Deepseek support

[irtiq7]

Hey Devs,
let me start by saying that this repo is great. Good job on your work, and thanks for sharing it.

Could we include support of DeepSeek V3 API in ChatDev? If there is a solution out there then when can it be implemented?

Thanks!

[SekiyoKana]

In fact, you can handle the compatibility yourself because the Deepseek official API provides OpenAI support. You only need to replace the two environment variables, BASE_URLand OPENAI_API_KEY, with the Deepseek API https://api.deepseek.com and your deepseek api_key, respectively. This will essentially connect you to the Deepseek model.
though some minor adjustments in the code are still required:

However, such modifications seem to render the ChatGPT model unusable, so this can only serve as a temporary solution.

typing.pt 54:

// add 'DEEPSEEK_CHAT ' to ModelType
class ModelType(Enum):
    GPT_3_5_TURBO = "gpt-3.5-turbo-16k-0613"
    GPT_3_5_TURBO_NEW = "gpt-3.5-turbo-16k"
    GPT_4 = "gpt-4"
    GPT_4_32k = "gpt-4-32k"
    GPT_4_TURBO = "gpt-4-turbo"
    GPT_4_TURBO_V = "gpt-4-turbo"
    GPT_4O = "gpt-4o"
    GPT_4O_MINI = "gpt-4o-mini"
    DEEPSEEK_CHAT = "deepseek-chat"

run.py 81:

// switch default model to 'DEEPSEEK_CHAT' and add 'DEEPSEEK_CHAT' into help
parser.add_argument('--model', type=str, default="DEEPSEEK_CHAT",
                    help="GPT Model, choose from {'GPT_3_5_TURBO', 'GPT_4', 'GPT_4_TURBO', 'GPT_4O', 'GPT_4O_MINI', 'DEEPSEEK_CHAT'}")

model_backend.py 68:

// use get_encoding if model is 'deepseek-chat'
if self.model_type.value == 'deepseek-chat':
            encoding = tiktoken.get_encoding('cl100k_base')
        else:
            encoding = tiktoken.encoding_for_model(self.model_type.value)

utils.py 80:

// same as model_backend.py
if value_for_tiktoken == 'deepseek-chat':
            encoding = tiktoken.get_encoding("cl100k_base")
        else:
            encoding = tiktoken.encoding_for_model(value_for_tiktoken)

finally, change all of the default model ModelType.GPT_3_5_TURBO to ModelType.DEEPSEEK_CHAT.

Looking forward to the official offering of a better integrated solution.

[irtiq7]

That's great @SekiyoKana!!! I will try your method. I was wondering if you were able to download and run the Deepseek model locally?