GPT Carry-On: Training Foundation Model for Customization Could Be Simple, Scalable and Affordable
This addresses the challenge of making LLM customization scalable and affordable for deployment in online services, though it appears incremental as it builds on existing transformer architectures.
The paper tackles the problem of customizing large language models (LLMs) for individual users or tasks by proposing a framework that trains a lightweight 'carry-on' module on top of pretrained LLMs, reducing GPU memory usage to less than 1GB for a 100M parameter module on a 30B LLM and enabling faster loss convergence with minimal data and parameters.
Modern large language foundation models (LLM) have now entered the daily lives of millions of users. We ask a natural question whether it is possible to customize LLM for every user or every task. From system and industrial economy consideration, general continue-training or fine-tuning still require substantial computation and memory of training GPU nodes, whereas most inference nodes under deployment, possibly with lower-end GPUs, are configured to make forward pass fastest possible. We propose a framework to take full advantages of existing LLMs and systems of online service. We train an additional branch of transformer blocks on the final-layer embedding of pretrained LLMs, which is the base, then a carry-on module merge the base models to compose a customized LLM. We can mix multiple layers, or multiple LLMs specialized in different domains such as chat, coding, math, to form a new mixture of LLM that best fit a new task. As the base model don't need to update parameters, we are able to outsource most computation of the training job on inference nodes, and only train a lightweight carry-on on training nodes, where we consume less than 1GB GPU memory to train a 100M carry-on layer on 30B LLM. We tested Qwen and DeepSeek opensourced models for continue-pretraining and got faster loss convergence. We use it to improve solving math questions with extremely small computation and model size, with 1000 data samples of chain-of-thoughts, and as small as 1 MB parameters of two layer layer carry-on, and the results are promising.