📡The Overview of Our Framework
Considering the above issues, in this paper, we propose a new framework for LLM training and knowledge exchange, namely The Internet of LLM. The figure below shows the overview of the framework. We implemented four innovations in the Internet of LLM, with some based on secondary development of Langchain and Ollama:
LLM Sharing Protocol: Given the substantial need for LLMs in using, constructing Workflows, and developing Agents, the rapid transfer of LLMs across different regions presents a significant challenge. To address this, we devised a universal LLM model protocol to facilitate one-click integration and swift sharing.
LLM Universal Environment: Diverse environments pose numerous adaptation challenges for developers and users. To reduce entry barriers, we established a unified platform for development, training, and execution, thereby minimizing the time developers spend on environment setup, version control, and troubleshooting.
Agent Optimal Path: When handling complex tasks, the system continually selects and combines models, conducts joint training of multiple models, and evaluates and provides feedback on the results. Due to the time-intensive nature of this process, we employed parallel computing techniques to expedite the search for the optimal pathway that meets the current requirements.
Joint Mining: To alleviate the initial computational costs for researchers, computing power providers can contribute by offering computational resources. In return, they share two benefits with model designers: breakthrough rewards for optimal model pathways and long-term returns from the models. This arrangement enables researchers to access discounted or complimentary computational support.
Last updated