6 | 12.4 |
| DeepSeek-Coder-Instruct | 6.7B | 21.6 | 19.2 |
| CodeQwen1.5-Chat | 7B | 25.0 | 23.2 |
MultiPL-E is a popular benchmark for evaluating multiple programming languages. You can find our reproduce process here.
| Model | Size | Python | C++ | Java | PHP | TS | C# | Bash | JS | Avg |
| Base Model | ||||||||||
| CodeLlama-Base | 7B | 31.7 | 29.8 | 34.2 | 23.6 | 36.5 | 36.7 | 12.0 | 29.2 | 29.2 |
| StarCoder2-Base | 7B | 35.3 | 40.9 | 37.3 | 29.2 | 37.7 | 40.5 | 9.4 | 36.0 | 33.3 |
| DeepSeek-Coder-Base | 6.7B | 49.4 | 50.3 | 43.0 | 38.5 | 49.7 | 50.0 | 28.5 | 48.4 | 44.7 |
| CodeQwen1.5 | 7B | 52.4 | 52.2 | 42.4 | 46.6 | 52.2 | 55.7 | 36.7 | 49.7 | 48.5 |
| Chat Model | ||||||||||
| GPT-3.5-Turbo | - | 76.2 | 63.4 | 69.2 | 60.9 | 69.1 | 70.8 | 42.4 | 67.1 | 64.9 |
| GPT-4 | - | 84.1 | 76.4 | 81.6 | 77.2 | 77.4 | 79.1 | 58.2 | 78.0 | 76.5 |
| DeepSeek-Coder-Instruct | 6.7B | 78.6 | 63.4 | 68.4 | 68.9 | 67.2 | 72.8 | 36.7 | 72.7 | 66.1 |
| CodeQwen1.5-Chat | 7B | 83.2 | 71.2 | 70.1 | 73.5 | 75.4 | 75.9 | 41.1 | 78.2 | 71.1 |
We evaluated CodeQwen1.5-7B-Chat on popular text-to-SQL benchmarks Spider and BIRD. Here you can find the prompts we used, sourced from Chang et al and Li et al.
| Model | Size |
Spider
Execution Accuracy
Dev Set
|
Bird
Execution Accuracy
Dev Set
|
| GPT-3.5-Turbo | - | 70.1 | 37.2 |
| GPT-4 | - | 85.3 | 50.7 |
| CodeLlama-Instruct | 7B | 59.5 | 22.4 |
| DeepSeek-Coder-Instruct | 6.7B | 70.1 | 39.4 |
| CodeQwen1.5-Chat | 7B | 77.9 | 42.0 |
If you find our work helpful, feel free to give us a cite.
@article{qwen,
title={Qwen Technical Report},
author={Jinze Bai and Shuai Bai and Yunfei Chu and Zeyu Cui and Kai Dang and Xiaodong Deng and Yang Fan and Wenbin Ge and Yu Han and Fei Huang and Binyuan Hui and Luo Ji and Mei Li and Junyang Lin and Runji Lin and Dayiheng Liu and Gao Liu and Chengqiang Lu and Keming Lu and Jianxin Ma and Rui Men and Xingzhang Ren and Xuancheng Ren and Chuanqi Tan and Sinan Tan and Jianhong Tu and Peng Wang and Shijie Wang and Wei Wang and Shengguang Wu and Benfeng Xu and Jin Xu and An Yang and Hao Yang and Jian Yang and Shusheng Yang and Yang Yao and Bowen Yu and Hongyi Yuan and Zheng Yuan and Jianwei Zhang and Xingxuan Zhang and Yichang Zhang and Zhenru Zhang and Chang Zhou and Jingren Zhou and Xiaohuan Zhou and Tianhang Zhu},
journal={arXiv preprint arXiv:2309.16609},
year={2023}
}If you are interested to leave a message to either our research team or product team, join our Discord or WeChat groups!