Quantization of large language models with overdetermined basis

Large Language Models (LLM) require large computational resources. In our work, it was shown that using Kashin decomposition (decomposition of a vector over an overdetermined basis), it is possible to obtain a significant reduction in the number of parameters while maintaining accuracy. At the same time, the decomposition method itself does not require large computational resources and provides theoretical guarantees. The report itself will also provide an overview of the field of compact representations of language models, highlighting key works and open problems.