Paper Link👁️

DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence

1. Introduction

We present DeepSeek-Coder-V2, an open-source Mixture-of-Experts (MoE) code language model that achieves performance comparable to GPT4-Turbo in code-specific tasks. Specifically, DeepSeek-Coder-V2 is further pre-trained from an intermediate checkpoint of DeepSeek-V2 with additional 6 trillion tokens. Through this continued pre-training, DeepSeek-Coder-V2 substantially enhances the coding and mathematical reasoning capabilities of DeepSeek-V2, while maintaining comparable performance in general language tasks. Compared to DeepSeek-Coder-33B, DeepSeek-Coder-V2 demonstrates significant advancements in various aspects of code-related tasks, as well as reasoning and general capabilities. Additionally, DeepSeek-Coder-V2 expands its support for programming languages from 86 to 338, while extending the context length from 16K to 128K.

In standard benchmark evaluations, DeepSeek-Coder-V2 achieves superior performance compared to closed-source models such as GPT4-Turbo, Claude 3 Opus, and Gemini 1.5 Pro in coding and math benchmarks. The list of supported programming languages can be found here.

2. Model Downloads

We release the DeepSeek-Coder-V2 with 16B and 236B parameters based on the DeepSeekMoE framework, which has actived parameters of only 2.4B and 21B , including base and instruct models, to the public.

Model	#Total Params	#Active Params	Context Length	Download
DeepSeek-Coder-V2-Lite-Base	16B	2.4B	128k	🤗 HuggingFace
DeepSeek-Coder-V2-Lite-Instruct	16B	2.4B	128k	🤗 HuggingFace
DeepSeek-Coder-V2-Base	236B	21B	128k	🤗 HuggingFace
DeepSeek-Coder-V2-Instruct	236B	21B	128k	🤗 HuggingFace

3. Evaluation Results

3.1 Code Generation

	#TP	#AP	HumanEval	MBPP+	LiveCodeBench	USACO
Closed-Source Models
Gemini-1.5-Pro	-	-	83.5	74.6	34.1	4.9
Claude-3-Opus	-	-	84.2	72.0	34.6	7.8
GPT-4-Turbo-1106	-	-	87.8	69.3	37.1	11.1
GPT-4-Turbo-0409	-	-	88.2	72.2	45.7	12.3
GPT-4o-0513	-	-	91.0	73.5	43.4	18.8
Open-Source Models
CodeStral	22B	22B	78.1	68.2	31.0	4.6
DeepSeek-Coder-Instruct	33B	33B	79.3	70.1	22.5	4.2
Llama3-Instruct	70B	70B	81.1	68.8	28.7	3.3
DeepSeek-Coder-V2-Lite-Instruct	16B	2.4B	81.1	68.8	24.3	6.5
DeepSeek-Coder-V2-Instruct	236B	21B	90.2	76.2	43.4	12.1

3.2 Code Completion

Model	#TP	#AP	RepoBench (Python)	RepoBench (Java)	HumanEval FIM
CodeStral	22B	22B	46.1	45.7	83.0
DeepSeek-Coder-Base	7B	7B	36.2	43.3	86.1
DeepSeek-Coder-Base	33B	33B	39.1	44.8	86.4
DeepSeek-Coder-V2-Lite-Base	16B	2.4B	38.9	43.3	86.4

3.3 Code Fixing

	#TP	#AP	Defects4J	SWE-Bench	Aider
Closed-Source Models
Gemini-1.5-Pro	-	-	18.6	19.3	57.1
Claude-3-Opus	-	-	25.5	11.7	68.4
GPT-4-Turbo-1106	-	-	22.8	22.7	65.4
GPT-4-Turbo-0409	-	-	24.3	18.3	63.9
GPT-4o-0513	-	-	26.1	26.7	72.9
Open-Source Models
CodeStral	22B	22B	17.8	2.7	51.1
DeepSeek-Coder-Instruct	33B	33B	11.3	0.0	54.5
Llama3-Instruct	70B	70B	16.2	-	49.2
DeepSeek-Coder-V2-Lite-Instruct	16B	2.4B	9.2	0.0	44.4
DeepSeek-Coder-V2-Instruct	236B	21B	21.0	12.7	73.7

3.4 Mathematical Reasoning

	#TP	#AP	GSM8K	MATH	AIME 2024	Math Odyssey
Closed-Source Models
Gemini-1.5-Pro	-	-	90.8	67.7	2/30	45.0
Claude-3-Opus	-	-	95.0	60.1	2/30	40.6
GPT-4-Turbo-1106	-	-	91.4	64.3	1/30	49.1
GPT-4-Turbo-0409	-	-	93.7	73.4	3/30	46.8
GPT-4o-0513	-	-	95.8	76.6	2/30	53.2
Open-Source Models
Llama3-Instruct	70B	70B	93.0	50.4	1/30	27.9
DeepSeek-Coder-V2-Lite-Instruct	16B	2.4B	86.4	61.8	0/30	44.4
DeepSeek-Coder-V2-Instruct	236B	21B	94.9	75.7	4/30	53.7

3.5 General Natural Language

Benchmark	Domain	DeepSeek-V2-Lite Chat	DeepSeek-Coder-V2-Lite Instruct	DeepSeek-V2 Chat	DeepSeek-Coder-V2 Instruct
BBH	English	48.1	61.2	79.7	83.9
MMLU	English	55.7	60.1	78.1	79.2
ARC-Easy	English	86.1	88.9	98.1	97.4
ARC-Challenge	English	73.4	77.4	92.3	92.8
TriviaQA	English	65.2	59.5	86.7	82.3
NaturalQuestions	English	35.5	30.8	53.4	47.5
AGIEval	English	42.8	28.7	61.4	60
CLUEWSC	Chinese	80.0	76.5	89.9	85.9
C-Eval	Chinese	60.1	61.6	78.0	79.4
CMMLU	Chinese	62.5	62.7	81.6	80.9
Arena-Hard	-	11.4	38.1	41.6	65.0
AlpaceEval 2.0	-	16.9	17.7	38.9	36.9
MT-Bench	-	7.37	7.81	8.97	8.77
Alignbench	-	6.02	6.83	7.91	7.84

3.6 Context Window

Evaluation results on the Needle In A Haystack (NIAH) tests. DeepSeek-Coder-V2 performs well across all context window lengths up to 128K.

4. Chat Website

You can chat with the DeepSeek-Coder-V2 on DeepSeek's official website: coder.deepseek.com

5. API Platform

We also provide OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com, and you can also pay-as-you-go at an unbeatable price.

6. How to run locally

Here, we provide some examples of how to use DeepSeek-Coder-V2-Lite model. If you want to utilize DeepSeek-Coder-V2 in BF16 format for inference, 80GB*8 GPUs are required.

Inference with Huggingface's Transformers

You can directly employ Huggingface's Transformers for model inference.

Code Completion

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
input_text = "#write a quick sort algorithm"
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_length=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Code Insertion

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Base", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
input_text = """<｜fim▁begin｜>def quick_sort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = []
    right = []
<｜fim▁hole｜>
        if arr[i] < pivot:
            left.append(arr[i])
        else:
            right.append(arr[i])
    return quick_sort(left) + [pivot] + quick_sort(right)<｜fim▁end｜>"""
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_length=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True)[len(input_text):])

Chat Completion

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct", trust_remote_code=True, torch_dtype=torch.bfloat16).cuda()
messages=[
    { 'role': 'user', 'content': "write a quick sort algorithm in python."}
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
# tokenizer.eos_token_id is the id of <｜end▁of▁sentence｜> token
outputs = model.generate(inputs, max_new_tokens=512, do_sample=False, top_k=50, top_p=0.95, num_return_sequences=1, eos_token_id=tokenizer.eos_token_id)
print(tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True))

The complete chat template can be found within tokenizer_config.json located in the huggingface model repository.

An example of chat template is as belows:

<｜begin▁of▁sentence｜>User: {user_message_1}

Assistant: {assistant_message_1}<｜end▁of▁sentence｜>User: {user_message_2}

Assistant:

You can also add an optional system message:

<｜begin▁of▁sentence｜>{system_message}

User: {user_message_1}

Assistant: {assistant_message_1}<｜end▁of▁sentence｜>User: {user_message_2}

Assistant:

In the last round of dialogue, note that "Assistant:" has no space after the colon. Adding a space might cause the following issues on the 16B-Lite model:

English questions receiving Chinese responses.
Responses containing garbled text.
Responses repeating excessively.

Older versions of Ollama had this bug (see deepseek-ai#12), but it has been fixed in the latest version.

Inference with vLLM (recommended)

To utilize vLLM for model inference, please merge this Pull Request into your vLLM codebase: vllm-project/vllm#4650.

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 8192, 1
model_name = "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
llm = LLM(model=model_name, tensor_parallel_size=tp_size, max_model_len=max_model_len, trust_remote_code=True, enforce_eager=True)
sampling_params = SamplingParams(temperature=0.3, max_tokens=256, stop_token_ids=[tokenizer.eos_token_id])

messages_list = [
    [{"role": "user", "content": "Who are you?"}],
    [{"role": "user", "content": "write a quick sort algorithm in python."}],
    [{"role": "user", "content": "Write a piece of quicksort code in C++."}],
]

prompt_token_ids = [tokenizer.apply_chat_template(messages, add_generation_prompt=True) for messages in messages_list]

outputs = llm.generate(prompt_token_ids=prompt_token_ids, sampling_params=sampling_params)

generated_text = [output.outputs[0].text for output in outputs]
print(generated_text)

7. License

This code repository is licensed under the MIT License. The use of DeepSeek-Coder-V2 Base/Instruct models is subject to the Model License. DeepSeek-Coder-V2 series (including Base and Instruct) supports commercial use.

8. Citation

@article{zhu2024deepseek,
  title={DeepSeek-Coder-V2: Breaking the Barrier of Closed-Source Models in Code Intelligence},
  author={Zhu, Qihao and Guo, Daya and Shao, Zhihong and Yang, Dejian and Wang, Peiyi and Xu, Runxin and Wu, Y and Li, Yukun and Gao, Huazuo and Ma, Shirong and others},
  journal={arXiv preprint arXiv:2406.11931},
  year={2024}
}

9. Contact

If you have any questions, please raise an issue or contact us at service@deepseek.com.

galachad/DeepSeek-Coder-V2