【Feature】add fd plugins && rm model_classes

[gzy19990617]

PR: Add FastDeploy Plugin Mechanism Support

Overview

This PR introduces the FastDeploy plugin mechanism, which allows users to extend functionality without modifying the core code. The key changes include:

1. New Feature: Added FastDeploy plugin mechanism support
2. Refactor: Removed deprecated mode_classed
3. Standardization: Standardized network architecture name
4. Documentation: Comprehensive plugin mechanism documentation (see below)

FastDeploy Plugin Mechanism Documentation

FastDeploy supports a plugin mechanism that allows users to extend functionality without modifying the core code. Plugins are automatically discovered and loaded through Python's entry_points mechanism.

How Plugins Work

Plugins are essentially registration functions that are automatically called when FastDeploy starts. The system uses the load_plugins_by_group function to ensure that all processes (including child processes in distributed training scenarios) have loaded the required plugins before official operations begin.

Plugin Discovery Mechanism

FastDeploy uses Python's entry_points mechanism to discover and load plugins. Developers need to register their plugins in the specified entry point group in their project.

Example: Creating a Plugin

1. Write Plugin Logic

Assuming you have a custom model class MyModelForCasualLM and a pretrained class MyPretrainedModel, you can write the following registration function:

# File: fd_add_dummy_model/__init__.py
from fastdeploy.model_registry import ModelRegistry
from my_custom_model import MyModelForCasualLM, MyPretrainedModel

def register():
    if "MyModelForCasualLM" not in ModelRegistry.get_supported_archs():
        ModelRegistry.register_model_class(MyModelForCasualLM)
        ModelRegistry.register_pretrained_model(MyPretrainedModel)

2. Register Plugin in setup.py

# setup.py
from setuptools import setup

setup(
    name="fastdeploy-plugins",
    version="0.1",
    packages=["fd_add_dummy_model"],
    entry_points={
        "fastdeploy.model_register_plugins": [
            "fd_add_dummy_model = fd_add_dummy_model:register",
        ],
    },
)

Plugin Structure

Plugins consist of three components:

Component	Description
Plugin Group	The functional group to which the plugin belongs, for example: - fastdeploy.model_register_plugins: for model registration - fastdeploy.model_runner_plugins: for model runner registration Users can customize groups as needed.
Plugin Name	The unique identifier for each plugin (e.g., fd_add_dummy_model), which can be controlled via the FD_PLUGINS environment variable to determine whether to load the plugin.
Plugin Value	Format is module_name:function_name, pointing to the entry function that executes the registration logic.

Controlling Plugin Loading Behavior

By default, FastDeploy loads all registered plugins. To load only specific plugins, you can set the environment variable:

export FD_PLUGINS=fastdeploy-plugins

Multiple plugin names can be separated by commas:

export FD_PLUGINS=plugin_a,plugin_b

Reference Example

Please refer to the example plugin implementation in the project directory:

./test/plugins/

It contains a complete plugin structure and setup.py configuration example.

Summary

Through the plugin mechanism, users can easily add custom models or functional modules to FastDeploy without modifying the core source code. This not only enhances system extensibility but also facilitates third-party developers in extending functionality.

For further plugin development, please refer to the model_registry and plugin_loader modules in the FastDeploy source code.

[paddle-bot]

Thanks for your contribution!

[yuanlehome]

PR描述详细一下，并且添加文档，这个PR属于新增plugins特性，最好添加方案和规范指导