GSOC 26 propsal idea

[ayushk687]

Hi @markgarnold @alex

I hope you’re both doing well. My name is Ayush, and I recently came across the project on adding support for Logarithmic Number Systems (LNS) in Large Language Models, and I found it extremely fascinating. The idea of exploring an alternative to floating point arithmetic—especially in the context of LLMs and attention mechanisms—really caught my interest.

I have experience working with Python, GitHub, and I’m actively building my understanding of systems-level concepts and performance-oriented computing. While I am still strengthening my C++ skills, I’m highly motivated to learn quickly and dive deep into this project, especially given its strong blend of numerical methods, low-level optimization, and modern AI systems.

What excites me most is the challenge of integrating xlnscpp into ggml and enabling LNS-based computation within frameworks like llama.cpp. The concept of designing a “virtual” LNS backend that behaves like a hardware target is particularly interesting, along with exploring how attention and matrix operations behave under LNS representation. I’m also curious about trade-offs in accuracy, performance, and conversion overhead, and how they impact real LLM outputs.

I would love to get started by understanding the current ggml architecture, exploring xlnscpp, and identifying smaller entry points where I can begin contributing. Please let me know how I can best prepare or where I should start.

Looking forward to your guidance and the opportunity to contribute!