mirror of
https://git.mirrors.martin98.com/https://github.com/langgenius/dify.git
synced 2025-08-12 23:29:02 +08:00
docs: add english versions for the files customizable_model_scale_out and predefined_model_scale_out (#8871)
This commit is contained in:
parent
c531b4a911
commit
42dfde6546
@ -0,0 +1,310 @@
|
||||
## Custom Integration of Pre-defined Models
|
||||
|
||||
### Introduction
|
||||
|
||||
After completing the vendors integration, the next step is to connect the vendor's models. To illustrate the entire connection process, we will use Xinference as an example to demonstrate a complete vendor integration.
|
||||
|
||||
It is important to note that for custom models, each model connection requires a complete vendor credential.
|
||||
|
||||
Unlike pre-defined models, a custom vendor integration always includes the following two parameters, which do not need to be defined in the vendor YAML file.
|
||||
|
||||

|
||||
|
||||
As mentioned earlier, vendors do not need to implement validate_provider_credential. The runtime will automatically call the corresponding model layer's validate_credentials to validate the credentials based on the model type and name selected by the user.
|
||||
|
||||
### Writing the Vendor YAML
|
||||
|
||||
First, we need to identify the types of models supported by the vendor we are integrating.
|
||||
|
||||
Currently supported model types are as follows:
|
||||
|
||||
- `llm` Text Generation Models
|
||||
|
||||
- `text_embedding` Text Embedding Models
|
||||
|
||||
- `rerank` Rerank Models
|
||||
|
||||
- `speech2text` Speech-to-Text
|
||||
|
||||
- `tts` Text-to-Speech
|
||||
|
||||
- `moderation` Moderation
|
||||
|
||||
Xinference supports LLM, Text Embedding, and Rerank. So we will start by writing xinference.yaml.
|
||||
|
||||
```yaml
|
||||
provider: xinference #Define the vendor identifier
|
||||
label: # Vendor display name, supports both en_US (English) and zh_Hans (Simplified Chinese). If zh_Hans is not set, it will use en_US by default.
|
||||
en_US: Xorbits Inference
|
||||
icon_small: # Small icon, refer to other vendors' icons stored in the _assets directory within the vendor implementation directory; follows the same language policy as the 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: # Supported model types. Xinference supports LLM, Text Embedding, and Rerank
|
||||
- llm
|
||||
- text-embedding
|
||||
- rerank
|
||||
configurate_methods: # Since Xinference is a locally deployed vendor with no predefined models, users need to deploy whatever models they need according to Xinference documentation. Thus, it only supports custom models.
|
||||
- customizable-model
|
||||
provider_credential_schema:
|
||||
credential_form_schemas:
|
||||
```
|
||||
|
||||
|
||||
Then, we need to determine what credentials are required to define a model in Xinference.
|
||||
|
||||
- Since it supports three different types of models, we need to specify the model_type to denote the model type. Here is how we can define it:
|
||||
|
||||
```yaml
|
||||
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
|
||||
```
|
||||
|
||||
- Next, each model has its own model_name, so we need to define that here:
|
||||
|
||||
```yaml
|
||||
- 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
|
||||
```
|
||||
|
||||
- Specify the Xinference local deployment address:
|
||||
|
||||
```yaml
|
||||
- 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
|
||||
```
|
||||
|
||||
- Each model has a unique model_uid, so we also need to define that here:
|
||||
|
||||
```yaml
|
||||
- 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
|
||||
```
|
||||
|
||||
Now, we have completed the basic definition of the vendor.
|
||||
|
||||
### Writing the Model Code
|
||||
|
||||
Next, let's take the `llm` type as an example and write `xinference.llm.llm.py`.
|
||||
|
||||
In `llm.py`, create a Xinference LLM class, we name it `XinferenceAILargeLanguageModel` (this can be arbitrary), inheriting from the `__base.large_language_model.LargeLanguageModel` base class, and implement the following methods:
|
||||
|
||||
- LLM Invocation
|
||||
|
||||
Implement the core method for LLM invocation, supporting both stream and synchronous responses.
|
||||
|
||||
```python
|
||||
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 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 usage
|
||||
:param stop: stop words
|
||||
:param stream: is the response a stream
|
||||
:param user: unique user id
|
||||
:return: full response or stream response chunk generator result
|
||||
"""
|
||||
```
|
||||
|
||||
When implementing, ensure to use two functions to return data separately for synchronous and stream responses. This is important because Python treats functions containing the `yield` keyword as generator functions, mandating them to return `Generator` types. Here’s an example (note that the example uses simplified parameters; in real implementation, use the parameter list as defined above):
|
||||
|
||||
```python
|
||||
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-compute Input Tokens
|
||||
|
||||
If the model does not provide an interface for pre-computing tokens, you can return 0 directly.
|
||||
|
||||
```python
|
||||
def get_num_tokens(self, model: str, credentials: dict, prompt_messages: list[PromptMessage],tools: Optional[list[PromptMessageTool]] = None) -> int:
|
||||
"""
|
||||
Get number of tokens for given prompt messages
|
||||
|
||||
:param model: model name
|
||||
:param credentials: model credentials
|
||||
:param prompt_messages: prompt messages
|
||||
:param tools: tools for tool usage
|
||||
:return: token count
|
||||
"""
|
||||
```
|
||||
|
||||
|
||||
Sometimes, you might not want to return 0 directly. In such cases, you can use `self._get_num_tokens_by_gpt2(text: str)` to get pre-computed tokens. This method is provided by the `AIModel` base class, and it uses GPT2's Tokenizer for calculation. However, it should be noted that this is only a substitute and may not be fully accurate.
|
||||
|
||||
- Model Credentials Validation
|
||||
|
||||
Similar to vendor credentials validation, this method validates individual model credentials.
|
||||
|
||||
```python
|
||||
def validate_credentials(self, model: str, credentials: dict) -> None:
|
||||
"""
|
||||
Validate model credentials
|
||||
|
||||
:param model: model name
|
||||
:param credentials: model credentials
|
||||
:return: None
|
||||
"""
|
||||
```
|
||||
|
||||
- Model Parameter Schema
|
||||
|
||||
Unlike custom types, since the YAML file does not define which parameters a model supports, we need to dynamically generate the model parameter schema.
|
||||
|
||||
For instance, Xinference supports `max_tokens`, `temperature`, and `top_p` parameters.
|
||||
|
||||
However, some vendors may support different parameters for different models. For example, the `OpenLLM` vendor supports `top_k`, but not all models provided by this vendor support `top_k`. Let's say model A supports `top_k` but model B does not. In such cases, we need to dynamically generate the model parameter schema, as illustrated below:
|
||||
|
||||
```python
|
||||
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
|
||||
```
|
||||
|
||||
- Exception Error Mapping
|
||||
|
||||
When a model invocation error occurs, it should be mapped to the runtime's specified `InvokeError` type, enabling Dify to handle different errors appropriately.
|
||||
|
||||
Runtime Errors:
|
||||
|
||||
- `InvokeConnectionError` Connection error during invocation
|
||||
- `InvokeServerUnavailableError` Service provider unavailable
|
||||
- `InvokeRateLimitError` Rate limit reached
|
||||
- `InvokeAuthorizationError` Authorization failure
|
||||
- `InvokeBadRequestError` Invalid request parameters
|
||||
|
||||
```python
|
||||
@property
|
||||
def _invoke_error_mapping(self) -> dict[type[InvokeError], list[type[Exception]]]:
|
||||
"""
|
||||
Map model invoke error to unified error
|
||||
The key is the error type thrown to the caller
|
||||
The value is the error type thrown by the model,
|
||||
which needs to be converted into a unified error type for the caller.
|
||||
|
||||
:return: Invoke error mapping
|
||||
"""
|
||||
```
|
||||
|
||||
For interface method details, see: [Interfaces](./interfaces.md). For specific implementations, refer to: [llm.py](https://github.com/langgenius/dify-runtime/blob/main/lib/model_providers/anthropic/llm/llm.py).
|
BIN
api/core/model_runtime/docs/en_US/images/index/image-1.png
Normal file
BIN
api/core/model_runtime/docs/en_US/images/index/image-1.png
Normal file
Binary file not shown.
After Width: | Height: | Size: 230 KiB |
BIN
api/core/model_runtime/docs/en_US/images/index/image-2.png
Normal file
BIN
api/core/model_runtime/docs/en_US/images/index/image-2.png
Normal file
Binary file not shown.
After Width: | Height: | Size: 205 KiB |
BIN
api/core/model_runtime/docs/en_US/images/index/image-3.png
Normal file
BIN
api/core/model_runtime/docs/en_US/images/index/image-3.png
Normal file
Binary file not shown.
After Width: | Height: | Size: 44 KiB |
BIN
api/core/model_runtime/docs/en_US/images/index/image.png
Normal file
BIN
api/core/model_runtime/docs/en_US/images/index/image.png
Normal file
Binary file not shown.
After Width: | Height: | Size: 262 KiB |
173
api/core/model_runtime/docs/en_US/predefined_model_scale_out.md
Normal file
173
api/core/model_runtime/docs/en_US/predefined_model_scale_out.md
Normal file
@ -0,0 +1,173 @@
|
||||
## Predefined Model Integration
|
||||
|
||||
After completing the vendor integration, the next step is to integrate the models from the vendor.
|
||||
|
||||
First, we need to determine the type of model to be integrated and create the corresponding model type `module` under the respective vendor's directory.
|
||||
|
||||
Currently supported model types are:
|
||||
|
||||
- `llm` Text Generation Model
|
||||
- `text_embedding` Text Embedding Model
|
||||
- `rerank` Rerank Model
|
||||
- `speech2text` Speech-to-Text
|
||||
- `tts` Text-to-Speech
|
||||
- `moderation` Moderation
|
||||
|
||||
Continuing with `Anthropic` as an example, `Anthropic` only supports LLM, so create a `module` named `llm` under `model_providers.anthropic`.
|
||||
|
||||
For predefined models, we first need to create a YAML file named after the model under the `llm` `module`, such as `claude-2.1.yaml`.
|
||||
|
||||
### Prepare Model YAML
|
||||
|
||||
```yaml
|
||||
model: claude-2.1 # Model identifier
|
||||
# Display name of the model, which can be set to en_US English or zh_Hans Chinese. If zh_Hans is not set, it will default to en_US.
|
||||
# This can also be omitted, in which case the model identifier will be used as the label
|
||||
label:
|
||||
en_US: claude-2.1
|
||||
model_type: llm # Model type, claude-2.1 is an LLM
|
||||
features: # Supported features, agent-thought supports Agent reasoning, vision supports image understanding
|
||||
- agent-thought
|
||||
model_properties: # Model properties
|
||||
mode: chat # LLM mode, complete for text completion models, chat for conversation models
|
||||
context_size: 200000 # Maximum context size
|
||||
parameter_rules: # Parameter rules for the model call; only LLM requires this
|
||||
- name: temperature # Parameter variable name
|
||||
# Five default configuration templates are provided: temperature/top_p/max_tokens/presence_penalty/frequency_penalty
|
||||
# The template variable name can be set directly in use_template, which will use the default configuration in entities.defaults.PARAMETER_RULE_TEMPLATE
|
||||
# Additional configuration parameters will override the default configuration if set
|
||||
use_template: temperature
|
||||
- name: top_p
|
||||
use_template: top_p
|
||||
- name: top_k
|
||||
label: # Display name of the parameter
|
||||
zh_Hans: 取样数量
|
||||
en_US: Top k
|
||||
type: int # Parameter type, supports float/int/string/boolean
|
||||
help: # Help information, describing the parameter's function
|
||||
zh_Hans: 仅从每个后续标记的前 K 个选项中采样。
|
||||
en_US: Only sample from the top K options for each subsequent token.
|
||||
required: false # Whether the parameter is mandatory; can be omitted
|
||||
- name: max_tokens_to_sample
|
||||
use_template: max_tokens
|
||||
default: 4096 # Default value of the parameter
|
||||
min: 1 # Minimum value of the parameter, applicable to float/int only
|
||||
max: 4096 # Maximum value of the parameter, applicable to float/int only
|
||||
pricing: # Pricing information
|
||||
input: '8.00' # Input unit price, i.e., prompt price
|
||||
output: '24.00' # Output unit price, i.e., response content price
|
||||
unit: '0.000001' # Price unit, meaning the above prices are per 100K
|
||||
currency: USD # Price currency
|
||||
```
|
||||
|
||||
It is recommended to prepare all model configurations before starting the implementation of the model code.
|
||||
|
||||
You can also refer to the YAML configuration information under the corresponding model type directories of other vendors in the `model_providers` directory. For the complete YAML rules, refer to: [Schema](schema.md#aimodelentity).
|
||||
|
||||
### Implement the Model Call Code
|
||||
|
||||
Next, create a Python file named `llm.py` under the `llm` `module` to write the implementation code.
|
||||
|
||||
Create an Anthropic LLM class named `AnthropicLargeLanguageModel` (or any other name), inheriting from the `__base.large_language_model.LargeLanguageModel` base class, and implement the following methods:
|
||||
|
||||
- LLM Call
|
||||
|
||||
Implement the core method for calling the LLM, supporting both streaming and synchronous responses.
|
||||
|
||||
```python
|
||||
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 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: is stream response
|
||||
:param user: unique user id
|
||||
:return: full response or stream response chunk generator result
|
||||
"""
|
||||
```
|
||||
|
||||
Ensure to use two functions for returning data, one for synchronous returns and the other for streaming returns, because Python identifies functions containing the `yield` keyword as generator functions, fixing the return type to `Generator`. Thus, synchronous and streaming returns need to be implemented separately, as shown below (note that the example uses simplified parameters, for actual implementation follow the above parameter list):
|
||||
|
||||
```python
|
||||
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-compute Input Tokens
|
||||
|
||||
If the model does not provide an interface to precompute tokens, return 0 directly.
|
||||
|
||||
```python
|
||||
def get_num_tokens(self, model: str, credentials: dict, prompt_messages: list[PromptMessage],
|
||||
tools: Optional[list[PromptMessageTool]] = None) -> int:
|
||||
"""
|
||||
Get number of tokens for given prompt messages
|
||||
|
||||
:param model: model name
|
||||
:param credentials: model credentials
|
||||
:param prompt_messages: prompt messages
|
||||
:param tools: tools for tool calling
|
||||
:return:
|
||||
"""
|
||||
```
|
||||
|
||||
- Validate Model Credentials
|
||||
|
||||
Similar to vendor credential validation, but specific to a single model.
|
||||
|
||||
```python
|
||||
def validate_credentials(self, model: str, credentials: dict) -> None:
|
||||
"""
|
||||
Validate model credentials
|
||||
|
||||
:param model: model name
|
||||
:param credentials: model credentials
|
||||
:return:
|
||||
"""
|
||||
```
|
||||
|
||||
- Map Invoke Errors
|
||||
|
||||
When a model call fails, map it to a specific `InvokeError` type as required by Runtime, allowing Dify to handle different errors accordingly.
|
||||
|
||||
Runtime Errors:
|
||||
|
||||
- `InvokeConnectionError` Connection error
|
||||
|
||||
- `InvokeServerUnavailableError` Service provider unavailable
|
||||
- `InvokeRateLimitError` Rate limit reached
|
||||
- `InvokeAuthorizationError` Authorization failed
|
||||
- `InvokeBadRequestError` Parameter error
|
||||
|
||||
```python
|
||||
@property
|
||||
def _invoke_error_mapping(self) -> dict[type[InvokeError], list[type[Exception]]]:
|
||||
"""
|
||||
Map model invoke error to unified error
|
||||
The key is the error type thrown to the caller
|
||||
The value is the error type thrown by the model,
|
||||
which needs to be converted into a unified error type for the caller.
|
||||
|
||||
:return: Invoke error mapping
|
||||
"""
|
||||
```
|
||||
|
||||
For interface method explanations, see: [Interfaces](./interfaces.md). For detailed implementation, refer to: [llm.py](https://github.com/langgenius/dify-runtime/blob/main/lib/model_providers/anthropic/llm/llm.py).
|
@ -58,7 +58,7 @@ provider_credential_schema: # Provider credential rules, as Anthropic only supp
|
||||
en_US: Enter your API URL
|
||||
```
|
||||
|
||||
You can also refer to the YAML configuration information under other provider directories in `model_providers`. The complete YAML rules are available at: [Schema](schema.md#Provider).
|
||||
You can also refer to the YAML configuration information under other provider directories in `model_providers`. The complete YAML rules are available at: [Schema](schema.md#provider).
|
||||
|
||||
### Implementing Provider Code
|
||||
|
||||
|
@ -117,7 +117,7 @@ model_credential_schema:
|
||||
en_US: Enter your API Base
|
||||
```
|
||||
|
||||
也可以参考 `model_providers` 目录下其他供应商目录下的 YAML 配置信息,完整的 YAML 规则见:[Schema](schema.md#Provider)。
|
||||
也可以参考 `model_providers` 目录下其他供应商目录下的 YAML 配置信息,完整的 YAML 规则见:[Schema](schema.md#provider)。
|
||||
|
||||
#### 实现供应商代码
|
||||
|
||||
|
Loading…
x
Reference in New Issue
Block a user