Integrate the Customizable Model

A custom model refers to an LLM that you deploy or configure on your own. This document uses the Xinference model as an example to demonstrate how to integrate a custom model into your model plugin.

By default, a custom model automatically includes two parameters—its model type and model name—and does not require additional definitions in the provider YAML file.

You do not need to implement validate_provider_credential in your provider configuration file. During runtime, based on the user’s choice of model type or model name, Dify automatically calls the corresponding model layer’s validate_credentials method to verify credentials.

Integrating a Custom Model Plugin

Below are the steps to integrate a custom model:

Create a Model Provider File Identify the model types your custom model will include.
Create Code Files by Model Type Depending on the model’s type (e.g., llm or text_embedding), create separate code files. Ensure that each model type is organized into distinct logical layers for easier maintenance and future expansion.
Develop the Model Invocation Logic Within each model-type module, create a Python file named for that model type (for example, llm.py). Define a class in the file that implements the specific model logic, conforming to the system’s model interface specifications.
Debug the Plugin Write unit and integration tests for the new provider functionality, ensuring that all components work as intended.

1. Create a Model Provider File

In your plugin’s /provider directory, create a xinference.yaml file.

The Xinference family of models supports LLM, Text Embedding, and Rerank model types, so your xinference.yaml must include all three.

Example:

provider: xinference  # Identifies the provider
label:                # Display name; can set both en_US (English) and zh_Hans (Chinese). If zh_Hans is not set, en_US is used by default.
  en_US: Xorbits Inference
icon_small:           # Small icon; store in the _assets folder of this provider’s directory. The same multi-language logic applies as with label.
  en_US: icon_s_en.svg
icon_large:           # Large icon
  en_US: icon_l_en.svg
help:                 # Help information
  title:
    en_US: How to deploy Xinference
    zh_Hans: 如何部署 Xinference
  url:
    en_US: https://github.com/xorbitsai/inference

supported_model_types:  # Model types Xinference supports: LLM/Text Embedding/Rerank
- llm
- text-embedding
- rerank

configurate_methods:     # Xinference is locally deployed and does not offer predefined models. Refer to its documentation to learn which model to use. Thus, we choose a customizable-model approach.
- customizable-model

provider_credential_schema:
  credential_form_schemas:

Next, define the provider_credential_schema. Since Xinference supports text-generation, embeddings, and reranking models, you can configure it as follows:

provider_credential_schema:
  credential_form_schemas:
  - variable: model_type
    type: select
    label:
      en_US: Model type
      zh_Hans: 模型类型
    required: true
    options:
    - value: text-generation
      label:
        en_US: Language Model
        zh_Hans: 语言模型
    - value: embeddings
      label:
        en_US: Text Embedding
    - value: reranking
      label:
        en_US: Rerank

Every model in Xinference requires a model_name:

  - variable: model_name
    type: text-input
    label:
      en_US: Model name
      zh_Hans: 模型名称
    required: true
    placeholder:
      zh_Hans: 填写模型名称
      en_US: Input model name

Because Xinference must be locally deployed, users need to supply the server address (server_url) and model UID. For instance:

  - variable: server_url
    label:
      zh_Hans: 服务器 URL
      en_US: Server url
    type: text-input
    required: true
    placeholder:
      zh_Hans: 在此输入 Xinference 的服务器地址，如 https://example.com/xxx
      en_US: Enter the url of your Xinference, for example https://example.com/xxx

  - variable: model_uid
    label:
      zh_Hans: 模型 UID
      en_US: Model uid
    type: text-input
    required: true
    placeholder:
      zh_Hans: 在此输入您的 Model UID
      en_US: Enter the model uid

Once you’ve defined these parameters, the YAML configuration for your custom model provider is complete. Next, create the functional code files for each model defined in this config.

2. Develop the Model Code

Since Xinference supports llm, rerank, speech2text, and tts, you should create corresponding directories under /models, each containing its respective feature code.

Below is an example for an llm-type model. You’d create a file named llm.py, then define a class—such as XinferenceAILargeLanguageModel—that extends __base.large_language_model.LargeLanguageModel. This class should include:

LLM Invocation

The core method for invoking the LLM, supporting both streaming and synchronous responses:

def _invoke(
    self,
    model: str,
    credentials: dict,
    prompt_messages: list[PromptMessage],
    model_parameters: dict,
    tools: Optional[list[PromptMessageTool]] = None,
    stop: Optional[list[str]] = None,
    stream: bool = True,
    user: Optional[str] = None
) -> Union[LLMResult, Generator]:
    """
    Invoke the large language model.

    :param model: model name
    :param credentials: model credentials
    :param prompt_messages: prompt messages
    :param model_parameters: model parameters
    :param tools: tools for tool calling
    :param stop: stop words
    :param stream: determines if response is streamed
    :param user: unique user id
    :return: full response or a chunk generator
    """

You’ll need two separate functions to handle streaming and synchronous responses. Python treats any function containing yield as a generator returning type Generator, so it’s best to split them:

def _invoke(self, stream: bool, **kwargs) -> Union[LLMResult, Generator]:
    if stream:
        return self._handle_stream_response(**kwargs)
    return self._handle_sync_response(**kwargs)

def _handle_stream_response(self, **kwargs) -> Generator:
    for chunk in response:
        yield chunk

def _handle_sync_response(self, **kwargs) -> LLMResult:
    return LLMResult(**response)

Pre-calculating Input Tokens

If your model doesn’t provide a token-counting interface, simply return 0:

def get_num_tokens(
    self,
    model: str,
    credentials: dict,
    prompt_messages: list[PromptMessage],
    tools: Optional[list[PromptMessageTool]] = None
) -> int:
    """
    Get the number of tokens for the given prompt messages.
    """
    return 0

Alternatively, you can call self._get_num_tokens_by_gpt2(text: str) from the AIModel base class, which uses a GPT-2 tokenizer. Remember this is an approximation and may not match your model exactly.

Validating Model Credentials

Similar to provider-level credential checks, but scoped to a single model:

def validate_credentials(self, model: str, credentials: dict) -> None:
    """
    Validate model credentials.
    """

Dynamic Model Parameters Schema

Unlike predefined models, no YAML is defining which parameters a model supports. You must generate a parameter schema dynamically.

For example, Xinference supports max_tokens, temperature, and top_p. Some other providers (e.g., OpenLLM) may support parameters like top_k only for certain models. This means you need to adapt your schema to each model’s capabilities:

def get_customizable_model_schema(self, model: str, credentials: dict) -> AIModelEntity | None:
    """
        used to define customizable model schema
    """
    rules = [
        ParameterRule(
            name='temperature', type=ParameterType.FLOAT,
            use_template='temperature',
            label=I18nObject(
                zh_Hans='温度', en_US='Temperature'
            )
        ),
        ParameterRule(
            name='top_p', type=ParameterType.FLOAT,
            use_template='top_p',
            label=I18nObject(
                zh_Hans='Top P', en_US='Top P'
            )
        ),
        ParameterRule(
            name='max_tokens', type=ParameterType.INT,
            use_template='max_tokens',
            min=1,
            default=512,
            label=I18nObject(
                zh_Hans='最大生成长度', en_US='Max Tokens'
            )
        )
    ]

    # if model is A, add top_k to rules
    if model == 'A':
        rules.append(
            ParameterRule(
                name='top_k', type=ParameterType.INT,
                use_template='top_k',
                min=1,
                default=50,
                label=I18nObject(
                    zh_Hans='Top K', en_US='Top K'
                )
            )
        )

    """
        some NOT IMPORTANT code here
    """

    entity = AIModelEntity(
        model=model,
        label=I18nObject(
            en_US=model
        ),
        fetch_from=FetchFrom.CUSTOMIZABLE_MODEL,
        model_type=model_type,
        model_properties={ 
            ModelPropertyKey.MODE:  ModelType.LLM,
        },
        parameter_rules=rules
    )

    return entity

Error Mapping

When an error occurs during model invocation, map it to the appropriate InvokeError type recognized by the runtime. This lets Dify handle different errors in a standardized manner:

Runtime Errors:

•	`InvokeConnectionError`
•	`InvokeServerUnavailableError`
•	`InvokeRateLimitError`
•	`InvokeAuthorizationError`
•	`InvokeBadRequestError`

@property
def _invoke_error_mapping(self) -> dict[type[InvokeError], list[type[Exception]]]:
    """
    Map model invocation errors to unified error types.
    The key is the error type thrown to the caller.
    The value is the error type thrown by the model, which needs to be mapped to a
    unified Dify error for consistent handling.
    """
    # return {
    #   InvokeConnectionError: [requests.exceptions.ConnectionError],
    #   ...
    # }