TinyGPT-V: Efficient Multimodal Large Language Model via Small Backbones (2312.16862v3)

Abstract: In recent years, multimodal LLMs (MLLMs) such as GPT-4V have demonstrated remarkable advancements, excelling in a variety of vision-language tasks. Despite their prowess, the closed-source nature and computational demands of such models limit their accessibility and applicability. This study introduces TinyGPT-V, a novel open-source MLLM, designed for efficient training and inference across various vision-language tasks, including image captioning (IC) and visual question answering (VQA). Leveraging a compact yet powerful architecture, TinyGPT-V integrates the Phi-2 LLM with pre-trained vision encoders, utilizing a unique mapping module for visual and linguistic information fusion. With a training regimen optimized for small backbones and employing a diverse dataset amalgam, TinyGPT-V requires significantly lower computational resources 24GB for training and as little as 8GB for inference without compromising on performance. Our experiments demonstrate that TinyGPT-V, with its LLM 2.8 billion parameters, achieves comparable results in VQA and image inference tasks to its larger counterparts while being uniquely suited for deployment on resource-constrained devices through innovative quantization techniques. This work not only paves the way for more accessible and efficient MLLMs but also underscores the potential of smaller, optimized models in bridging the gap between high performance and computational efficiency in real-world applications. Additionally, this paper introduces a new approach to multimodal LLMs using smaller backbones. Our code and training weights are available in the supplementary material.

• The paper introduces TinyGPT-V, a resource-efficient multimodal LLM that leverages small backbones for effective vision-language fusion.
• It employs a four-stage training process and novel normalization techniques like QK Normalization, achieving competitive results with lower computational cost.
• Experimental evaluations confirm robust visual reasoning and fast inference, enabling deployment on devices with limited memory.

TinyGPT-V: Efficient Multimodal LLM via Small Backbones

The paper "TinyGPT-V: Efficient Multimodal LLM via Small Backbones" introduces TinyGPT-V, an open-source Multimodal LLM (MLLM) specifically designed to balance accessibility with performance across various vision-language tasks. Unlike large proprietary models such as GPT-4V, TinyGPT-V emphasizes efficiency, making it suitable for deployment on devices with limited computational resources while maintaining competitive task performance.

Model Architecture and Innovations

TinyGPT-V's architecture features a compact yet scalable design, incorporating the Phi-2 LLM alongside pre-trained vision encoders such as CLIP. This model utilizes a unique mapping module to integrate visual information with linguistic inputs, facilitating effective vision-language fusion without demanding excessive computational power.

Figure 1: TinyGPT-V narrows the occupancy ratio of the LLM compared to MiniGPT-4.

A key aspect of TinyGPT-V's architecture is its reliance on small backbones, utilizing techniques like LoRA for efficient fine-tuning (Figure 3a). The model's training process incorporates a sequence of pre-training and fine-tuning stages, leveraging datasets like LAION and Conceptual Captions for multi-task learning (Table 3). By adopting novel normalization strategies, including QK Normalization, the model enhances training stability for multimodal data (Figure 3d).

Figure 3: Despite fewer parameters, TinyGPT-V achieves comparable performance in visual language tasks.

Training Methodology

TinyGPT-V employs a structured four-stage training process, optimizing its performance on image-text pairs through a combination of warm-up, pre-training, instruction tuning, and multi-task fine-tuning. This modular approach allows for effective resource utilization, requiring only 24GB of memory for training and enabling deployment with as little as 8GB for inference.

The introduction of sophisticated normalization techniques, including input Layer Norm and RMS Norm, is critical in mitigating gradient vanishing issues common in smaller models. This ensures robust performance during multimodal data interactions.

Experimental Evaluation

The experimental evaluation confirms TinyGPT-V's capability to perform on par with larger LLMs while operating under more constrained resource conditions. It exhibits proficiency across benchmarks like GQA and IconVQ, demonstrating strong visual reasoning and language generation abilities.

Figure 2: Architecture of TinyGPT-V, featuring advanced mechanisms like LoRA and QK Normalization.

Qualitative assessments highlight TinyGPT-V's proficiency in concise and accurate responses, outperforming other models in tasks requiring nuanced interpretation and contextual understanding. Moreover, the model's rapid inference time and low resource consumption make it highly suitable for practical applications, as shown in comparison metrics with models like LLaVA and MiniGPT-4.

Conclusion

TinyGPT-V exemplifies a significant step toward creating high-performing MLLMs that prioritize efficiency and accessibility. By leveraging small backbones and incorporating advanced architectural techniques, TinyGPT-V achieves a balance between computational footprint and task performance. This model sets a precedent for future developments in multimodal AI applications, emphasizing the potential for compact models to achieve substantial theoretical and practical benefits.