Keywords
LLM, Model Compression, Pruning, CNN, Efficiency

Technologies
PyTorch, Hugging Face, LoRA, Flan-T5, VGG16

Large language models achieve strong performance but are computationally expensive to deploy due to their size and large number of parameters. This project investigates whether compression techniques originally developed Convolutional neural networks transfer to transformer-based models. VGG16 and RegNet were used to evaluate pruning, low-rank compression and quantisation using accuracy and model size. These methods were then applied to FLAN-T5 with LoRA fine-tuning and evaluated using ROUGE-L metrics. Results show that quantisation and low-rank compression remain stable across architectures, while pruning is more sensitive in transformers, highlighting the need for architecture-aware compression strategies.