ROS2 部署大语言模型节点

4GB GPU的DeepSeek-Coder 1.3B模型，并且它已经被量化或优化过。以下是具体的步骤：

安装必要的依赖项：

pip install transformers torch grpcio googleapis-common-protos

创建一个新的ROS 2包：

cd ~/ros2_ws/src
ros2 pkg create --build-type ament_python llm_ros2_node --dependencies rclpy std_msgs grpcio googleapis-common-protos torch transformers

编辑setup.py文件以包含所需的依赖项：

from setuptools import setup

package_name = 'llm_ros2_node'

setup(
    name=package_name,
    version='0.0.0',
    packages=[package_name],
    data_files=[
        ('share/ament_index/resource_index/packages', ['resource/' + package_name]),
        ('share/' + package_name, ['package.xml']),
    ],
    install_requires=['setuptools'],
    zip_safe=True,
    maintainer='your_name',
    maintainer_email='your_email@example.com',
    description='TODO: Package description',
    license='Apache License 2.0',
    tests_require=['pytest'],
    entry_points={
        'console_scripts': [
            'llm_node = llm_ros2_node.llm_node:main',
        ],
    },
)

编写ROS 2节点代码：在这个节点中，我们将订阅一个话题并发送消息到本地的大语言模型，然后将结果发布到另一个话题。

import rclpy
from rclpy.node import Node
from std_msgs.msg import String
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

class LLMNode(Node):
    def __init__(self):
        super().__init__('llm_node')
        self.subscription = self.create_subscription(
            String,
            'input_text',
            self.listener_callback,
            10)
        self.publisher_ = self.create_publisher(String, 'output_text', 10)

        # Load the DeepSeek-Coder model and tokenizer
        self.model_name_or_path = "path/to/deepseek-coder-1.3b-optimized"
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.tokenizer = AutoTokenizer.from_pretrained(self.model_name_or_path)
        self.model = AutoModelForCausalLM.from_pretrained(self.model_name_or_path).to(self.device)
        self.model.eval()

    def listener_callback(self, msg):
        self.get_logger().info(f'Received input text: {msg.data}')
        response = self.call_llm(msg.data)
        self.publisher_.publish(String(data=response))

    def call_llm(self, prompt):
        inputs = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)
        outputs = self.model.generate(inputs, max_length=50, num_return_sequences=1)
        reply = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
        return reply

def main(args=None):
    rclpy.init(args=args)
    llm_node = LLMNode()
    rclpy.spin(llm_node)
    llm_node.destroy_node()
    rclpy.shutdown()

if __name__ == '__main__':
    main()