DeepSeek-V2.5：兼具通用能力和编码能力的新型开源模型

DeepSeek-V2.5 将聊天和编码功能合二为一，现已开源。 针对写作、编码和人性化偏好进行了增强。 通过网络和 API 提供。

今天，我们成功合并了 DeepSeek-V2-Chat 和 DeepSeek-Coder-V2 模型，正式发布了DeepSeek-V2.5。

DeepSeek-V2.5 保留了Chat模型的一般对话能力和Coder模型的编码优势，同时更符合人类的偏好。 此外，DeepSeek-V2.5 在编写任务和指令跟踪方面也有了显著的改进。

DeepSeek-V2.5 现在已经完全可以在Web和API平台上使用。 API保持向后兼容，允许用户通过 deepseek-coder 或 deepseek-chat 访问新模型。 函数调用、FIM 补全和 JSON 输出功能保持不变。 一体化的 DeepSeek-V2.5 提供了更简单、更智能、更高效的用户体验。

Metric	DeepSeek-V2-0628	DeepSeek-Coder-V2-0724	DeepSeek-V2.5
AlpacaEval 2.0	46.6	44.5	50.5
ArenaHard	68.3	66.3	76.2
AlignBench	7.88	7.91	8.04
MT-Bench	8.85	8.91	9.02
HumanEval python	84.5	87.2	89
HumanEval Multi	73.8	74.8	73.8
LiveCodeBench(01-09)	36.6	39.7	41.8
Aider	69.9	72.9	72.2
SWE-verified	N/A	19	16.8
DS-FIM-Eval	N/A	73.2	78.3
DS-Arena-Code	N/A	49.5	63.1

如何在本地运行

要利用BF16格式的DeepSeek-V2.5进行推理，需要80GB*8个GPU。 利用Huggingface的Transformers进行推理

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig

model_name = "deepseek-ai/DeepSeek-V2.5"
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)

完整的聊天模板可在 huggingface 模型仓库中的 tokenizer_config.json 中找到。

注：与之前的 DeepSeek-V2-Chat 版本相比，聊天模板已经更新。

聊天模板示例如下：

<｜begin▁of▁sentence｜><｜User｜>{user_message_1}<｜Assistant｜>{assistant_message_1}<｜end▁of▁sentence｜><｜User｜>{user_message_2}<｜Assistant｜>

您还可以添加一条可选的系统信息：

<｜begin▁of▁sentence｜>{system_message}<｜User｜>{user_message_1}<｜Assistant｜>{assistant_message_1}<｜end▁of▁sentence｜><｜User｜>{user_message_2}<｜Assistant｜>

使用 vLLM 进行推理（推荐）

要利用 vLLM 进行模型推理，请将此 Pull Request 合并到您的 vLLM 代码库中：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-V2.5"
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)

函数调用

函数调用允许模型调用外部工具来增强其功能。
下面是一个例子：

# Assume that `model` and `tokenizer` are loaded
model.generation_config = GenerationConfig(do_sample=False, max_new_tokens=128, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id)

tool_system_prompt = """You are a helpful Assistant.

## Tools

### Function

You have the following functions available:

- `get_current_weather`:
```json
{
    "name": "get_current_weather",
    "description": "Get the current weather in a given location",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "The city and state, e.g. San Francisco, CA"
            },
            "unit": {
                "type": "string",
                "enum": [
                    "celsius",
                    "fahrenheit"
                ]
            }
        },
        "required": [
            "location"
        ]
    }
}
```"""

tool_call_messages = [{"role": "system", "content": tool_system_prompt}, {"role": "user", "content": "What's the weather like in Tokyo and Paris?"}]
tool_call_inputs = tokenizer.apply_chat_template(tool_call_messages, add_generation_prompt=True, return_tensors="pt")
tool_call_outputs = model.generate(tool_call_inputs.to(model.device))
# Generated text: '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>get_current_weather\n```json\n{"location": "Tokyo"}\n```<｜tool▁call▁end｜>\n<｜tool▁call▁begin｜>function<｜tool▁sep｜>get_current_weather\n```json\n{"location": "Paris"}\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜><｜end▁of▁sentence｜>'

# Mock response of calling `get_current_weather`
tool_messages = [{"role": "tool", "content": '{"location": "Tokyo", "temperature": "10", "unit": null}'}, {"role": "tool", "content": '{"location": "Paris", "temperature": "22", "unit": null}'}]
tool_inputs = tokenizer.apply_chat_template(tool_messages, add_generation_prompt=False, return_tensors="pt")[:, 1:]
tool_inputs = torch.cat([tool_call_outputs, tool_inputs.to(model.device)], dim=1)
tool_outputs = model.generate(tool_inputs)
# Generated text: The current weather in Tokyo is 10 degrees, and in Paris, it is 22 degrees.<｜end▁of▁sentence｜>

JSON 输出

您可以使用 JSON 输出模式来确保模型生成有效的 JSON 对象。 要激活此模式，应在系统提示中添加一条特殊指令。

# Assume that `model` and `tokenizer` are loaded
model.generation_config = GenerationConfig(do_sample=False, max_new_tokens=128, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id)

user_system_prompt = 'The user will provide some exam text. Please parse the "question" and "answer" and output them in JSON format.'
json_system_prompt = f"""{user_system_prompt}

## Response Format

Reply with JSON object ONLY."""

json_messages = [{"role": "system", "content": json_system_prompt}, {"role": "user", "content": "Which is the highest mountain in the world? Mount Everest."}]
json_inputs = tokenizer.apply_chat_template(json_messages, add_generation_prompt=True, return_tensors="pt")
json_outpus = model.generate(json_inputs.to(model.device))
# Generated text: '```json\n{\n  "question": "Which is the highest mountain in the world?",\n  "answer": "Mount Everest."\n}\n```<｜end▁of▁sentence｜>'

FIM 补全

在 FIM（中间填充）补全中，您可以提供一个前缀和一个可选的后缀，模型将完成中间的内容。

# Assume that `model` and `tokenizer` are loaded
model.generation_config = GenerationConfig(do_sample=False, max_new_tokens=128, eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.eos_token_id)

prefix = """def quick_sort(arr):
    if len(arr) <= 1:
        return arr
    pivot = arr[0]
    left = []
    right = []
"""

suffix = """
        if arr[i] < pivot:
            left.append(arr[i])
        else:
            right.append(arr[i])
    return quick_sort(left) + [pivot] + quick_sort(right)"""

fim_prompt = f"<｜fim▁begin｜>{prefix}<｜fim▁hole｜>{suffix}<｜fim▁end｜>"
fim_inputs = tokenizer(fim_prompt, add_special_tokens=True, return_tensors="pt").input_ids
fim_outputs = model.generate(fim_inputs.to(model.device))
# Generated text: "    for i in range(1, len(arr)):<｜end▁of▁sentence｜>"