Awesome

DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model

1. Introduction

Today, we’re introducing DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token. Compared with DeepSeek 67B, DeepSeek-V2 achieves stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times.

We pretrained DeepSeek-V2 on a diverse and high-quality corpus comprising 8.1 trillion tokens. This comprehensive pretraining was followed by a process of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unleash the model's capabilities. The evaluation results validate the effectiveness of our approach as DeepSeek-V2 achieves remarkable performance on both standard benchmarks and open-ended generation evaluation.

2. News

• 2024.05.16: We released the DeepSeek-V2-Lite.
• 2024.05.06: We released the DeepSeek-V2.

3. Model Downloads

Due to the constraints of HuggingFace, the open-source code currently experiences slower performance than our internal codebase when running on GPUs with Huggingface. To facilitate the efficient execution of our model, we offer a dedicated vllm solution that optimizes performance for running our model effectively.

4. Evaluation Results

Base Model

Standard Benchmark (Models larger than 67B)

Standard Benchmark (Models smaller than 16B)

Context Window

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

Chat Model

Standard Benchmark (Models larger than 67B)

Standard Benchmark (Models smaller than 16B)

English Open Ended Generation Evaluation

We evaluate our model on AlpacaEval 2.0 and MTBench, showing the competitive performance of DeepSeek-V2-Chat-RL on English conversation generation.

Chinese Open Ended Generation Evaluation

Alignbench (https://arxiv.org/abs/2311.18743)

Coding Benchmarks

We evaluate our model on LiveCodeBench (0901-0401), a benchmark designed for live coding challenges. As illustrated, DeepSeek-V2 demonstrates considerable proficiency in LiveCodeBench, achieving a Pass@1 score that surpasses several other sophisticated models. This performance highlights the model's effectiveness in tackling live coding tasks.

5. Model Architecture

DeepSeek-V2 adopts innovative architectures to guarantee economical training and efficient inference：

• For attention, we design MLA (Multi-head Latent Attention), which utilizes low-rank key-value union compression to eliminate the bottleneck of inference-time key-value cache, thus supporting efficient inference.
• For Feed-Forward Networks (FFNs), we adopt DeepSeekMoE architecture, a high-performance MoE architecture that enables training stronger models at lower costs.

6. Chat Website

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

7. API Platform

We also provide OpenAI-Compatible API at DeepSeek Platform: platform.deepseek.com. Sign up for over millions of free tokens. And you can also pay-as-you-go at an unbeatable price.

8. How to run locally

To utilize DeepSeek-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.

Text Completion

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "deepseek-ai/DeepSeek-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
# `device_map` cannot be set to `auto`
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="sequential", torch_dtype=torch.bfloat16, max_memory=max_memory, attn_implementation="eager")
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id

text = "An attention function can be described as mapping a query and a set of key-value pairs to an output, where the query, keys, values, and output are all vectors. The output is"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs.to(model.device), max_new_tokens=100)

result = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(result)

Chat Completion

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "deepseek-ai/DeepSeek-V2-Chat"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
# `max_memory` should be set based on your devices
max_memory = {i: "75GB" for i in range(8)}
# `device_map` cannot be set to `auto`
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True, device_map="sequential", torch_dtype=torch.bfloat16, max_memory=max_memory, attn_implementation="eager")
model.generation_config = GenerationConfig.from_pretrained(model_name)
model.generation_config.pad_token_id = model.generation_config.eos_token_id

messages = [
    {"role": "user", "content": "Write a piece of quicksort code in C++"}
]
input_tensor = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
outputs = model.generate(input_tensor.to(model.device), max_new_tokens=100)

result = tokenizer.decode(outputs[0][input_tensor.shape[1]:], skip_special_tokens=True)
print(result)

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:

Inference with SGLang (recommended)

SGLang currently supports MLA optimizations, FP8 (W8A8), FP8 KV Cache, and Torch Compile, offering the best latency and throughput among open-source frameworks. Here are some example commands to launch an OpenAI API-compatible server:

# BF16, tensor parallelism = 8
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V2-Chat --tp 8 --trust-remote-code

# BF16, w/ torch.compile (The compilation can take several minutes)
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V2-Lite-Chat --trust-remote-code --enable-torch-compile

# FP8, tensor parallelism = 8, FP8 KV cache
python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-V2-Chat --tp 8 --trust-remote-code --quant fp8 --kv-cache-dtype fp8_e5m2

After launching the server, you can query it with OpenAI API

import openai
client = openai.Client(
    base_url="http://127.0.0.1:30000/v1", api_key="EMPTY")

# Chat completion
response = client.chat.completions.create(
    model="default",
    messages=[
        {"role": "system", "content": "You are a helpful AI assistant"},
        {"role": "user", "content": "List 3 countries and their capitals."},
    ],
    temperature=0,
    max_tokens=64,
)
print(response)

Inference with vLLM (recommended)

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

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 8192, 8
model_name = "deepseek-ai/DeepSeek-V2-Chat"
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": "Translate the following content into Chinese directly: DeepSeek-V2 adopts innovative architectures to guarantee economical training and efficient inference."}],
    [{"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)

LangChain Support

Since our API is compatible with OpenAI, you can easily use it in langchain. Here is an example:

from langchain_openai import ChatOpenAI
llm = ChatOpenAI(
    model='deepseek-chat',
    openai_api_key=<your-deepseek-api-key>,
    openai_api_base='https://api.deepseek.com/v1',
    temperature=0.85,
    max_tokens=8000)

9. License

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

10. Citation

@misc{deepseekv2,
      title={DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model}, 
      author={DeepSeek-AI},
      year={2024},
      eprint={2405.04434},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

11. Contact

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